Wearable Technologies and Artificial Intelligence in Sports Medicine: A Systematic Narrative Review of Injury Prevention and Rehabilitation

ABSTRACT

Background and Aims

Wearable technologies and artificial intelligence (AI) are shifting sports medicine from episodic, subjective assessment toward continuous, data‐driven injury surveillance. This review aimed to map the main wearable sensor modalities used in injury prevention and rehabilitation, evaluate AI architectures applied to biomechanical and physiological data, summarize clinically relevant applications by anatomical region, and identify methodological, ethical, and implementation challenges.

Methods

A systematic narrative review was conducted across PubMed, Scopus, Web of Science, and Google Scholar for studies published from 2016 to 2026, supplemented by manual reference screening. Search terms combined wearable sensors, AI and machine learning, and sports medicine outcomes. After title and abstract screening of 614 records and full‐text assessment of 87 articles, 28 studies met the inclusion criteria.

Results

Inertial measurement units, GPS/GNSS, surface electromyography, and photoplethysmography were the most frequently reported sensor modalities. Deep learning, ensemble methods, and multimodal fusion were commonly used to process biomechanical and physiological signals. One hybrid IMU‐sEMG study involving 50 athletes reported 92.3% accuracy, 90.5% recall, and an area under the receiver operating characteristic curve of 0.93 for classifying biomechanical injury‐risk states. An ankle‐monitoring study using data from 87 athletes across 12 sports reported 87.4% overall detection accuracy, a 12.1% false‐alarm rate, and 17.2 ms latency. However, substantial heterogeneity in study design, populations, outcomes, and validation methods prevented direct comparison or pooled quantitative interpretation.

Conclusion

Wearable‐AI systems show promising potential for injury‐risk monitoring, rehabilitation, and return‐to‐play support. Wider clinical adoption will require more diverse populations, longitudinal and field‐based validation, improved model interpretability, standardized protocols, and robust ethical, legal, and governance frameworks.

1. Introduction

The integration of wearable sensors into athletes' daily training and competitive environments has fundamentally transformed the landscape of modern sports medicine [1, 2]. Traditionally, injury risk assessment and rehabilitation monitoring relied heavily on periodic clinical evaluations, observational judgments, and self‐reported symptoms. Although such approaches remain valuable, they are inherently limited by subjectivity, intermittent sampling, and their inability to capture subtle physiological or biomechanical deviations as they emerge over time. As a result, early warning signs of overload, neuromuscular fatigue, asymmetrical movement, or maladaptive recovery may go undetected until they culminate in significant injury or performance decline. In contrast, the emergence of modern wearable technologies has enabled continuous, high‐frequency monitoring of athlete status, creating a care model that shifts from reactive to proactive, preventive sports medicine practice [3, 4].

This transition is especially important in contemporary sport, where training intensity, fixture congestion, and performance expectations place athletes under unprecedented physical and psychological demands. In such high‐stakes settings, a single injury can have consequences far beyond temporary absence from play, including reduced competitive success, disrupted career trajectories, prolonged rehabilitation costs, and substantial financial burdens for athletes, teams, and medical staff. Consequently, the dual objectives of maximizing performance and minimizing injury incidence have become central pillars of elite and subelite sports management. Wearable technologies help bridge this gap by providing individualized, real‐time data on movement quality, training load, physiological readiness, and recovery dynamics, thereby allowing practitioners to detect emerging risk factors before they progress into clinically significant problems [5, 6].

The clinical value of wearable monitoring is particularly evident in rehabilitation and return‐to‐play settings. Historically, rehabilitation progression often relied on broad timelines, therapist observation, and the athlete's own perception of readiness. While these indicators are useful, they may not fully reflect whether the recovering athlete has restored symmetrical movement mechanics, sufficient neuromuscular control, or the capacity to tolerate sport‐specific loading. Wearable systems now enable tracking objective rehabilitation benchmarks, such as limb symmetry indices, joint kinematics, acceleration patterns, muscle activation timing, impact exposure, and physiological recovery markers, in real time [7, 8]. This supports more precise decision‐making, allowing clinicians to tailor rehabilitation progression to the athlete's actual functional status rather than to generalized protocols or subjective impressions alone.

However, the growing adoption of wearable monitoring has also introduced major analytical challenges. Modern sports wearables generate dense, multimodal streams of biomechanical and physiological data, often collected continuously across training sessions, rehabilitation exercises, and competition. Inertial measurement units (IMUs) capture acceleration, angular velocity, and orientation; surface electromyography (sEMG) reflects muscle activation behavior; GPS and Global Navigation Satellite Systems (GNSS) systems quantify external workload; and optical or cardiovascular sensors provide insights into internal physiological responses. The scale, complexity, and heterogeneity of these signals make them difficult to interpret using conventional statistical methods alone. Robust Artificial Intelligence (AI) and machine learning frameworks are therefore increasingly required to denoise raw signals, detect latent patterns, classify risk states, and generate clinically meaningful predictions from complex sensor ecosystems [9, 10].

Advanced computational approaches, including deep learning, temporal modeling, ensemble learning, and graph‐based methods, have become central to this evolving paradigm [11, 12]. These models can identify high‐risk kinematic signatures, quantify deviations from normative movement patterns, and integrate multiple sensor channels into unified injury‐risk or rehabilitation‐readiness assessments. Their value lies not only in prediction accuracy, but also in their ability to support individualized medicine in sport, where each athlete presents a unique combination of anatomical, physiological, and training‐related factors. As wearable sensing becomes more sophisticated, the convergence of hardware innovation and AI‐driven analytics is reshaping how clinicians, coaches, and sports scientists understand injury mechanisms, monitor recovery, and optimize long‐term athlete health.

Against this background, the present review aims to provide a comprehensive synthesis of contemporary wearable injury‐tracking systems, with particular emphasis on the sensor technologies, computational frameworks, and practical implementation strategies currently shaping the field. It examines the state of the art in wearable hardware, the artificial intelligence models used to process multimodal sports data, and the key clinical applications spanning injury prevention, rehabilitation, and return‐to‐play monitoring. In addition, this review identifies persistent methodological limitations, translational barriers, and emerging research directions. By critically analyzing the technological landscape and its clinical implications, this paper advocates for a more integrated, data‐driven model of sports medicine that prioritizes athlete longevity, safer rehabilitation, and sustainable peak performance.

1.1. Literature Review

From 2020 to 2026, the literature on wearable technology in sports medicine moved from basic athlete monitoring toward integrated systems that support injury prevention, rehabilitation, and decision‐making. Early work in this period focused strongly on workload surveillance and field‐based monitoring, whereas later reviews increasingly described wearables as part of wider digital ecosystems that combine biomechanical sensing, physiological monitoring, and artificial intelligence. This transition reflects a broader change in sports medicine from descriptive observation toward predictive and individualized monitoring frameworks [4, 6, 13, 14].

A recurring finding across the review literature is the consolidation of a core sensor set for sports applications. Inertial measurement units remain the dominant devices for tracking joint motion, acceleration, angular velocity, and movement asymmetry. Global Positioning System (GPS) and GNSS systems are widely used in field and team sports for external‐load assessment, while physiological systems such as heart‐rate monitors, wearable ECG devices, surface electromyography, and heat‐monitoring technologies extend monitoring toward internal load, recovery, autonomic response, and thermophysiological stress. Recent reviews consistently show that sports medicine is moving away from reliance on a single sensor family and toward combined biomechanical and physiological monitoring strategies [4, 6, 14, 15, 16, 17].

The lower‐limb injury literature, especially research on the anterior cruciate ligament (ACL), has been one of the most active areas in this field. Reviews on ACL rehabilitation and portable sensing show growing interest in out‐of‐lab monitoring of gait, landing mechanics, asymmetry, and functional progression after injury or reconstruction. This is important because ACL assessment has traditionally depended on intermittent clinic visits and isolated functional tests, whereas wearable systems allow more continuous and ecologically valid tracking of movement quality and loading tolerance. The literature, therefore, increasingly positions wearables not only as risk‐detection tools but also as clinically useful systems for rehabilitation monitoring and return‐to‐sport support [16, 18, 19, 20].

A second major literature stream concerns team‐sport training load and injury prevention. The acute:chronic workload ratio review helped establish the association between recent and cumulative load patterns and non‐contact injury risk, while later reviews broadened this discussion through wearable monitoring, biomechanical assessment, and digital injury‐management tools in team sports. Across these publications, wearables are presented not merely as tracking devices but as instruments that can support training adaptation, workload governance, and reinjury prevention when their outputs are interpreted within an appropriate clinical and coaching framework. At the same time, these reviews also show that methodological and practical inconsistencies remain common across sports, devices, and analytical approaches [6, 13, 14, 21].

Artificial intelligence became a defining topic in the review literature from 2022 onward. Narrative, scoping, and systematic reviews consistently argue that wearable data streams are too dense, multimodal, and context‐dependent to be interpreted adequately with simple thresholds alone. AI and machine‐learning models are therefore increasingly used to detect latent movement signatures, classify biomechanical states, estimate risk, personalize interventions, and support rehabilitation analysis. The recent literature also makes clear that AI in sports is no longer limited to performance analytics; it is now closely tied to injury prediction, clinical interpretation, and individualized athlete management [3, 22, 23, 24].

Within this AI‐focused literature, several methodological themes recur. Deep learning and neural network approaches are frequently used when high‐dimensional temporal or multimodal data must be modeled, while more traditional machine learning approaches remain attractive when data sets are smaller or more structured. Recent reviews also point toward multimodal fusion, explainable AI, and emerging digital‐twin frameworks as important next steps because they address key limitations of earlier systems, especially interpretability, personalization, and the integration of multiple physiological and biomechanical streams. More recent 2026 reviews further suggest that real‐time feedback and AI‐enhanced monitoring systems are becoming central to future sports‐medicine platforms [7, 16, 22, 24, 25].

Sport‐specific studies help clarify the practical value of this technology. Real‐life tennis applications demonstrate how wearable devices can quantify training load and recovery in an individualized setting, while cardiac monitoring and heat illness reviews highlight the expanding role of physiological wearables in athlete safety and medical surveillance. Team‐sport reviews also confirm that wearable systems are increasingly embedded in field environments rather than being limited to laboratory‐based experimentation. Together, these studies indicate that the usefulness of wearable systems depends heavily on matching the sensor choice and analytical framework to the actual injury mechanism, sport context, and clinical question being addressed [15, 17, 20, 21, 26].

Despite rapid progress, the literature repeatedly identifies major translational barriers. Strong performance in laboratory studies does not automatically ensure real‐world robustness, and many reviews note persistent problems involving heterogeneous validation methods, inconsistent sampling and preprocessing, modest sample sizes, limited population diversity, and incomplete reporting of interpretability and uncertainty. Ethical and governance concerns are also becoming more prominent, especially regarding data ownership, privacy, third‐party platform use, athlete autonomy, and the responsible deployment of predictive systems. These concerns are increasingly central to the literature, but they still receive less systematic treatment than sensor hardware and algorithmic performance [4, 6, 14, 21, 27].

Against this background, the present review is justified by the need for a broader and more clinically organized synthesis. Previous reviews have often emphasized only one side of the field, such as wearable biomechanics, AI methods, rehabilitation use, or team‐sport monitoring. A stronger review should connect sensor technologies, AI architectures, anatomy‐specific injury applications, rehabilitation pathways, implementation barriers, and ethical, legal, and social considerations within a single translational framework. That is the gap this review is designed to fill, particularly by integrating injury prevention and rehabilitation monitoring with emerging paradigms such as explainable AI, federated or privacy‐aware approaches, and digital‐twin thinking [3, 4, 14, 27].

2. Methods

2.1. Literature Search Strategy

A systematic narrative review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA 2020) guidelines to ensure a transparent, structured, and reproducible study selection process. The literature search strategy was designed to comprehensively capture studies at the intersection of wearable sensing technologies, artificial intelligence, and sports medicine applications. To achieve this, Boolean operators such as AND and OR were used to combine search terms across three major conceptual domains. The first domain focused on wearable technologies and included keywords such as wearable sensors, inertial measurement unit (IMU), GPS, surface electromyography (sEMG), accelerometer, and gyroscope. The second domain addressed artificial intelligence methodologies and incorporated terms including machine learning, deep learning, neural network, data fusion, random forest, and XGBoost. The third domain targeted sports medicine outcomes and clinical‐performance applications, using terms such as injury prevention, rehabilitation, biomechanics, ACL, concussion, and ankle sprain. By integrating these three domains, the search was structured to identify studies that specifically examined how wearable sensor data, when analyzed using AI‐driven methods, can support injury risk assessment, performance monitoring, rehabilitation tracking, and biomechanical evaluation in athletic populations.

2.2. Databases and Search Period

The literature search was conducted across four major scientific databases, namely PubMed, Scopus, Web of Science, and Google Scholar, to ensure broad and multidisciplinary coverage of the relevant literature. These databases were selected because they collectively index high‐impact research in biomedical sciences, engineering, computer science, rehabilitation, and sports medicine, thereby providing comprehensive access to studies related to wearable technologies, artificial intelligence, and clinical or performance‐oriented outcomes in athletic populations. The search period spanned from January 2016 to April 2026, a timeframe chosen to capture both the early foundational work in wearable sensing and machine learning applications, as well as the most recent developments in deep learning, multimodal sensor integration, and real‐time sports medicine analytics. In addition to the database search, the reference lists of all selected full‐text articles were manually screened to identify potentially relevant studies that may not have been retrieved through the primary search strategy. This supplementary backward‐reference searching helped enhance the completeness of the review and reduce the likelihood of omitting influential or contextually important publications.

2.3. Inclusion and Exclusion Criteria

Studies were considered eligible for inclusion if they satisfied several predefined criteria aligned with the aims of this review. First, the study had to focus on the use of wearable technologies within a sports or athletic context, including applications involving performance monitoring, biomechanical assessment, movement analysis, or physiological tracking during training, competition, or recovery.

Second, the study needed to address at least one relevant sports medicine outcome, such as injury prevention, injury risk screening, monitoring of physical load, rehabilitation progress, return‐to‐play evaluation, or functional recovery after musculoskeletal injury.

Third, eligible studies were required to incorporate artificial intelligence techniques or quantitative analytical approaches, including machine learning, deep learning, statistical modeling, signal processing, sensor fusion, or other data‐driven methods used to interpret wearable sensor outputs.

Fourth, eligible publication types included peer‐reviewed original research articles and systematic or scoping reviews that reported a reproducible search and selection methodology. Narrative reviews, opinion articles, editorials, conference abstracts without complete data, dissertations, preprints, and non‐peer‐reviewed publications were excluded.

Finally, only studies published in English were considered for analysis. Studies were excluded if they did not align with the review's thematic scope or methodological standards. Specifically, articles were excluded if they were unrelated to sports, exercise science, athletic performance, or human biomechanics, or if they focused on unrelated consumer health or general wellness applications without clear sports medicine relevance.

Publications were also excluded when they lacked sufficient methodological clarity, did not report empirical findings, or failed to provide enough detail regarding wearable devices, analytical methods, or study outcomes. In addition, opinion pieces, editorials, conference abstracts without full data, dissertations, magazine articles, and other non‐scientific publications were excluded to preserve the quality and consistency of the evidence base. Duplicate records identified across databases were removed during the screening process to ensure that each study was assessed only once.

2.4. Study Selection and PRISMA Flow

Figure 1 presents the PRISMA 2020 flow diagram summarizing the identification, screening, eligibility assessment, and final inclusion of studies in this narrative review. In total, 1072 records were identified through the initial search process, including 312 records from PubMed, 287 from Scopus, 241 from Web of Science, 198 from Google Scholar, and 34 additional records from manual reference screening. After removing 458 duplicate records, 614 unique articles remained for title and abstract screening.

Figure 1.

PRISMA 2020 flow diagram of the systematic literature search and study selection process. Records (n = 1072) were identified from four databases (PubMed: n = 312; Scopus: n = 287; Web of Science: n = 241; Google Scholar: n = 198) and manual reference screening (n = 34). After duplicate removal (n = 458), 614 records proceeded to title/abstract screening, of which 527 were excluded. Full‐text assessment of 87 articles resulted in 28 studies meeting inclusion criteria for the final narrative synthesis. This flow diagram follows the PRISMA 2020 reporting standards, documenting the number of records at each stage of identification, screening, eligibility, and inclusion. PRISMA, Preferred Reporting Items for Systematic Reviews and Meta‐Analyses; n, number of records.

During title and abstract screening, 527 of 614 records were excluded for the following mutually exclusive primary reasons: lack of relevance to wearable technologies in sports medicine (n = 218); absence of AI‐based or quantitative analytical methods (n = 156); nonempirical or non‐peer‐reviewed publication status (n = 89); and insufficient methodological detail to assess eligibility (n = 64). The remaining 87 articles were retrieved for full‐text assessment. At the full‐text stage, 59 of 87 articles were excluded for the following mutually exclusive primary reasons: insufficient description or application of AI/ML methods (n = 18; 30.5% of full‐text exclusions); absence of a clearly defined sports medicine outcome (n = 12; 20.3%); insufficient methodological information for critical appraisal (n = 10; 16.9%); ineligible publication type without original empirical data or reproducible systematic review methods (n = 8; 13.6%); non‐peer‐reviewed publication status (n = 6; 10.2%); and a population not relevant to sports or athletic contexts (n = 5; 8.5%). The full‐text exclusion rate of 67.8% reflects the requirement that eligible studies address all three core dimensions of the review: wearable sensing technology, AI/ML or quantitative analytical methods, and an application related to sports injury prevention, rehabilitation, or return‐to‐play monitoring. Following full‐text assessment, 28 studies met all eligibility criteria and were included in the final narrative synthesis.

No original statistical analyses were performed as part of this review. Therefore, no statistical tests, a priori significance thresholds, one‐sided or two‐sided testing procedures, or statistical software were used. Performance measures are reported as presented by the authors of the included studies. Measures of uncertainty, including confidence intervals, standard deviations, or interquartile ranges, are reported only when they were explicitly provided in the original publication. No pooled estimates were calculated.

3. Core Technologies in Wearable Injury Tracking

Modern wearable injury tracking systems rely on the integration of multiple sensing modalities capable of capturing complementary biomechanical, physiological, and positional information during athletic activity. Rather than depending on a single data source, contemporary systems combine motion‐based, location‐based, and physiological sensors to generate a more comprehensive picture of athlete status, workload exposure, and injury risk. This multimodal approach is particularly important in sports medicine because injury development is rarely attributable to a single variable; instead, it emerges from the interaction of mechanical loading, neuromuscular fatigue, movement quality, recovery state, and environmental stress. Figure 2 illustrates the anatomical placement of key wearable sensor types across the body, highlighting how different body regions can be instrumented to monitor specific functional and physiological parameters. The following subsections summarize the principal technologies that underpin wearable injury tracking systems and explain their roles in performance monitoring, injury prevention, and rehabilitation.

Figure 2.

Anatomical placement of wearable sensor modalities used in sports injury prevention and rehabilitation monitoring. The figure illustrates a human figure with six sensor types positioned at their typical anatomical locations: (1) Inertial Measurement Units (IMUs; blue) placed on the trunk, thigh, shank, and foot to capture joint kinematics, acceleration, and orientation; (2) GPS/GNSS vest (green) worn on the torso for external load and positional tracking; (3) Surface electromyography (sEMG) electrodes (orange) on major muscle groups (quadriceps, hamstrings, gastrocnemius) to measure neuromuscular activation; (4) Photoplethysmography/heart rate monitor (red) at the wrist for cardiovascular and autonomic assessment; (5) Smart helmet accelerometers (purple) for head impact quantification in contact sports; and (6) Pressure insoles (yellow) within footwear for ground reaction force and gait symmetry assessment. Color coding facilitates rapid identification of each sensor modality. GNSS, Global Navigation Satellite System; GPS, Global Positioning System; IMU, Inertial Measurement Unit; PPG, Photoplethysmography; sEMG, surface electromyography.

3.1. Inertial Measurement Units (IMUs)

IMUs are among the most widely used technologies in wearable sports monitoring because they provide rich, continuous information about human movement in real‐world environments. A typical IMU combines a three‐axis accelerometer, a three‐axis gyroscope, and, in many cases, a three‐axis magnetometer to quantify linear acceleration, angular velocity, and magnetic orientation. When attached to specific anatomical landmarks, such as the trunk, thigh, shank, foot, or upper limb, IMUs can capture detailed movement signatures of posture, gait, joint kinematics, impact loading, and dynamic limb coordination. Depending on the device configuration, these sensors may operate at sampling frequencies of up to 1000 Hz, allowing them to detect both gross movement patterns and subtle high‐frequency biomechanical events such as foot strike impacts, rapid decelerations, or asymmetrical landing strategies [28, 29]. In clinical and sports performance contexts, IMUs are particularly valuable because they enable objective assessment outside of laboratory motion‐capture systems. They can be used to quantify movement asymmetries, monitor jump mechanics and landing control, assess sprint acceleration profiles, and provide real‐time biomechanical feedback during rehabilitation exercises. For example, in volleyball and other jumping sports, IMU‐based systems such as the Vert Classic device are used to track jump count, jump height, and cumulative loading, thereby helping practitioners manage fatigue and identify excessive mechanical stress. In rehabilitation settings, IMUs also allow clinicians to monitor whether athletes are restoring symmetrical movement patterns following injury, which is critical during return‐to‐play decision‐making. Their portability, relatively low cost, and capability for field‐based deployment make IMUs a foundational component of modern wearable injury tracking systems.

3.2. GPS/GNSS and External Load Management

GNSS, including the GPS, are widely used to quantify external load during outdoor sports training and competition. These systems estimate athlete position, velocity, acceleration, and total distance covered by continuously tracking spatial movement over time. In team sports such as football, rugby, and American football, GNSS‐based wearables are routinely employed to characterize running demands, high‐speed efforts, sprint frequency, deceleration loads, and positional movement profiles. Such information provides coaches and sports medicine staff with an objective means of understanding how much physical work an athlete has performed during a given session or competition. One of the most frequently discussed workload metrics derived from GPS/GNSS monitoring is the Acute‐to‐Chronic Workload Ratio (ACWR), which compares short‐term training load, often represented by a 7‐day rolling average, with longer‐term habitual load, commonly represented by a 28‐day rolling average [29]. The rationale behind this metric is that rapid spikes in short‐term workload relative to an athlete's established training base may increase injury susceptibility, particularly when tissue adaptation and recovery are insufficient. ACWR values exceeding 1.5 have been associated with elevated injury probability and are often interpreted as indicative of a potentially hazardous workload imbalance. As a result, external load monitoring using GNSS has become a central part of training management in many professional sporting environments, including the NFL, elite football, and professional rugby [30]. Beyond injury prevention, these data also support return‐to‐play planning by enabling gradual re‐exposure to sport‐specific demands and ensuring that recovering athletes reacquire appropriate load tolerance before full reintegration.

3.3. Physiological and Biochemical Monitoring

In addition to movement and positional sensing, wearable systems increasingly incorporate physiological and biochemical monitoring to assess internal load, recovery status, neuromuscular readiness, and environmental strain. These measures provide insight into how the athlete's body is responding to training and competition, thereby complementing biomechanical indicators with markers of systemic stress and adaptation. Because injury risk is influenced not only by external mechanical exposure but also by fatigue, autonomic regulation, thermoregulation, and muscle activation quality, physiological monitoring has become an essential dimension of wearable injury tracking.

• Heart Rate (HR) and Heart Rate Variability (HRV) are among the most commonly used cardiovascular and autonomic measures. Heart rate reflects the immediate cardiovascular response to exercise intensity, whereas HRV provides information about the balance between sympathetic and parasympathetic nervous system activity. Higher HRV is generally interpreted as a marker of better recovery, improved autonomic flexibility, and greater readiness for high‐intensity training, while suppressed HRV may indicate fatigue, insufficient recovery, or excessive physiological stress [16, 25]. In practice, HR and HRV are used to guide training intensity, monitor adaptation over time, and identify periods in which athletes may be more vulnerable to overload or suboptimal performance [26, 31].

• Surface electromyography (sEMG) measures the electrical activity associated with skeletal muscle activation and is particularly useful for assessing neuromuscular coordination, fatigue development, and bilateral asymmetry. In sports medicine, sEMG can reveal altered recruitment patterns, compensatory activation strategies, and declining neuromuscular efficiency, all of which may precede musculotendinous injury or indicate incomplete recovery after injury [19, 27]. For instance, asymmetrical activation between dominant and nondominant limbs, or delayed recruitment in stabilizing muscle groups, may signal elevated mechanical vulnerability during high‐load movements such as sprinting, cutting, or jumping. sEMG is therefore valuable not only in injury detection but also in rehabilitation monitoring and functional reassessment [7, 32].

• Galvanic skin response (GSR) and skin temperature provide information about sweat gland activity, psychophysiological arousal, and thermal strain. These measures are especially relevant in high‐intensity or hot‐environment sports, where thermoregulatory stress can impair performance and increase the risk of exertional heat illness [12, 15]. Monitoring changes in skin conductance and temperature can help identify periods of excessive physiological burden, dehydration risk, or abnormal heat accumulation. When integrated with workload and heart‐rate metrics, these signals can contribute to a more holistic understanding of athlete strain and readiness [9, 33].

• Photoplethysmography (PPG) is another widely adopted sensing modality, particularly in wrist‐worn and consumer‐grade wearable devices. PPG uses optical light transmission or reflection to estimate blood volume changes in the microvascular bed, enabling indirect measurement of heart rate and, in some systems, peripheral oxygen saturation [4, 14]. Its noninvasive and user‐friendly design makes it attractive for continuous monitoring in both training and recovery settings. Although PPG may be more sensitive to motion artifacts than electrocardiogram‐based systems, it remains a practical solution for real‐time physiological tracking in applied sports environments.

Taken together, these physiological and biochemical sensing approaches expand wearable injury tracking beyond simple movement quantification. By capturing internal responses to training load, neuromuscular status, and environmental stress, they enhance the ability of clinicians and performance staff to detect early warning signs of overload, personalize rehabilitation programs, and make more informed return‐to‐play decisions.

4. Anatomical Assessment of Injuries and Sports Applications

This section organizes the clinical applications of wearable injury tracking systems according to anatomical region, thereby providing a structured framework for understanding how different sensing technologies are applied to distinct injury profiles across the body. Rather than discussing wearable applications in a purely technical manner, this region‐based approach emphasizes their practical relevance in sports medicine by linking each technology to specific tissues, movement patterns, and injury mechanisms commonly encountered in athletic populations. Attention is given to how wearable systems support the detection of abnormal biomechanics, excessive loading, neuromuscular deficits, and recovery‐related changes in different body regions, including the lower extremity, upper extremity, trunk, and head. This organization also facilitates clearer interpretation of how injury risk factors vary depending on the demands of the sport, the anatomical structure involved, and the type of movement being performed. In addition, real‐world sports examples are incorporated throughout the section to highlight the translational value of wearable monitoring in training environments, competition settings, and rehabilitation programs. Figure 3 illustrates the principal biomechanical mechanisms underlying the most prevalent sports injuries monitored by wearable systems, providing a visual summary of how movement errors, overload patterns, impact forces, and neuromuscular imbalances contribute to injury development across anatomical regions.

Figure 3.

Biomechanical mechanisms of the six most prevalent sports injuries monitored by wearable sensor systems. The figure is organized as a 3 × 2 panel layout (left to right, top to bottom): (A) Anterior Cruciate Ligament (ACL) injury: Depicts dynamic knee valgus collapse during single‐leg landing, characterized by excessive knee abduction angle, reduced knee flexion, and elevated ground reaction force loading. (B) Ankle sprain: Illustrates the subtalar inversion mechanism during cutting or landing, showing excessive inversion angle and plantar flexion that exceed the ligamentous restraint capacity of the lateral ankle ligaments. (C) Hamstring strain injury: Shows eccentric overload of the biceps femoris during the terminal swing phase of sprinting, with muscle‐tendon unit elongation exceeding safe limits. (D) Concussion: Displays linear and rotational head acceleration vectors resulting from impact to the helmeted head, with direction and magnitude indicated by arrows. (E) Tibial stress fracture: Represents cumulative ground reaction force loading transmitted through the tibia during running, with stress concentration indicated on the anterior tibial cortex. (F) Shoulder overuse injury: Illustrates repetitive stroke mechanics in tennis, highlighting excessive shoulder internal rotation velocity and elbow valgus stress during the serving motion. Each panel uses arrows, force vectors, and anatomical highlights to indicate the primary biomechanical dysfunction. ACL, anterior cruciate ligament; GRF, ground reaction force.

4.1. Head and Torso: Impact Biomechanics and Systemic Risks

In contact and collision sports such as rugby, American football, ice hockey, and soccer, head trauma and concussion remain major concerns because of their potential short‐term neurological effects and long‐term cumulative consequences. Wearable monitoring systems have therefore been increasingly developed to quantify head impact exposure in real time and support more objective sideline assessment. Smart helmet systems, instrumented mouthguards, and head‐mounted sensor platforms typically incorporate arrays of accelerometers and gyroscopes capable of measuring both linear acceleration and rotational acceleration, the two principal biomechanical components associated with concussive and subconcussive events [34, 35]. These systems provide immediate quantitative data on impact magnitude, direction, and frequency, allowing medical teams to identify potentially dangerous impacts that may not be obvious from visual observation alone. This is especially important because athletes may underreport symptoms or continue playing despite subtle neurological impairment.

Beyond acute event detection, wearable impact sensors have also contributed to research on the biomechanical signatures of repetitive sub‐concussive loading. Systematic reviews have used these data to characterize sport‐specific impact profiles, identify high‐risk collision scenarios, and distinguish between isolated severe impacts and repetitive lower‐magnitude exposures that may accumulate over time [30]. Such insights are valuable not only for diagnosis and return‐to‐play decision‐making, but also for prevention, as they inform rule modifications, helmet design improvements, and safer tackling or heading techniques. In this way, wearable head‐impact monitoring systems are increasingly positioned as both clinical and preventive tools within sports medicine.

Torso and Systemic Health (Heat Stroke and Fatigue): Wearable technologies are also highly relevant for monitoring whole‐body physiological strain, particularly in relation to exertional heat illness, systemic fatigue, and cumulative overload. These conditions often develop progressively and may not be recognized early through observation alone, making continuous physiological monitoring especially valuable in high‐performance sport. One example is the Wearable Heat‐Stroke‐Detection Device (WHDD), which integrates multiple physiological and environmental signals, including galvanic skin response (GSR), heart rate, and ambient condition monitoring, to identify patterns associated with rising thermal stress and increased risk of exertional heat stroke [7, 9]. By combining internal physiological responses with external environmental conditions, such systems provide a more comprehensive assessment of an athlete's heat burden during training or competition, particularly in hot and humid settings.

Systemic fatigue monitoring extends beyond thermal strain and increasingly incorporates indicators such as heart rate variability (HRV), sleep quality measures, and movement efficiency indices [12, 33]. HRV offers insight into autonomic recovery and readiness, while sleep‐related metrics help identify insufficient recovery that may predispose the athlete to impaired neuromuscular control and reduced physiological resilience. Movement efficiency indices, derived from wearable motion sensors, can reveal subtle biomechanical deterioration during repeated effort, such as altered posture, reduced coordination, or less economical movement strategies. These changes may reflect the onset of fatigue before overt performance decline becomes clinically apparent.

A particularly valuable application is cumulative load monitoring, in which external mechanical workload is interpreted alongside internal physiological or neuromuscular responses. For example, GPS‐derived workload measures can be combined with sEMG‐based muscle activity data to determine whether increasing training demands are being matched by appropriate physiological adaptation or are instead producing inefficient movement patterns and excessive neuromuscular strain [28, 36]. This multimodal monitoring approach helps identify the point at which loading transitions from productive training stress to maladaptive overload or non‐functional overreaching. Clinically, this information is useful for adjusting training volume, preventing fatigue‐related injuries, and optimizing return‐to‐play progression after illness or musculoskeletal injury. Overall, wearable systems for torso and systemic health monitoring broaden the scope of injury surveillance by capturing not only local biomechanical dysfunction, but also the whole‐body physiological conditions that influence athletic readiness and injury susceptibility.

4.2. Upper Limbs: Overuse and Stroke Mechanics

In individual sports such as tennis, wearable devices can monitor stroke mechanics and serve speed to support the prevention of shoulder and elbow overuse injuries [24, 26]. By identifying at‐risk movement patterns during repetitive strokes, characterized by excessive shoulder internal rotation velocity and elbow valgus stress, these devices enable early intervention before conditions such as lateral epicondylitis become chronic [4, 6].

4.3. Lower Limbs: Biomechanics and Specific Injury Focus

4.3.1. ACL Injuries

ACL injuries are characterized by identifiable biomechanical risk patterns during dynamic tasks, including dynamic knee valgus, excessive tibial internal rotation, insufficient hip‐knee control, and stiff landing mechanics during pivoting, cutting, deceleration, and single‐leg landings [32, 36]. Wearable systems combining IMUs with sEMG capture both mechanical execution and neuromuscular activation, providing a more complete risk assessment than kinematic observation alone [21, 28]. The primary sensor configuration for ACL monitoring comprises lower‐limb IMUs (typically placed on the thigh, shank, and foot) paired with quadriceps and hamstring sEMG electrodes. This configuration enables detection of high‐risk movement signatures during training and rehabilitation. A hybrid IMU‐sEMG framework evaluated in 50 athletes reported 92.3% accuracy, 90.5% recall, and an area under the receiver operating characteristic curve of 0.93 for classifying biomechanical injury‐risk states [7, 8]. The system also achieved an average feedback latency of 188 ± 15 ms. These results relate to the classification of biomechanical risk patterns and should not be interpreted as prospective prediction of confirmed ACL injuries. Further longitudinal validation is required to determine whether such classifications predict subsequent injury incidence.

4.3.2. Ankle Sprains

Lateral ankle sprains are among the most common injuries in field‐ and court‐based sports, arising from rapid direction changes, unstable foot placement, and repeated jumping and landing [7, 10]. Ankle‐mounted IMUs continuously monitor inversion angle, plantar flexion, and angular velocity, the key biomechanical components of sprain mechanisms, identifying when the subtalar joint approaches or exceeds safe kinematic ranges [33, 37]. Some systems provide near‐immediate (17.2 ms latency) haptic feedback when dangerous transitions are detected [33, 37]. An adaptive ankle‐monitoring framework was evaluated using synchronized inertial‐sensor and video‐labeled data from 87 athletes across 12 sports [7, 10]. The complete real‐time system reported 87.4% overall detection accuracy, a 12.1% false‐alarm rate, and an average latency of 17.2 ms. It also identified 76.3% of labeled injury‐related events at least 150 ms in advance. These results demonstrate the feasibility of low‐latency detection of high‐risk ankle‐motion events, although prospective clinical validation against subsequent ankle‐sprain incidence remains necessary.

4.3.3. Hamstring Strain Injuries

Hamstring strain injuries are strongly linked to sprinting mechanics, eccentric overload, and fatigue‐related disruption of neuromuscular coordination [11, 36]. Wearable systems focus on capturing the relationship between muscle activation timing and mechanical demand during the late swing and early stance phases of running. Random Forest models and other machine learning methods analyze wearable sensor outputs to detect fatigue‐related changes and loss of synchrony between neural activation and mechanical loading [11, 36]. Machine‐learning methods have been explored for identifying fatigue‐related changes in sprint mechanics, eccentric loading, and neuromuscular coordination. However, the available studies differ substantially in their sensor configurations, outcome definitions, and validation methods. The evidence reviewed does not support a single pooled or independently validated accuracy value for prospective hamstring‐strain prediction. These systems should therefore be described as promising monitoring and risk‐stratification tools rather than established injury‐prediction models.

4.3.4. Tibial Stress Fractures

Tibial stress fractures develop through cumulative mechanical loading without adequate recovery [14, 24]. Wearable accelerometers monitor Peak Tibial Acceleration (PTA) as a surrogate marker of repetitive impact stress. Some systems incorporate haptic biofeedback cues triggered when tibial loading thresholds are exceeded, prompting runners to increase step frequency and reduce vertical ground reaction forces [14, 24]. This closed‐loop feedback transforms passive monitoring into active injury‐prevention intervention.

4.4. Professional Team Versus Individual Sports

The practical implementation of wearable injury monitoring differs substantially between team sports and individual sports, largely because of differences in movement demands, tactical context, athlete volume, and monitoring priorities. In professional team sports, GPS/GNSS monitoring is commonly used to support workload management, rehabilitation progression, and return‐to‐play planning. However, the magnitude of its effect on injury and reinjury incidence remains difficult to establish because implementation strategies and outcome definitions vary across studies [29, 38]. In this setting, wearable technologies function not only as injury surveillance tools but also as population‐level decision‐support systems for performance and medical departments.

In individual sports, wearable monitoring is generally more focused on fine‐grained technical analysis, fatigue detection, and personalized biomechanical optimization. In combat sports, for example, IMUs and pressure sensors are used to differentiate skill levels, quantify strike mechanics, monitor balance and postural control, and detect fatigue‐related deterioration in movement precision [24, 32]. In running, wearable applications often emphasize gait symmetry, impact loading, cadence, and lower‐limb coordination to reduce the risk of overuse conditions such as plantar fasciitis, stress fractures, and tendon‐related disorders [4, 6]. Because individual sports often involve repetitive movement patterns performed at high volume, wearable systems are particularly valuable for detecting subtle deviations from efficient technique before they evolve into injury. Overall, while team sports emphasize group‐level workload management and reinjury reduction, individual sports place greater importance on technique‐specific, athlete‐specific monitoring, demonstrating the versatility of wearable systems across different competitive contexts.

4.5. Rehabilitation and Return‐to‐Play Monitoring

Wearable systems have substantially improved the rehabilitation process by enabling clinicians, physiotherapists, and sports medicine practitioners to monitor recovery objectively against predefined functional benchmarks rather than relying solely on subjective assessment or time elapsed since injury [7, 8]. This represents an important shift in rehabilitation philosophy, as successful recovery is increasingly understood to depend not only on symptom resolution, but also on the restoration of biomechanical symmetry, neuromuscular control, and tissue load tolerance. By continuously acquiring movement and physiological data, wearable devices provide detailed information about how the recovering athlete performs during rehabilitation exercises, allowing subtle deficits to be identified even when overt clinical symptoms appear to have resolved.

Real‐time monitoring frameworks are particularly valuable for assessing joint‐angle asymmetry, movement quality, and muscle‐force imbalances during progressive rehabilitation tasks such as squatting, lunging, landing, sprint reintroduction, and sport‐specific agility drills. By comparing the performance of the injured limb with that of the contralateral limb or with previously recorded baseline values, these systems can determine whether the athlete is regaining symmetrical function and whether compensatory movement strategies persist under increasing physical demand. This is critical because premature return to sport in the presence of unresolved asymmetry or neuromuscular deficits may increase the risk of reinjury, chronic dysfunction, or reduced long‐term performance. Wearable monitoring, therefore, supports more precise progression of rehabilitation intensity, ensuring that the recovering limb can tolerate loading demands that approach 100% of pre‐injury functional intensity before the athlete is cleared for unrestricted competition [7, 28]. Table 1 summarizes wearable technologies, key metrics, and clinical applications in sports injury rehabilitation.

Table 1.

Summary of wearable technologies, key metrics, and clinical applications in sports injury rehabilitation.

Technology	Metric(s) in rehabilitation	What It measures	Clinical benefit	Typical rehabilitation use	Example in sports medicine
IMUs	Joint angles, range of motion, limb symmetry, angular velocity, and impact force	Kinematic quality of movement and side‐to‐side biomechanical differences	Detects movement asymmetries, compensatory strategies, and abnormal loading during recovery	Monitoring return‐to‐run, jump‐landing mechanics, gait retraining, and postsurgical movement quality	Used after ACL reconstruction to identify asymmetrical knee loading and faulty landing patterns before return to play
sEMG	Neuromuscular activation magnitude, timing, co‐contraction, muscle onset/offset patterns	Muscle recruitment behavior and coordination deficits between injured and non‐injured limbs	Identifies muscle fatigue, delayed activation, and neural recruitment imbalances that may increase reinjury risk	Tracking quadriceps inhibition, hamstring activation, shoulder stabilization, and muscle re‐education programs	Applied in hamstring and quadriceps rehabilitation to assess whether the injured limb shows reduced or delayed activation compared with the healthy side
PPG/HR	Heart rate, Heart Rate Variability (HRV), pulse waveform trends, recovery rate	Autonomic nervous system recovery and cardiovascular readiness	Assesses systemic recovery status to determine readiness for progression to higher intensity rehabilitation loads	Monitoring daily recovery, exercise tolerance, fatigue status, and return‐to‐training readiness	Used in endurance and team‐sport athletes to determine whether physiological recovery supports progression from light rehabilitation to sport‐specific drills
GPS/GNSS	Total distance, high‐intensity running (HIR), sprint exposure, accelerations/decelerations, ACWR	External training load and movement volume during field‐based rehabilitation	Manages workload progression, reduces overuse risk, and identifies fatigue through speed or workload decline	Controlling running volume, reconditioning after muscle injury, and phased return‐to‐play programs	Common in football and rugby rehabilitation to ensure athletes progressively regain match‐related running demands without exceeding safe load thresholds
Pressure/Haptic Sensors	Plantar pressure distribution, gait symmetry, stance time, balance shifts, real‐time haptic feedback	Weight‐bearing behavior, posture control, and movement correction needs	Provides immediate alerts to correct posture, reduce stress on healing tissues, and improve gait mechanics	Gait retraining, postural correction, balance recovery, and lower‐limb unloading control	Used after ankle, foot, or knee injury to detect uneven weight distribution and guide safer movement through real‐time feedback

A particularly important concept in this context is the Limb Symmetry Index (LSI), which is commonly used to quantify functional equivalence between the injured and uninjured limbs. Thresholds of ≥ 90% are widely regarded as a minimum benchmark for return‐to‐play readiness, especially after lower‐limb injuries such as ACL reconstruction, ankle sprain, or muscle strain. The integration of LSI‐based decision criteria with continuous wearable monitoring represents a major advancement over traditional time‐based return‐to‐play protocols, which often assume recovery based on healing duration rather than demonstrated functional competence. In contrast, wearable‐guided rehabilitation provides an evidence‐based and individualized framework in which clearance decisions are grounded in real‐time biomechanical performance, progressive loading capacity, and objective symmetry restoration. As a result, these systems improve both the safety and precision of rehabilitation, while supporting more confident clinical decisions regarding return to training and competition. Table 2 presents a synthesis of the included studies on wearable technology for injury prevention.

Table 2.

Systematic synthesis of included studies on wearable technology for injury prevention.

Clinical domain	References	Sensor/analytical approach	Key injury mechanism detected	Diagnostics/monitoring metric	Clinical benefit	Typical rehabilitation or prevention use
ACL and Knee Stability	[4, 7, 36]	IMU‐based lower‐limb kinematic monitoring and landing biomechanics analysis	Detection of dynamic knee valgus, reduced knee flexion control, and inter‐limb asymmetry during high‐impact landing tasks	92.3% accuracy, 90.5% recall, and AUC 0.93 for classification of biomechanical injury‐risk states	Enables early identification of athletes at elevated ACL injury risk	Used in jump‐landing screening, neuromuscular retraining, and return‐to‐sport decision support
Ankle Sprains	[7, 10]	Real‐time wearable motion sensing with inversion‐velocity tracking	Monitoring of subtalar inversion velocity, unstable foot mechanics, and high‐risk directional transitions	87.4% overall detection accuracy, 12.1% false‐alarm rate, and 17.2 ms average latency	Supports immediate detection of hazardous ankle movement patterns	Applied in balance retraining, lateral movement drills, and recurrence prevention programs
Hamstring (HSI)	[11, 36]	AI‐driven biomechanical and neuromuscular pattern analysis	Assessment of eccentric braking deficits and altered neural recruitment during terminal swing phase	No single independently validated prospective injury‐prediction estimate was identified	Helps identify persistent deficits associated with hamstring reinjury risk	Used during sprint rehabilitation, return‐to‐run progression, and load‐management monitoring
Stress Fractures	[14, 24]	Tibial loading sensors and gait‐retraining systems with haptic feedback	Monitoring of peak tibial acceleration, repetitive impact loading, and maladaptive gait mechanics	Significant reduction in ground reaction forces	Reduces excessive skeletal loading and improves safer movement patterns	Applied in gait correction, progressive reloading, and prevention of recurrent bone stress injury
Head Trauma	[30, 34, 35]	Smart helmet systems and wearable head‐impact sensing platforms	Quantification of linear and rotational head acceleration to identify subconcussive and potentially concussive exposure	Real‐time impact magnitude mapping	Improves objective monitoring of head‐impact burden in contact sports	Used in sideline surveillance, exposure tracking, and impact‐informed return‐to‐play protocols
Heat and Systemic Risk	[9, 33]	Multi‐sensor physiological monitoring integrating GSR, temperature, and cardiovascular strain	Detection of thermal stress, autonomic overload, and early fatigue‐related physiological deterioration	Real‐time physiological alert systems	Supports early intervention before heat illness or systemic fatigue escalates	Used in hot‐environment training, recovery supervision, and exertional heat‐risk prevention
AI and Data Fusion	[21, 28]	Multimodal machine learning combining IMU, GPS, sEMG, and PPG data	Spatio‐temporal integration of internal physiological load with external biomechanical workload	Multimodal models showed promising discrimination in individual studies, although differences in outcomes and validation methods prevent direct comparison.	Provides more comprehensive and individualized injury‐risk assessment	Used in personalized monitoring systems, workload optimization, and integrated injury forecasting

5. AI Architecture and Data Fusion

Advanced artificial intelligence architectures are essential for transforming the high‐volume, high‐velocity, and multimodal data streams generated by wearable sensor systems into clinically meaningful insights. Unlike conventional monitoring approaches that rely on isolated variables or manually interpreted summaries, modern wearable platforms continuously acquire heterogeneous data from multiple sources, including inertial sensors, physiological monitors, positional trackers, and biochemical or environmental sensors. These data are often noisy, asynchronous, and highly dimensional, making them difficult to interpret using traditional analytical techniques alone. As a result, AI‐driven frameworks have become central to wearable sports medicine, enabling automated feature extraction, temporal pattern recognition, multimodal data fusion, and predictive modeling across complex sensor inputs.

The need for advanced AI is particularly pronounced because injury risk, fatigue, recovery status, and rehabilitation progression are not determined by a single measurement, but rather by dynamic interactions between biomechanical, physiological, and contextual variables. Machine learning and deep learning models are therefore used to identify latent relationships within the data, detect nonlinear dependencies, and classify athlete states with greater sensitivity than rule‐based systems. For example, AI models can integrate joint kinematics, muscle activation timing, heart rate variability, workload exposure, and movement symmetry into a unified analytical framework that supports injury‐risk screening, rehabilitation monitoring, and return‐to‐play decision‐making. In addition, these models can operate continuously and adaptively, allowing the monitoring system to move beyond retrospective analysis toward real‐time decision support.

Figure 4 illustrates the complete AI processing pipeline, beginning with raw sensor data acquisition and progressing through signal preprocessing, synchronization, noise reduction, feature extraction, multimodal fusion, and predictive modeling, before ultimately delivering clinically interpretable outputs for decision support. This end‐to‐end framework highlights how wearable data are converted from raw physiological and biomechanical signals into actionable information that can assist clinicians, physiotherapists, coaches, and sports scientists in identifying risk patterns, monitoring recovery trajectories, and personalizing intervention strategies. In this context, AI serves not merely as a computational tool but as the analytical core that enables wearable systems to function as intelligent, responsive platforms for modern sports medicine.

Figure 4.

Conceptual workflow showing how wearable sensor data are transformed through signal processing, feature extraction, and AI‐based modeling into clinically relevant decision support for injury prevention, rehabilitation monitoring, workload adjustment, and return‐to‐play assessment. The framework also includes a continuous feedback loop in which clinical use informs future data collection and model refinement. AI disclosure: The initial version of Figure 4 was generated using OpenAI image‐generation assistance and was subsequently reviewed and revised by the authors for scientific accuracy, clarity, and final presentation. AI, artificial intelligence; GNSS, Global Navigation Satellite System; GPS, Global Positioning System; HRV, heart rate variability; IMU, inertial measurement unit; PPG, photoplethysmography; sEMG, surface electromyography.

5.1. Established AI Paradigms

Several advanced artificial intelligence architectures have emerged as particularly well‐suited for analyzing wearable sensor data in sports medicine, owing to their ability to model complex, high‐dimensional, and time‐varying signals. Deep learning approaches, especially one‐dimensional Convolutional Neural Networks (1D‐CNNs) and Long Short‐Term Memory (LSTM) networks, are widely used to extract meaningful spatial and temporal patterns from high‐frequency biomechanical and physiological sensor streams. 1D‐CNNs are effective at learning localized discriminative features directly from raw sequential data, such as acceleration, angular velocity, or muscle activation signals, whereas LSTM networks are specifically designed to capture temporal dependencies and long‐range sequential relationships [11, 39]. This makes them particularly valuable for applications involving gait‐cycle analysis, fatigue progression, rehabilitation monitoring, and movement‐phase classification, where injury‐related abnormalities may unfold dynamically rather than appear as isolated events. In many cases, these deep architectures have demonstrated superior performance over conventional machine learning models in sequential biomechanical signal analysis because they reduce the need for manual feature engineering and can learn complex representations directly from sensor data.

In parallel, ensemble learning methods such as Random Forest and XGBoost remain highly relevant in wearable sports medicine due to their strong predictive performance, interpretability, and robustness with relatively small or moderately sized data sets. By aggregating the outputs of multiple decision trees, these algorithms can model non‐linear relationships between workload, kinematics, physiological markers, and injury outcomes while reducing the risk of overfitting. This is particularly advantageous in sports medicine studies, where dataset sizes are often limited by cohort availability, injury incidence, or the difficulty of collecting large volumes of well‐annotated clinical data. Ensemble methods are therefore commonly used for injury‐risk stratification, fatigue detection, return‐to‐play classification, and feature‐importance analysis, offering a practical balance between predictive accuracy and model transparency [10, 36].

Graph Neural Networks (GNNs) have attracted growing interest because they provide a natural framework for representing the interconnected structure of the human musculoskeletal system. Unlike scalar or independently processed sensor models, GNNs can explicitly encode the relationships between anatomical segments, joints, and functional movement chains, thereby capturing the interdependence of body regions during athletic activity. This is especially important in injury analysis, where pathological movement patterns often arise not from a single joint in isolation, but from dysfunctional coordination across the kinematic chain. For example, abnormal hip control may influence knee mechanics, which in turn affects ankle loading during landing or cutting tasks. By modeling these relational dependencies, GNNs enable a more physiologically meaningful analysis of movement and may improve the detection of subtle biomechanical risk patterns that conventional approaches fail to represent. Collectively, deep learning, ensemble learning, and graph‐based models form the core analytical foundation of modern AI‐enabled wearable systems in sports medicine [9, 37].

5.2. Emerging AI Paradigms

The field is rapidly evolving beyond established methods toward paradigms that address current limitations:

• Explainable AI (XAI): SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model‐agnostic Explanations) frameworks decompose model predictions into individual feature contributions, enabling clinicians to understand which biomechanical variables drive injury risk scores [11, 24].

• Federated Learning: Enables AI model training across multiple institutions without sharing raw athlete data, addressing privacy concerns while improving model generalizability across diverse populations [28].

• Large Language Models (LLMs): Emerging applications include natural language interfaces for wearable data interpretation, automated clinical report generation, and multimodal data synthesis combining sensor data with electronic health records.

• Digital Twins: Personalized computational models of individual athletes that simulate injury scenarios, optimize training loads, and predict rehabilitation timelines based on real‐time wearable data integration.

• Data Fusion: Multimodal architectures combine biomechanical and physiological signals, including IMU, sEMG, PPG, and GPS data. Individual studies have reported promising discrimination, although differences in target outcomes, populations, and validation methods prevent direct comparison [2, 21].

Table 3 compares existing reviews of wearable technologies and AI in sports medicine.

Table 3.

Comparison with existing reviews on wearable technologies and AI in sports medicine.

Review	Main scope/focus	Type of wearable or digital technology covered	AI coverage level	Ethical, legal, and social implications (ELSI) coverage	Main limitation	Distinct contribution relative to the present review
Preatoni et al. [4]	Musculoskeletal injury assessment using wearable sensors	Primarily wearable biomechanical sensing for injury‐related assessment	Limited—Mainly descriptive discussion of analytics rather than advanced AI pipelines	Absent	Focused mainly on injury biomechanics and sensor‐based assessment, without broad coverage of modern AI frameworks or governance issues	Provides an important foundation in wearable injury monitoring, but does not extend to explainable, federated, or clinically integrated AI systems
Souaifi et al. [24]	AI in sports biomechanics with emphasis on motion analysis	Motion‐analysis systems and sensor‐driven biomechanical evaluation	Moderate—Includes machine learning methods, but limited broader clinical translation	Minimal	Stronger on biomechanics than on full clinical deployment, rehabilitation integration, or ethical governance	Advances discussion of AI‐based sports motion analysis, but remains narrower than a full sensor‐to‐clinic review
Gadiri et al. [21]	Digital technologies in team sports	Broad digital and wearable technologies in team‐sport contexts	Moderate—Systematic treatment of analytics and performance‐related applications	Brief mention	More oriented toward team‐sport technologies in general than toward anatomically structured injury prevention and rehabilitation pathways	Useful for understanding team‐sport digital ecosystems, but less detailed in injury‐specific AI workflows and ELSI depth
Chidambaram et al. [3]	AI‐enhanced sensing and sports performance applications	Sensor systems linked to AI‐enhanced monitoring and performance analysis	Good—Includes deep learning and data‐driven sensing approaches	Absent	Greater emphasis on performance optimization than on injury prevention, rehabilitation monitoring, and responsible deployment	Highlights the promise of AI‐enhanced sensing, but does not provide a dedicated injury‐rehabilitation‐governance framework
Present review	End‐to‐end pipeline from wearable sensing to AI interpretation, clinical decision support, injury prevention, and rehabilitation monitoring	Full spectrum of wearable and digital sensing modalities integrated with clinical and sports medicine applications	Comprehensive—Includes ML, DL, XAI, federated learning, digital twins, and LLM‐oriented perspectives	Dedicated section	None in comparative context; specifically designed to address gaps in prior reviews	Offers an anatomically organized, clinically grounded, and ethically aware synthesis that connects sensors, AI models, decision support, and rehabilitation practice in one unified framework

6. Strengths and Limitations of Current Evidence

6.1. High Predictive Performance Versus Limited Generalizability

Before examining the reported performance metrics, it is important to note that the accuracy, AUC, and prediction rates presented throughout this review are derived from studies that differ considerably in their designs, participant populations, sensor configurations, preprocessing pipelines, and validation methodologies [4, 14, 28]. Direct comparisons between these values should therefore be made with caution, as they may reflect methodological differences rather than genuine differences in model performance. With this caveat in mind, the following observations apply.

Several studies reported promising performance for wearable and AI‐based systems used to identify biomechanical injury‐risk patterns. However, the reported values were obtained from studies with different populations, sensor configurations, outcome definitions, class distributions, and validation methods. They should therefore not be interpreted as directly comparable estimates of clinical injury‐prediction performance [7, 36]. In addition, high accuracy in the classification of biomechanical risk states does not necessarily demonstrate that a system can prospectively predict which athletes will subsequently sustain an injury. Sensitivity, specificity, precision, F1 score, calibration, and class distribution should be considered whenever these measures are available in the original study.

However, these encouraging results should be interpreted with caution. A major limitation across the current literature is that many predictive models have been developed and validated under highly controlled laboratory or semi‐controlled training environments, where sensor placement, task execution, and data quality are carefully standardized [4, 6]. While this improves internal consistency and enhances signal interpretability, it may also inflate performance estimates relative to the complexity of real‐world sport settings, where movement variability, environmental noise, fatigue state, and inconsistent adherence can substantially affect model reliability. In addition, many studies rely on small and relatively homogeneous cohorts, often composed of elite or highly trained athletes, whose movement patterns, physical conditioning, and rehabilitation compliance may differ significantly from those of recreational athletes, youth players, or general clinical populations.

This dependence on restricted cohorts raises important concerns regarding external validity and generalizability. Models trained on optimized biomechanical data from elite performers may learn patterns that do not adequately represent the broader end‐user population, including amateur athletes, older individuals, female athletes in underrepresented sports, or patients undergoing rehabilitation after injury [14, 28]. As a result, a model that performs well in a narrowly defined cohort may fail to maintain the same predictive value when applied in less controlled, more heterogeneous, or clinically complex settings. For wearable injury prediction to achieve meaningful translational impact, future studies must therefore prioritize larger multicenter data sets, more diverse participant populations, and real‐world validation protocols that reflect the variability of actual sports and rehabilitation environments. Only through such broader validation can the field determine whether the high accuracies reported in experimental studies can be translated into reliable and scalable clinical decision‐support tools.

6.2. Model Complexity Versus Clinical Interpretability

Advanced machine learning and deep learning models have shown strong predictive performance when applied to multimodal wearable sensor data, particularly when combining biomechanical, physiological, and contextual inputs into a unified analytical framework [11, 36, 39]. These models can detect subtle, nonlinear relationships that may not be apparent through conventional statistical analysis, allowing them to identify injury‐risk patterns, fatigue states, and rehabilitation‐related changes with considerable sensitivity. Their strength lies in the ability to process large volumes of heterogeneous data simultaneously, integrate signals across multiple body systems, and generate highly accurate classifications or predictions. For this reason, they are increasingly viewed as a promising foundation for next‐generation decision‐support systems in sports medicine. Despite these advantages, limited interpretability remains one of the most important obstacles to their routine clinical adoption. In real‐world sports medicine practice, clinicians, physiotherapists, and performance staff generally require more than a risk score or binary prediction; they need to understand why the model reached a given conclusion and which biomechanical or physiological variables contributed most strongly to that decision. For example, actionable clinical interpretation may depend on knowing whether the predicted injury risk is primarily driven by dynamic knee valgus, reduced knee flexion, asymmetrical muscle activation, excessive cumulative load, or impaired recovery status. Many current black‐box models, however, do not consistently provide this level of explanatory detail, limiting their practical value even when predictive performance is high [24, 28].

This lack of transparency creates a gap between algorithmic accuracy and clinical usability. A model may achieve excellent numerical performance in research settings yet still be difficult to trust or implement if practitioners cannot link its predictions to recognizable biomechanical mechanisms or rehabilitation targets. As a result, limited interpretability can undermine clinician confidence, hinder communication with athletes, and reduce the willingness to rely on AI outputs in high‐stakes decisions, such as training modifications or return‐to‐play clearance. The relatively sparse integration of explainable artificial intelligence (XAI) approaches in the current literature therefore represents a major limitation. Without methods that highlight feature importance, clarify temporal decision patterns, or visualize the biomechanical basis of risk predictions, many AI systems remain insufficiently transparent for widespread clinical translation. Future research should place greater emphasis on interpretable and explainable modeling frameworks so that predictive accuracy is matched by mechanistic insight, practitioner trust, and real‐world clinical relevance.

6.3. Validation Rigor and Statistical Considerations

The reported performance measures should be interpreted cautiously and within the methodological context of each source study. The hybrid IMU‐sEMG study reported 92.3% accuracy, 90.5% recall, and an area under the receiver operating characteristic curve of 0.93 for the classification of biomechanical injury‐risk states. The ankle‐monitoring study reported 87.4% overall detection accuracy, a 12.1% false‐alarm rate, and an average latency of 17.2 ms. These values are not directly comparable because the studies differed in their participant populations, target outcomes, sensor configurations, class distributions, and validation procedures. The reviewed evidence did not support a single independently validated accuracy estimate for prospective hamstring‐strain prediction.

A second critical factor is the validation strategy used during model development and testing. Performance estimates obtained through leave‐one‐out cross‐validation, k‐fold cross‐validation, or evaluation on a fully independent external test set are not directly equivalent and may yield substantially different impressions of model robustness. In general, more isolated and independent validation procedures provide stronger evidence of generalizability, whereas repeated internal resampling methods may overestimate performance, particularly in small data sets or highly standardized experimental settings. Consequently, two studies reporting similar accuracy values may differ substantially in methodological rigor depending on how the models were trained, tuned, and evaluated. Another important limitation is the limited reporting of confidence intervals or other measures of statistical uncertainty around performance metrics. Many studies present single‐point estimates without indicating the variability associated with those values, making it difficult to judge whether reported differences between models are statistically meaningful or potentially due to sampling variability. In clinical and translational research, uncertainty estimation is essential because it provides a clearer indication of reliability and supports more responsible interpretation of predictive claims.

Finally, there is a fundamental distinction between laboratory‐validated performance and field‐validated performance. Models developed under controlled experimental conditions often benefit from optimized sensor placement, standardized movement tasks, and cleaner data acquisition, all of which can enhance apparent predictive success. However, these conditions do not fully reflect the complexity of real sporting or rehabilitation environments, where noise, fatigue, movement variability, inconsistent adherence, and environmental factors may reduce performance. As such, strong laboratory results should not automatically be interpreted as evidence of real‐world clinical readiness. To improve methodological transparency and translational value, future studies should align more closely with established reporting frameworks such as the TRIPOD guidelines for prediction model development and validation. Adherence to these standards would encourage clearer reporting of cohort composition, predictor selection, outcome definitions, validation procedures, uncertainty estimates, and generalizability limitations, thereby strengthening the credibility and clinical relevance of wearable‐AI injury prediction research.

Throughout this review, performance measures are reported as presented by the authors of the original studies. Measures of uncertainty are included only when they were explicitly reported in the source publication. Because the included studies used different designs, populations, outcomes, and validation methods, no pooled performance estimates were calculated. A formal quantitative synthesis would require sufficiently homogeneous studies and was beyond the scope of this narrative review.

6.4. Laboratory Accuracy Versus Ecological Validity

A persistent and important discrepancy remains between laboratory validation and real‐world implementation in wearable sports medicine systems. Although many studies report high levels of predictive accuracy under controlled experimental conditions, these results do not always translate directly to field‐based environments, where data acquisition is inherently less stable and more variable. In practical use, wearable systems are frequently affected by motion artifacts, sensor displacement, soft‐tissue movement, signal dropout, and broader forms of environmental noise which can degrade signal quality and reduce model reliability [9, 33]. These issues are especially relevant during high‐intensity athletic actions, where rapid acceleration, sweat, repeated impacts, or equipment movement may alter sensor position and compromise measurement consistency. As a result, performance metrics obtained in laboratory settings may overestimate the true effectiveness of wearable systems during live training, competition, or unsupervised rehabilitation.

An additional challenge is the substantial methodological heterogeneity across the literature. Studies often differ in their sensor placement protocols, attachment methods, sampling frequencies, preprocessing pipelines, feature extraction strategies, and model validation procedures [4, 14]. Even when targeting the same injury type or movement task, these variations make it difficult to determine whether differences in reported performance reflect genuine model superiority or simply differences in experimental setup. This lack of standardization reduces reproducibility, limits cross‐study comparability and slows the development of transferable clinical guidelines. For wearable monitoring systems to achieve broader clinical and sporting adoption, future research must move toward more standardized acquisition protocols, transparent reporting of signal‐processing workflows, and robust field‐based validation designs that reflect the complexity of real athletic environments. Only then can the field bridge the gap between promising laboratory performance and reliable real‐world deployment.

7. Real‐World Implementation and Evidence

7.1. Implementation Case Studies

NFL Load Management using GPS: GPS‐based monitoring frameworks have been integrated into professional sports training environments to quantify external workload, identify exposure fluctuations, and support individualized load management. These systems may assist medical and performance staff in adjusting training intensity and monitoring post‐injury reintegration. However, the direct effect of GPS‐based workload monitoring on injury and reinjury incidence remains uncertain because available studies use heterogeneous implementation strategies, outcome definitions, and study designs [29, 38]. These applications illustrate the feasibility of integrating wearable‐derived workload measures into routine sports medicine decision‐making. However, stronger prospective evidence is required to determine whether such monitoring directly improves clinical injury outcomes.

Rugby Smart Helmet Programme: In rugby, the implementation of accelerometer‐instrumented smart helmets has provided an important real‐world model for monitoring head impact exposure in collision sports. The World Rugby Smart Helmet initiative has deployed these systems across professional competitions to quantify the magnitude, frequency, and directional characteristics of head impacts sustained during match play and training. These data have enabled the construction of detailed biomechanical impact profiles, allowing researchers and governing bodies to better understand the mechanisms underlying both concussive and subconcussive events. Importantly, the practical value of this program extends beyond passive monitoring. Findings from smart helmet data sets have directly informed rule modifications related to tackle technique, with the goal of reducing dangerous contact scenarios and improving player safety. In addition, the initiative has contributed to the establishment of more evidence‐based thresholds for concussion substitution and sideline assessment protocols, thereby strengthening the medical governance of head injury management in rugby. This program illustrates how wearable monitoring can shape both clinical practice and sport policy when deployed at scale [30, 35].

Tennis Wearable Monitoring: Compared with contact team sports, wearable monitoring in tennis has been explored primarily through single‐case investigations and small‐cohort studies, yet these studies have demonstrated strong feasibility and practical relevance. Wrist‐worn IMU devices have been used to monitor serve mechanics, stroke load, and upper‐limb movement dynamics in professional and high‐performance tennis players [26]. Because tennis involves repetitive, high‐velocity shoulder and elbow movements, wearable monitoring offers an attractive method for detecting technique‐related overload before symptoms progress to clinically significant injury. Real‐time tracking of shoulder rotation velocity, racket‐arm acceleration, and cumulative stroke exposure has enabled more individualized adjustments to training volume and technical execution. These systems may therefore help reduce the risk of overuse injuries, especially in the shoulder, elbow, and wrist, by identifying excessive mechanical stress during repetitive serving and groundstroke activity. Although the evidence base in tennis remains smaller than in sports such as football or rugby, these early applications highlight the versatility of wearable monitoring across different sporting contexts and support its growing value in individualized performance and injury‐risk management.

7.2. Cost‐Effectiveness Considerations

The economic rationale for implementing wearable injury prevention systems is becoming increasingly persuasive, particularly as device costs continue to decline while the financial consequences of sports injuries remain substantial. Current wearable technologies span a broad cost spectrum, ranging from approximately $300 for consumer‐grade GPS units to $5000 for research‐grade multimodal sensor platforms that integrate inertial, physiological, and biomechanical monitoring capabilities. Although this initial investment may appear significant, especially for high‐fidelity systems, the potential financial savings associated with injury prevention and optimized rehabilitation can be considerable. In elite and professional sport, where player availability has direct competitive and economic consequences, the prevention of even a single major injury may offset the cost of an entire monitoring program.

This is particularly evident in the case of severe injuries such as ACL rupture, which may generate total costs of $500,000 to $2,000,000 when accounting not only for direct medical and surgical expenses, but also for rehabilitation, time lost from competition, reduced athletic performance, replacement costs, and broader organizational impact. Within this context, the return on investment (ROI) for wearable systems supported by evidence‐based monitoring and individualized intervention strategies can exceed 50:1, making them highly attractive from a sports medicine and management perspective. Importantly, the value of these systems is not limited to injury prevention alone; they may also reduce reinjury rates, shorten rehabilitation inefficiencies, improve load management, and support more precise return‐to‐play decisions, all of which contribute additional economic benefit.

However, despite this compelling case in professional sport, the broader cost‐effectiveness of wearable monitoring remains less clearly established in amateur, youth, and clinical rehabilitation populations. In these settings, device acquisition and maintenance costs often represent a much larger proportional burden, while the direct economic consequences of injury may be less readily quantified in financial terms. Moreover, access to expert interpretation, infrastructure for data analysis, and long‐term monitoring support may be more limited outside elite performance environments. As a result, although wearable systems show strong economic promise, published cost‐effectiveness analyses in non‐professional settings remain relatively sparse. Future research should therefore evaluate not only technical performance, but also affordability, scalability, and health‐economic impact across broader populations to determine where wearable injury prevention systems can deliver the greatest practical value.

7.2.1. Practical Implementation Barriers

Beyond cost considerations, several practical challenges constrain the routine adoption of wearable‐AI systems in clinical practice. Device interoperability remains a persistent issue, as most manufacturers employ proprietary data formats and communication protocols, limiting seamless integration with electronic health records and complicating data sharing across multidisciplinary care teams [4, 14]. Clinician training represents another critical barrier; the effective use of these systems requires competencies in data interpretation, AI model outputs, and technical troubleshooting that are not routinely included in medical or rehabilitation curricula [6, 27]. Workflow integration poses additional difficulties, as continuous data streams may increase clinical burden, contribute to alert fatigue, and create role ambiguity regarding who is responsible for monitoring, interpreting, and acting upon wearable‐derived information [7, 30, 38]. Finally, athlete acceptance and adherence are essential for sustained monitoring, yet may be undermined by concerns about device comfort, data privacy, perceived surveillance, or feedback that is not actionable [5, 14]. Addressing these barriers will require investment in training infrastructure, development of standardized interoperability frameworks, user‐centered device design, and clear protocols for clinical workflow integration.

7.3. Hardware Readiness and the Field‐Clinic Divide

A recurring but under‐discussed theme in the literature is the distinction between field‐ready and clinic‐confined technologies. This distinction has substantial implications for the translational potential of wearable systems, as it determines whether a technology can be deployed at scale in real‐world sporting environments or remains limited to research and supervised rehabilitation settings.

Field‐Ready Technologies (IMUs, GPS/GNSS, PPG): These systems have reached technical maturity and are routinely deployed in professional and elite sport. Their advantages include ease of use, athlete acceptability, continuous data acquisition, and compatibility with existing training workflows. However, they have limitations: IMUs are susceptible to motion artifacts and sensor drift; GPS is unreliable in indoor or covered environments; and PPG is compromised by motion and skin tone. Despite these limitations, their field readiness has enabled large‐scale and longitudinal monitoring in professional sports, although evidence concerning their direct effect on injury and reinjury incidence remains limited and heterogeneous [29, 38].

Clinic‐Confined Technologies (sEMG, high‐fidelity pressure insoles): These systems provide deeper biomechanical and neuromuscular insight but are limited by technical requirements that preclude routine field deployment. sEMG requires skin preparation, secure electrode attachment, and careful artifact management, making it challenging outside supervised settings [7, 27]. As a result, sEMG applications in the reviewed literature were uniformly conducted in clinical or laboratory environments. While these studies provide essential mechanistic understanding, their translational impact is limited by the gap between controlled research and field implementation.

Bridging the Divide: The future of wearable sports medicine depends on bridging this field‐clinic divide. Emerging technologies, such as dry‐electrode sEMG, textile‐based sensors, and miniaturized high‐fidelity IMUs, may enable more reliable field deployment of currently clinic‐confined modalities. In parallel, hybrid approaches (brief clinic‐based sEMG calibration followed by IMU‐based field monitoring) offer a pragmatic pathway that leverages the strengths of both settings. Researchers should explicitly report the deployment context of their sensors and discuss the implications for real‐world translation.

8. Ethical, Legal, and Social Implications (ELSI)

The widespread adoption of wearable technologies in sports medicine raises profound ethical, legal, and social questions that the field has been relatively slow to address in a systematic and standardized manner. Although wearable systems offer substantial benefits for injury prevention, rehabilitation monitoring, performance optimization, and return‐to‐play decision‐making, they also introduce important concerns related to privacy, autonomy, fairness, data governance, and accountability. These concerns are especially significant because wearable devices collect continuous, highly granular, and often sensitive physiological and biomechanical data that extend beyond traditional clinical assessment. Accordingly, this section examines five critical ELSI domains that must be addressed to ensure that wearable technologies are implemented in a manner that is ethically responsible, legally compliant, and socially acceptable within both professional and non‐professional sports contexts.

8.1. Data Privacy and Surveillance Risks

Wearable systems continuously collect highly detailed physiological, biomechanical, and behavioral data, raising major concerns regarding privacy and surveillance. Unlike conventional medical records, wearable data are often collected in real time and over prolonged periods, enabling the construction of comprehensive profiles of an athlete's health status, fatigue level, injury susceptibility, recovery progression, and overall physical readiness. These data are deeply personal because they may reveal prior injury history, physical vulnerabilities, performance decline, or incomplete recovery, all of which may have consequences beyond clinical care. In competitive sports environments, such information could potentially be exploited by employers, insurers, sponsors, or even opposing teams if appropriate safeguards are not in place. As a result, wearable monitoring may blur the boundary between legitimate health supervision and continuous bodily surveillance [5, 14]. The legal implications of such data collection are also substantial. Regulatory frameworks such as the General Data Protection Regulation (GDPR) in Europe, along with comparable privacy laws in other jurisdictions, require explicit informed consent, data minimization, clear limitation of purpose, secure data handling, and, where applicable, the right of individuals to access or delete their data. However, many commercial wearable platforms do not consistently demonstrate full compliance with these principles, particularly when data are stored in proprietary cloud systems, shared across multiple stakeholders, or reused for secondary analytics without sufficient transparency. Therefore, stronger governance frameworks are needed to clarify who owns athlete data, who may access it, how long it is retained, and under what conditions it may be used. Without such protections, the expansion of wearable technologies in sports medicine risks normalizing intrusive surveillance practices that may undermine athlete autonomy, trust, and data rights.

8.2. Algorithmic Bias and Health Equity

A major concern in applying artificial intelligence to wearable sports medicine is the risk of algorithmic bias. Many existing AI models are developed using data sets predominantly derived from elite male athletes in high‐income countries, where access to advanced wearable technology, sports science support, and standardized monitoring protocols is more readily available. Although such data sets may be convenient for model development, they do not adequately represent the broader populations to whom these systems may eventually be applied. As a result, models trained under these conditions may perform less reliably when used with female athletes, youth populations, athletes from different ethnic or anthropometric backgrounds, or individuals in lower‐resource sporting and clinical environments.

This imbalance is not merely a technical limitation; it raises a broader issue of digital health inequity. If injury prediction systems systematically underperform for underrepresented groups, the consequence may be unequal quality of prevention, rehabilitation guidance, or return‐to‐play assessment. In practical terms, one group of athletes may benefit from early detection of biomechanical risk, while another may receive less accurate or less trustworthy predictions because the model was not trained on data reflecting their physiological and movement characteristics. Such disparities may reinforce existing inequities in access to high‐quality sports medicine and undermine confidence in AI‐enabled care. Addressing this challenge requires a deliberate commitment to dataset diversity, transparent reporting of participant demographics, and prospective validation in underrepresented populations. Future studies should move beyond narrow, convenience‐based cohorts and evaluate whether model performance remains stable across sex, age, ethnicity, level of competition, and clinical context. Without this broader validation, the growing use of wearable‐AI systems risks advancing technological sophistication without ensuring equitable clinical benefit [14, 28].

8.3. Regulatory Pathways

Wearable systems that provide injury‐risk estimates, rehabilitation guidance, or other clinically relevant recommendations operate in a complex, often uncertain regulatory space. In many jurisdictions, these technologies fall into a gray zone between consumer wellness devices and formal medical tools. This ambiguity becomes especially important when wearable systems move beyond passive monitoring and begin generating outputs that could influence diagnosis, treatment planning, or return‐to‐play decisions. At that point, the device may carry responsibilities and risks more closely aligned with regulated medical technologies than with ordinary fitness trackers. Within the European Union, wearable systems that generate clinical recommendations may need to comply with the Medical Device Regulation (MDR 2017/745), particularly when their outputs are intended to support medical judgment or therapeutic action. In addition, the emergence of the AI Act introduces further scrutiny for systems that rely on algorithmic decision‐making in health‐related contexts. In the United States, comparable concerns arise under FDA guidance on Software as a Medical Device (SaMD), which applies to software systems that provide information used for diagnostic or therapeutic purposes. These frameworks highlight that regulatory approval is not merely a technical formality but a critical component of ensuring safety, reliability, accountability, and appropriate clinical use. Despite the importance of these pathways, the current literature offers relatively limited discussion of regulatory compliance, approval strategy, postmarket oversight, and documentation requirements. This represents a significant gap between technological innovation and real‐world deployment. Many studies focus heavily on predictive performance while giving comparatively little attention to whether the resulting systems could realistically satisfy the standards required for use in routine clinical or professional sports settings. Future research should therefore consider regulation as a core translational issue rather than a downstream administrative concern, particularly if wearable‐AI systems are to progress from research prototypes to trusted clinical decision‐support tools [14].

8.4. Athlete Autonomy and Coercive Monitoring

The growing use of wearable monitoring in sport also raises fundamental questions about athlete autonomy and the possibility of coercive participation. In principle, wearable systems are often presented as tools for protecting athlete health, optimizing recovery, and supporting evidence‐based care. In practice, however, the decision to participate in monitoring programs may not always be fully voluntary. Professional and semiprofessional athletes operate within structures shaped by significant power asymmetries, where clubs, federations, coaching staff, medical departments, and national bodies may exert direct or indirect pressure on athletes to accept continuous data collection. Under these conditions, consent can become ethically ambiguous. Athletes may feel that declining wearable monitoring could be interpreted as a lack of commitment, reduced transparency, or unwillingness to cooperate with performance and medical staff. They may also fear that opting out could affect selection decisions, contract opportunities, rehabilitation trust, or social standing within the team environment. As a result, even when monitoring programs are formally described as voluntary, the surrounding institutional culture may make refusal difficult in practice. This undermines the concept of meaningful informed consent, which depends not only on information disclosure but also on genuine freedom to choose without penalty. Ethical frameworks for wearable sports monitoring should therefore explicitly protect the athlete's right to opt out, limit the use of collected data to legitimate health and performance purposes, and establish safeguards against data‐driven discrimination [5]. These protections are necessary to ensure that wearable adoption does not shift from supportive care to surveillance‐based control. Respect for autonomy requires that athletes remain active decision‐makers in how their bodily data are collected, interpreted, and shared, rather than becoming passive sources of performance intelligence for institutions with greater bargaining power.

9. Obstacles, Research Gaps, and Future Directions

9.1. Current Obstacles

Despite the rapid progress of wearable sensing and AI‐driven analytics in sports medicine, several major obstacles continue to limit the reliable translation of these systems into routine clinical and field‐based practice. One of the most persistent challenges is the technical instability of data acquisition in real‐world environments. Wearable sensors are highly vulnerable to motion artifacts, signal noise, calibration drift, sweat‐induced contact degradation, and environmental interference, particularly during high‐intensity or contact‐based activities. These factors can substantially reduce signal fidelity and compromise the reproducibility of biomechanical or physiological measurements, thereby weakening model robustness outside controlled laboratory conditions [9, 14]. In practice, systems that perform well in pilot studies may exhibit reduced reliability when deployed during actual training sessions or competition, where movement complexity and contextual variability are significantly greater.

A second major obstacle lies in the clinical and predictive limitations of current systems. Although many wearable‐AI frameworks can detect fatigue accumulation, biomechanical asymmetries, altered neuromuscular control, or other pre‐injury states, they remain far less successful in predicting acute traumatic injuries such as ligament ruptures or sudden impact‐related events. These injuries often arise from complex, rapidly evolving interactions among intrinsic and extrinsic factors, including opponent behavior, surface conditions, decision‐making, and chance events, which are difficult to capture with physiological and biomechanical monitoring alone [12, 28]. As a result, many current systems are better described as risk‐monitoring tools than true injury‐prediction platforms, limiting their preventive power in highly dynamic sports settings. Another critical barrier is the lack of standardization across wearable research and implementation protocols. There is substantial heterogeneity in sensor placement, attachment methods, sampling frequencies, filtering strategies, feature extraction pipelines, model validation procedures, and reporting metrics. This inconsistency makes it difficult to compare results across studies, reproduce findings, or establish reliable benchmarks for clinical adoption [4, 14]. Without agreed methodological standards, the field remains fragmented, and even promising results may not be generalizable across teams, sports, or institutions.

A further practical challenge concerns athlete adherence and usability. Many proposed systems require multiple sensors positioned across different body regions, which may reduce comfort, restrict movement, or interfere with sport‐specific performance demands. In competitive environments, athletes and coaches may resist using equipment that is intrusive, time‐consuming, or perceived as distracting. Compliance issues are especially relevant in longitudinal monitoring programs, where sustained wearability and routine acceptance are essential for collecting meaningful data sets over time. Thus, the success of wearable injury prevention systems depends not only on technical sophistication but also on ergonomic design, minimal burden, and seamless integration into existing training workflows. In addition, an often‐overlooked obstacle is the gap between algorithm development and clinical decision‐making. Many AI models are optimized for predictive performance but are not designed with clinician usability, interpretability, or workflow compatibility in mind. This creates a translational bottleneck in which technically impressive systems fail to influence practice because coaches, physicians, and rehabilitation specialists cannot easily interpret or operate the outputs. Consequently, adoption barriers are not merely computational but also organizational, behavioral, and clinical.

9.2. Identified Research Gaps

Alongside these implementation obstacles, the literature reveals several important research gaps that must be addressed to advance the field. One of the most prominent is the limited diversity of study populations. A large proportion of existing research has been conducted on elite or semi‐elite male athletes, often in high‐income countries and within a narrow range of team sports [4, 6]. Female athletes, youth athletes, masters' athletes, para‐athletes, and participants from low‐ and middle‐income regions remain markedly underrepresented. This restricts the external validity of current findings and may produce models that are poorly calibrated for broader or more heterogeneous sporting populations. A second major gap is the lack of longitudinal validation. Many studies are cross‐sectional, short‐term, or limited to a single training phase or competitive season. While these studies are useful for proof‐of‐concept development, they do not establish whether wearable‐based monitoring systems can sustain predictive value over time, adapt to evolving athlete conditions, or improve long‐term clinical outcomes [12, 28]. Longitudinal evidence is especially important because injury risk is dynamic and influenced by cumulative training load, recovery patterns, seasonal progression, and repeated exposure to stressors. Without long‐term follow‐up, it remains unclear whether these technologies deliver durable preventive benefit or simply reflect transient associations.

Another substantial gap is the insufficient integration of explainable artificial intelligence (XAI). Although machine learning and deep learning models increasingly achieve high classification or risk‐detection accuracy, many remain opaque “black‐box” systems that provide little insight into how predictions are generated. In clinical sports medicine, this lack of transparency can undermine trust, reduce interpretability, and limit the practical use of model outputs in decision‐making [11, 38]. Clinicians and coaches often need understandable explanations, such as which biomechanical variables contributed most strongly to a risk alert, whether a specific asymmetry drove the result, or how the recommendation aligns with known injury mechanisms. The absence of such explanatory capacity represents a significant translational gap. The literature also demonstrates a persistent underrepresentation of holistic and contextual factors. Most wearable‐based systems focus heavily on biomechanical, physiological, or workload‐derived signals, while systematically excluding psychological readiness, stress, sleep quality, motivation, environmental conditions, travel burden, coaching context, and social influences [9, 33]. Yet injury risk is fundamentally multifactorial. A model that ignores mental fatigue, behavioral adaptation, or contextual stressors may fail to capture the true complexity of athlete vulnerability. This reductionist approach limits both ecological validity and clinical usefulness.

In addition, there is a clear gap in multimodal data fusion. Many current systems rely on a single sensor type, such as inertial measurement units, pressure insoles, heart rate monitors, or GPS devices, or at most combine two modalities [21, 28]. However, injury processes often emerge from interactions between biomechanics, physiological load, neuromuscular fatigue, and contextual exposure. The underdevelopment of robust multimodal fusion architecture means that many systems still provide only partial representations of the athlete's state. More comprehensive integration strategies are needed to move beyond isolated indicators toward holistic, personalized risk modeling. A final research gap concerns clinical outcome validation. Several studies evaluate algorithmic performance using metrics such as accuracy, AUC, or sensitivity, but do not demonstrate whether model‐informed interventions reduce injury incidence, shorten rehabilitation duration, or improve return‐to‐play outcomes. This disconnect between computational success and real clinical benefit remains one of the most significant unanswered questions in the field.

9.3. Future Research Priorities

Future research should move beyond proof‐of‐concept experimentation and prioritize the development of clinically actionable, scalable, and ethically robust systems. A central priority is the creation of explainable AI frameworks that provide clinician‐facing interpretations alongside risk predictions. Rather than merely outputting a probability score, next‐generation systems should indicate which movement patterns, physiological markers, or workload abnormalities contributed to the prediction, thereby improving transparency, practitioner trust, and decision support [11, 24]. XAI will be especially important in multidisciplinary environments where sports scientists, physicians, physiotherapists, and coaches must collaboratively interpret monitoring outputs. Another urgent priority is the establishment of standardized methodologies for wearable sensor deployment, signal preprocessing, feature engineering, validation, and reporting [4, 14]. Consensus‐based standards would improve reproducibility, support meta‐analysis, and accelerate regulatory and clinical acceptance. Standardization should extend not only to hardware configuration and data acquisition, but also to model evaluation, including externally validated benchmarks, clinically meaningful endpoints, and transparent documentation of dataset composition and algorithm limitations.

The field also requires large‐scale, longitudinal, and multi‐site studies involving diverse athlete populations [12, 28]. Such studies are essential for assessing generalizability across sex, age, sport type, competition level, geographic region, and injury mechanism. Multicenter collaboration would also allow the creation of larger and more heterogeneous data sets, which are necessary for training robust models and minimizing bias. Importantly, future studies should evaluate not just predictive performance, but also intervention effectiveness and real‐world impact on injury prevention and rehabilitation outcomes. An especially promising direction is the integration of Digital Twin technology into sports medicine. Digital Twins could provide dynamic virtual representations of individual athletes by continuously updating biomechanical, physiological, and contextual data streams. Such systems may enable personalized simulation of injury risk trajectories, scenario‐based adjustments to training, and individualized rehabilitation planning. In this framework, injury prevention would shift from reactive monitoring to proactive and adaptive risk modeling, potentially transforming the way athletes are assessed and managed across the training‐to‐recovery continuum.

Another major avenue for future work is the development of strong ethical, legal, and regulatory frameworks governing data collection, ownership, consent, model transparency, and accountability [14]. Because wearable systems produce highly granular and sensitive health‐related data, research must address not only technical performance but also fairness, privacy preservation, informed consent, and responsible deployment. Algorithmic decisions that influence selection, training load, rehabilitation access, or return‐to‐play timing must be transparent and auditable to avoid misuse or unintended discrimination. Closely related to this is the exploration of federated learning and privacy‐preserving collaborative AI. Federated learning offers a pathway for cross‐institutional model development without requiring the centralized sharing of raw athlete data, thereby reducing privacy risks while supporting larger, more diverse training data sets. This may be particularly valuable in elite sport, where data confidentiality is a major concern and institutions may be reluctant to share sensitive performance and health records.

Future research should also prioritize multimodal, context‐aware architectures that integrate biomechanical, physiological, psychological, and environmental data into unified frameworks. Rather than focusing narrowly on isolated injury correlates, these models should aim to capture the full ecosystem of injury risk. In parallel, user‐centered design principles should guide system development to ensure that devices are comfortable, minimally intrusive, and practically compatible with sport‐specific demands. Ultimately, the next phase of progress in wearable‐AI sports medicine will depend on a shift from demonstrating technical novelty to proving clinical value, implementation feasibility, and ethical legitimacy. The most impactful systems will likely be those that are not only accurate, but also interpretable, standardized, equitable, athlete‐centered, and validated in real‐world practice. Only through this broader translational perspective can wearable technologies fulfill their potential to support injury prevention, optimize rehabilitation, and improve athlete health at scale.

10. Discussion

The present review highlights a clear shift in sports medicine from episodic and subjective assessment toward continuous, sensor‐driven monitoring supported by artificial intelligence. Across the included studies, wearable technologies demonstrated promising utility for both injury prevention and rehabilitation, particularly when biomechanical and physiological signals were integrated into data‐driven decision frameworks. The evidence synthesized in this review suggests that inertial measurement units, GPS/GNSS systems, surface electromyography, and photoplethysmography now form the technological core of this transformation, while AI models increasingly serve as the analytical layer that converts raw signals into clinically actionable risk indicators. Taken together, these developments indicate that wearable‐AI systems are no longer merely performance‐enhancement tools but are becoming serious candidates for routine use in injury surveillance, return‐to‐play decision support, and individualized training prescription.

A major finding of this review is that predictive performance aligns with the specificity of the sensor‐injury mechanism relationship. Lower‐limb injuries as ACL, ankle sprain, and hamstring strain, demonstrate the strongest reported performance because their biomechanical precursors are well‐characterized and directly measurable by IMU and sEMG sensors. However, high classification accuracy does not necessarily translate into reliable prospective injury prediction, particularly when outcome prevalence, class distribution, and false‐positive rates are not reported consistently. This creates a clinical paradox: the sensors are excellent at detecting abnormal biomechanics, but translating these detections into reliable injury predictions remains challenging because many athletes exhibiting abnormal movement patterns do not subsequently sustain injury. The practical implication is that lower‐limb wearable systems may be most valuable not as binary injury predictors, but as continuous monitoring tools that detect deviations from an athlete's baseline movement patterns. This shift in framing, from “injury prediction” to “deviation detection,” reduces the burden of false positives and aligns with rehabilitation monitoring, where the goal is to track restoration of symmetrical function rather than to predict a specific injury event. In this context, the lower‐limb evidence supports a monitoring paradigm that prioritizes trend analysis, individualized baseline comparisons, and progressive loading guidance over binary risk classification. These examples illustrate an important principle: wearable systems are most effective when they are not used as generic monitoring devices, but as mechanism‐sensitive tools aligned with the specific biomechanics of the target injury. In practical terms, this means that future system design should prioritize injury‐specific signal selection and athlete‐specific model calibration rather than one‐size‐fits‐all architectures.

Another important observation is that the most promising clinical performance emerges from multimodal integration rather than isolated sensor streams. Single‐sensor approaches remain useful because they are more affordable, easier to implement, and often better tolerated by athletes. However, the strongest discriminative performance in the reviewed literature was observed with multimodal fusion, in which internal load signals, such as sEMG or physiological recovery markers, are combined with external load measures, such as IMU‐ or GPS‐derived movement data. This is clinically meaningful because many sports injuries do not result from a single abnormal variable, but from the interaction between cumulative workload, altered mechanics, neuromuscular fatigue, and insufficient recovery. Findings from individual studies suggest that multimodal integration may improve discrimination between normal and potentially hazardous movement states. However, differences in target outcomes, populations, and validation procedures prevent direct comparison of model performance. For rehabilitation, this also supports the use of integrated monitoring frameworks that interpret joint symmetry, muscle activation timing, and cardiovascular recovery together before progression to higher training loads or return to competition.

Despite these encouraging results, the review also shows that high reported accuracy should be interpreted cautiously. Many of the strongest findings were obtained in controlled laboratory settings, under supervised sensor placement conditions, and in relatively small or homogeneous cohorts. This introduces a substantial translational gap between algorithmic success and real‐world clinical robustness. Models trained on elite athletes performing standardized tasks may not generalize well to amateur athletes, youth populations, female athletes, or individuals undergoing rehabilitation after injury. Moreover, injury data sets are frequently imbalanced because true injury events are rare, which can inflate conventional accuracy values if not accompanied by more informative measures such as sensitivity, specificity, AUC, calibration, and uncertainty estimates. The distinction between detecting abnormal biomechanics in a laboratory and prospectively preventing injury in competition must therefore remain central to interpretation. From a methodological perspective, future studies need more rigorous external validation, clearer reporting of cohort composition, better handling of class imbalance, and stronger prospective study designs that evaluate whether model outputs improve clinical outcomes.

Clinical adoption also depends on interpretability, not only prediction. In sports medicine practice, clinicians and performance staff must understand why a model is flagging an athlete as high risk. A risk score without biomechanical explanation has limited practical value, especially in high‐stakes return‐to‐play decisions. This explains why the limited integration of explainable AI remains a major weakness in the field. Systems that can identify the relative importance of variables such as knee abduction angle, landing asymmetry, eccentric hamstring activation, or cumulative sprint load are more likely to support intervention planning than opaque models that provide only a binary alert. Explainability is therefore not simply a technical preference; it is a prerequisite for clinical trust, interdisciplinary communication, and ethical accountability. Future wearable‐AI systems should move beyond black‐box prediction toward hybrid decision‐support tools that preserve strong performance while also offering interpretable, clinician‐facing biomechanical reasoning.

The review further suggests that the practical value of wearable monitoring may be highest not in predicting isolated acute traumatic events, but in identifying progressive pre‐injury states and guiding rehabilitation load management. Acute injuries often involve context‐dependent and unpredictable interactions that may not be fully captured by wearable data alone. By contrast, overuse conditions, fatigue‐related breakdown, asymmetrical recovery patterns, and cumulative loading errors are more amenable to continuous tracking. This may explain why current systems appear particularly useful for monitoring ACL rehabilitation, hamstring fatigue, tibial stress loading, and workload‐related reinjury risk. In this sense, the role of wearables may be less about replacing clinicians and more about extending clinical observation into daily training environments, where early deviations can be detected before they become clinically obvious. Current evidence suggests that the value of wearables may lie primarily in ongoing load governance and individualized adjustment rather than in automated injury prediction. However, prospective studies are still required to determine whether wearable‐guided interventions directly reduce injury or reinjury incidence.

At the same time, wider deployment raises unresolved ethical, legal, and social concerns. Wearable systems collect highly granular data that may reveal not only performance capacity, but also fatigue status, injury vulnerability, and rehabilitation trajectory. In competitive environments, such information can influence selection, contract decisions, and perceived athlete reliability. As a result, data ownership, informed consent, opt‐out rights, and safeguards against coercive monitoring must be treated as central design requirements rather than secondary policy issues. Similar concerns apply to algorithmic bias. If models are trained predominantly on elite male athletes from specific sports or regions, then their predictions may systematically underperform in underrepresented populations, thereby reproducing inequities in care and decision‐making. Regulatory uncertainty adds a further challenge, particularly when wearable systems move from descriptive monitoring toward clinical recommendations. For these reasons, successful translation will require not only better engineering but also governance frameworks that address privacy, fairness, accountability, and the legal status of AI‐assisted recommendations.

From a future‐research perspective, the field now needs a transition from proof‐of‐concept studies to robust implementation science. Several priorities emerge from this review. First, longitudinal multisite cohorts are needed to evaluate whether wearable‐AI systems remain reliable across seasons, contexts, and athlete populations. Second, standardization of sensor placement, sampling protocols, feature definitions, and validation reporting would greatly improve reproducibility and cross‐study synthesis. Third, explainable AI, federated learning, and digital twin approaches deserve further development because they directly address current barriers related to interpretability, privacy, and personalization. Fourth, broader biopsychosocial integration is needed. Most current systems focus heavily on biomechanical and physiological inputs while underrepresenting psychological stress, sleep disruption, training context, and environmental exposures, all of which shape injury risk in the real world. A more complete injury model will likely require combining biomechanical sensing with contextual and behavioral data rather than relying on movement signals alone.

Overall, the literature reviewed here supports a cautiously optimistic conclusion. Wearable‐AI systems have already demonstrated meaningful clinical promise, particularly for mechanism‐specific monitoring, rehabilitation surveillance, and multimodal risk stratification. However, the field remains in a transitional stage. The strongest current evidence shows what these systems can do under favorable conditions, but not yet what they can reliably deliver across diverse sports, settings, and populations at scale. The next phase of progress will depend on whether researchers and practitioners can move beyond isolated accuracy reports toward interpretable, standardized, ethically governed, and prospectively validated systems that improve real clinical decisions. Only then will wearable technologies fulfill their potential to transform sports medicine from reactive injury management to proactive, individualized prevention and rehabilitation.

11. Conclusions

This review has synthesized 28 studies to map the current landscape of wearable monitoring systems for sports injury prevention and rehabilitation. The convergence of multimodal sensor technologies, IMUs, GPS/GNSS, sEMG, and PPG, with advanced AI architectures has enabled more accurate detection, monitoring, and prevention of injuries across a wide range of sports and anatomical regions.

The principal findings are threefold. First, wearable systems show promising performance in detecting biomechanical risk patterns and monitoring rehabilitation‐related changes. A hybrid IMU‐sEMG framework reported 92.3% accuracy, 90.5% recall, and an area under the receiver operating characteristic curve of 0.93 for classifying biomechanical injury‐risk states, while an ankle‐monitoring framework reported 87.4% overall detection accuracy, a 12.1% false‐alarm rate, and an average latency of 17.2 ms. These findings should not be interpreted as evidence of prospective prediction of confirmed injuries. Second, wearable systems are increasingly used to support workload monitoring, rehabilitation progression, and return‐to‐play decision‐making, although evidence concerning their direct effect on injury and reinjury incidence remains limited and heterogeneous. Third, the field is progressing from established machine‐learning and deep‐learning methods toward explainable AI, federated learning, digital twins, and other approaches intended to improve interpretability, privacy, and personalization.

Despite these advances, significant challenges remain: limited population diversity, insufficient longitudinal validation, model interpretability deficits, and unresolved ethical‐legal‐social implications regarding data privacy, algorithmic bias, and athlete autonomy. Addressing these challenges requires interdisciplinary collaboration spanning sports science, clinical medicine, computer science, bioethics, and regulatory affairs. The ultimate goal, maximizing athletic longevity and performance through evidence‐based, ethically governed wearable‐AI systems, demands nothing less.

Author Contributions

Margarida Silva: conceptualization, investigation, formal analysis, data curation, writing – original draft, writing – review and editing, methodology. Saif Al‐jumaili: conceptualization, methodology, formal analysis, data curation, writing – original draft, writing – review and editing. Mário Lopes: conceptualization, methodology, data curation, formal analysis, writing – original draft, writing – review and editing. Carlos Albuquerque: conceptualization, methodology, formal analysis, data curation, writing – original draft, writing – review and editing, funding acquisition. António D. Reis: conceptualization, methodology, writing – original draft, writing – review and editing, data curation, formal analysis. Paulo Jorge Coelho: conceptualization, methodology, funding acquisition, writing – original draft, writing – review and editing, data curation. Ivan Miguel Pires: supervision, conceptualization, methodology, data curation, validation, writing – original draft, writing – review and editing, funding acquisition. All authors have read and approved the final version of the manuscript.

Ethics Statement

The authors have nothing to report.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Transparency Statement

Ivan Miguel Pires, as manuscript guarantor, affirms that this manuscript is an honest, accurate, and transparent account of the review being reported; that no important aspects of the review have been omitted; and that any discrepancies from the review as planned have been explained.

Data Availability Statement

Data sharing not applicable to this article as no data sets were generated or analyzed during the current study.