Application of Artificial Intelligence in Physical Rehabilitation of Patients Admitted to the Intensive Care Unit: A Scoping Review

Abstract

Background and aims

Artificial intelligence (AI) has proven to be a highly useful tool in the clinical setting, especially in the Intensive Care Unit (ICU). The use of various AI-mediated instruments to guide medical treatments and even support surgical procedures has been previously described, but there is still no aggregated evidence on its usefulness in assisting the physical rehabilitation process of critically ill patients, understanding that this is extremely important to prevent the development of muscle weakness in the ICU. This review, therefore, aimed to describe the usefulness of AI in supporting the physical rehabilitation of patients admitted to the ICU.

Materials and methods

This scoping review was conducted following the Joanna Briggs Institute (JBI) methodology, originally developed by Arksey and O'Malley. A structured search strategy based on a Population, Concept, and Context (PCC) framework was used to search PubMed, Web of Science, Scopus, and the Virtual Health Library (VHL) databases.

Results

The initial search yielded 116 articles. After removing duplicates and applying exclusion criteria during title and abstract screening, eight studies were included in the final analysis. Identified tools included noninvasive mobility sensors, robotic assistance systems, machine learning algorithms, and software to support musculoskeletal ultrasound assessment.

Conclusion

Artificial intelligence is emerging as a key tool for ICU rehabilitation, offering objective data, enhancing patient monitoring, and streamlining assessment processes.

How to cite this article

Payán-Salcedo HA, Castro Aguilera AM, Salinas Batioja MF, Castillo Diaz LM. Application of Artificial Intelligence in Physical Rehabilitation of Patients Admitted to the Intensive Care Unit: A Scoping Review. Indian J Crit Care Med 2025;29(10):851–860.

Highlights

This scoping review analyzes and describes the usefulness of tools powered by artificial intelligence (AI) to assist with physical rehabilitation in the Intensive Care Unit (ICU), a process fundamental to improving patient outcomes. Our results show promising tools that can improve the efficiency of physical assessment and interventions for critically ill patients.

Introduction

Artificial intelligence is defined as a complex and constantly evolving structure based on the use of algorithms and software designed to emulate human cognitive functions, such as reasoning, learning, and problem-solving.¹ Artificial intelligence and healthcare have followed a convergent path in recent years, with growing evidence suggesting that technological and digital advances facilitate clinical decision-making and, in specific cases, guide patient assessment, treatment, and follow-up through process automation, strengthening and improving clinical practice.^2,3

One of the scenarios where these tools are particularly useful is the ICU spaces where highly qualified professionals and cutting-edge technology are required to obtain accurate and reliable data, enabling the implementation of advanced strategies to administer effective and safe therapies to patients.^4–7 In this context, AI enables real-time multiparametric monitoring of critical patients, facilitating the individualization and precision of treatment plans, becoming increasingly integrated into the clinical practice of the professionals who work there.⁸

One of the challenges faced by ICU clinicians is a lack of patient mobilization, which can lead to physical deconditioning and muscle atrophy.⁹ In young adults, 14 days of immobilization has been associated with a 5–9% loss of quadriceps muscle mass and a 20–27% decrease in contractile strength. This process is further accelerated in older adults, whose rate of strength and muscle mass loss is 3–6 times higher.¹⁰ In patients on mechanical ventilation (MV), muscle mass may decrease by up to 12.5% during the first week in the ICU stay. Of particular concern, more than one-third of patients who require MV for at least 5 days may develop ICU-acquired weakness (ICU-AW).¹⁰ This condition may delay ventilator weaning, prolong hospital stay, and increase morbidity and mortality.¹¹

In response to the previously described problem, physical rehabilitation has emerged as a type of intervention that has demonstrated significant benefits for bedridden patients.¹¹ Evidence shows that it facilitates weaning from MV, reduces the use of sedative medications, shortens ICU stay, and lowers mortality.¹² This process includes active and passive mobilization; progressive physical exercise with or without resistance; respiratory muscle training; positional transitions; weight-bearing support, among other interventions, which may be guided or facilitated by AI tools.^13–15

Considering the above, traditional rehabilitation methods rely heavily on manual techniques, subjective clinical scales, and visual assessments, which can be time-consuming, labor-intensive, and prone to interobserver variability.^3,11 The integration of AI tools can overcome these limitations by providing objective, reproducible, and real-time data to optimize patient mobilization and rehabilitation strategies.^16,17 Although its inclusion in different clinical settings is described as an emerging need, consolidated and up-to-date information specifically addressing its applicability to the physical rehabilitation process in the ICU is lacking. Therefore, the objective of this review was to describe the usefulness of AI in supporting the physical rehabilitation of patients admitted to the ICU.

Materials and Methods

Protocol

This scoping review was conducted following the methodology of the Joanna Briggs Institute (JBI) originally developed by Arksey and O'Malley, in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) checklist.^18–20 To map the literature, this guideline recommends: establishing clear inclusion and exclusion criteria; providing detailed characteristics of the participants; defining a concept to guide the scope and breadth of the review; and outlining a context that delimits the factors involved in the research.

Eligibility Criteria

To guide this review, a Population, Concept, and Context (PCC) question (Fig. 1) was formulated, including three elements: Population (P), Concept (C), and Context (C). Regarding the Population (P), studies involving adults and pediatric patients admitted to the ICU were included.¹⁸ Based on the Concept (C), studies that described AI strategies or tools, such as software, neural networks, assistive robots, machine learning methods, and algorithms, were considered. For the Context (C), we included studies in which AI tools were applied to the physical rehabilitation process of patients admitted to the ICU. In terms of source types, experimental and quasi-experimental study designs were considered, including randomized controlled trials, nonrandomized controlled trials, descriptive or analytical observational studies, and case reports or case series.

Fig. 1.

Inclusion criteria based on the PCC question

As exclusion criteria, we excluded letters to the editor, review articles, gray literature documents, and studies in which AI tools were applied to processes unrelated to physical rehabilitation.

Data Sources

In accordance with JBI guidelines, a systematic and structured literature search was conducted in the databases of the Virtual Health Library (VHL), PubMed, Web of Science (Clarivate Analytics), and Scopus (Elsevier) in June 2024.¹⁸

Search Strategy

The search strategy aimed to identify studies in English and Spanish published without date restrictions. For this review, searches were conducted in the selected databases using Medical Subject Headings (MeSH) and Health Sciences Descriptors (DeCS), combined with Boolean operators, and structured according to the PCC framework. The goal was to retrieve articles describing the use of AI to support physical rehabilitation in adult or pediatric patients admitted to the ICU.

Based on the PCC framework and using MeSH and DeCS terms with appropriate Boolean operators, the search strategy was applied in English across the selected databases (Supplementary Table 1).

Table 1.

Characteristics of the studies included

Author	Type of study/ country	Sample	Type of patients	Age in years (SD)–(interquartile ranges)	Target
Jensen et al.23	Quasi-experimental/Germany	14Healthy subjects: 12 Patients: 2	Healthy subjects and ICU patients	31 (4)	Present and compare two metrics (Robotic algorithm and surface EMG) to monitor active patient participation during robotic rehabilitation.
Huebner et al.24	Prospective cohort/Germany	29IG: 16 patients CG:13 patients	Adults presurgery for lung transplantation	IG: 61 (55–63)CG: 58 (49–63)	To explore the effects of robot-assisted mobilization and verticalization in intensive care.
Warmbein et al.25	Cross-sectional observational/Germany	16 patients F:8M:8	Adults presurgery for lung transplantation	58 (8.4)	To evaluate the feasibility of using a VEMOTION robotic system to support early mobilization of ICU patients, observing its safety and acceptance in clinical practice.
Reiter et al.26	Quasi-experimental/E.E.U.U	8 patients	Adults in a surgical ICU	NR	Assess ICU patient mobility through noninvasive sensors and compare the results with traditional mobility assessment.
Ma et al.27	Cross-sectional observational/E.E.U.U	8 patients	Adults in a surgical ICU	67 (52–77)	To compare the level of mobility of patients in an ICU obtained through motion sensors and a mobility scale.
Yeung et al.14	Cross-sectional observational/E.E.U. U	NR	Adult ICU patients	NR	Develop and validate computer vision algorithms to detect patient mobilization activities in the adult ICU, as well as to measure the duration of such activities and the number of personnel involved.
Huy Nhat et al.28	Prospective cohort/Vietnam	20 patients	Adults with severe tetanus	Group AI: 67 (13) Group without AI: 56 (17)	Develop and evaluate an AI tool that performs automated recognition and measurement of RFCSA to assist non-expert operators in their measurement by muscle ultrasound.
Fuest et al.29	Longitudinal observational/Alemania	Seriously ill and frail: 188 Middle-aged: 331 Young trauma: 108Old non-frail: 321	Adults in ICU	Seriously ill and frail: 78 (73–82) Middle-aged: 59 (54–63) Young trauma: 34 (27–40) Old non-frail: 75 (71–80)	Demonstrate that AI-based patient grouping can provide individualized mobility recommendations that increase the opportunity to return home.

AI, artificial intelligence; CG, control group; EMG, electromyography; F, female; ICU, intensive care unit; IG, intervention group; M, male; NR, does not report; RFCSA, transverse area of the rectus femoris; SD, standard deviation

The reference lists of all included studies were carefully reviewed to identify additional studies that met the inclusion criteria.

Study Selection/Source of Evidence

After a comprehensive search, all identified citations were compiled using Mendeley software (Elsevier), and duplicate records were removed. Two independent reviewers screened the titles and abstracts to assess eligibility based on the predefined inclusion criteria. Subsequently, both reviewers independently assessed the full-text articles to confirm whether they met the inclusion criteria. Any reason for study exclusion was documented in the review. In cases of disagreement between reviewers during the selection process, consensus was reached through discussion or, if necessary, with the involvement of a third reviewer. The final search results were presented using a flow diagram in accordance with the PRISMA-ScR guidelines (Fig. 2).²⁰ This process was carried out to ensure the integrity of the evidence included in the review.

Fig. 2.

Study selection flowchart (PRISMA methodology)

Data Extraction and Synthesis

Data were extracted from the included articles by two independent reviewers, following a meticulous process to minimize the risk of losing relevant information. The extracted data were compiled in a Microsoft Excel spreadsheet and are presented in descriptive tables, which include specific variables of interest, such as study type, country, year of publication, sample size, patient population, study design, AI tool employed, purpose or function of the tool, and the healthcare professional responsible for its use.

Analysis and Assessment of Methodological Quality

The Newcastle-Ottawa Scale (NOS) was used to assess the methodological quality of descriptive observational and cohort studies (adapted version), as recommended by the Cochrane Collaboration. The NOS comprises eight items grouped into three domains, with a total score ranging from 0 to 9 points. Accordingly, the studies were classified as low quality (0–2 points), moderate quality (3–5 points), or good/high quality (6–9 points).²¹ Quasi-experimental studies were evaluated using the JBI critical appraisal checklist.²² This checklist consists of nine items assessing aspects, such as intervention and outcome variables, participant selection, control group presence, pre- and postintervention measurements, participant follow-up, outcome measures, and statistical analysis. Discrepancies between reviewers during the quality assessment process were resolved through discussion. If consensus could not be reached, a third reviewer was consulted.

Results

The initial search yielded 116 studies. Fifteen duplicate studies were excluded, and 89 studies were excluded after screening their titles and abstracts. An additional four studies were excluded for not meeting the inclusion criteria. Ultimately, eight studies were included in the qualitative analysis (Fig. 2). Of these, 50% were published in Germany, 37.5% in the United States, and 12.5% in Vietnam.

Among these studies, 50% were descriptive observational, 25% were prospective cohort studies, and 25% were quasi-experimental. The studies were published between 2016 and 2024, and included patients aged 31–78 years. Half of the studies were conducted in surgical ICUs (Table 1). In 37.5% of the studies, the VEMOTION® Robotic Assistance System (Reactive Robotics, Munich) was used as the AI tool to support rehabilitation in the ICU.^23–25 Additional technologies described included noninvasive mobility sensors to assess the degree and level of patients' active mobility, software to facilitate and accelerate ultrasound evaluation of the cross-sectional area of the rectus femoris (RF-CSA), and machine learning algorithms to classify and group patients and evaluate the effectiveness of physical rehabilitation interventions (Fig. 3). In 50% of the studies, the AI tool was operated by nursing staff, whereas in 37.5% it was used by physiotherapists (Table 2).^14,23–29

Fig. 3.

Artificial intelligence tools to assist physical rehabilitation in the ICU

Table 2.

Variables of interest and results of the included studies

Author	Tool for AI	Duration of the intervention	Adverse events	Utility of the tool	Professional in- charge	Results	Main findings/conclusion
Jensen et al.23	Robotic assistance system VEMOTION® (Reactive Robotics, Munich) Energy-based control algorithm “assist-as-needed” (AAN)	Healthy subjects: About 2 hours per session Patients: 20 minutes. 4 interventions of 1 minute (8–16 steps per leg) with 2–3 minutes rest between each intervention	NR	Quantify the work performed by the patient during robot-assisted lower extremity rehabilitation to determine the patient's active muscle participation in therapy	Physiotherapists	Robot-based metrics and EMG-based metrics showed a high correlation (R2 > 0.80) for quantifying active patient participation in mobilization	The VEMOTION® robot-based metric is a reliable, noninvasive indicator of active patient participation during ICU rehabilitation
Huebner et al.24	Robotic assistance system VEMOTION® (Reactive Robotics, Munich)	IG: robot-assisted verticalization and mobility 2 times daily, 20 minutesCG: conventional mobilization	NR	Achieve verticalization of up to 70° and assist cyclic stepping movements of the patient	Physiotherapists and nurses	There were no statistically significant differences in duration of mechanical ventilation (p = 0.181), ICU stay (p = 0.187), muscle parameters assessed by ultrasound, and quality of life assessed at 3 months between IG and CG	Robot-assisted mobilization is feasible and safe, but does not demonstrate clinical superiority over conventional mobilization
Warmbein et al.25	Robotic assistance system VEMOTION® (Reactive Robotics, Munich)	20 minutes per session, twice a day, for 10 sessions. Assisted gait from a verticalization of 14–54°	Pain	Assist in the early mobilization of critically ill patients; however, it requires trained personnel and adjustments in the process	Nurses	Nurse mobilizers rated the intervention as feasible (mean of 5.3 ± 1.6) and their physical stress as low (mean of 2.0 ± 1.3) according to a Likert scale	Robot-assisted mobilization is well tolerated, although it requires specialized human resources and can cause discomfort.
Reiter et al.26	NIMS: NonInvasive Mobility Sensor	NR	NR	Detect through sensors the postures and mobility of the patient in the ICU.	NR	A weighted Kappa of 0.64 (CI:95%) was obtained when comparing the results of the two methods for assessing mobility (NIMS-Traditional mobility assessment)	The sensor allowed mobility to be quantified and protocols to be standardized, overcoming the subjectivity of clinical scales
Ma et al.27	Motion sensors Microsoft Kinect (Microsoft, E.E.U.U)	NR	NR	Perform continuous, novel, feasible, and automated assessment of ICU patient mobility	Doctors	A concordance of 0.86 (95% CI, 0.72–1.0) was obtained according to the weighted Kappa index, between the two tools for quantifying mobility	The sensor has high accuracy in detecting posture and movement changes, providing objective and continuous monitoring of the mobility of critically ill patients
Yeung et al.14	Computer vision algorithm based on deep neural networks, ResNet y YOLOV2	NR	NR	Automatically detects patient mobility activities and provides valuable data to help clinicians improve and standardize mobilization protocols in the ICU	Nurses	The algorithm achieved an average specificity of 89.2%, a sensitivity of 87.2% and an average accuracy of 68.8% in the measurement of ICU mobility activities	AI algorithms allow for analyzing workflows (number of mobilization activities, duration, and number of professionals involved) and standardizing clinical practice
Huy Nhat et al.28	Software RAIMUS	3 measurements on each leg, 1 at ICU admission, another at day 7, and another at ICU discharge (26 ± 11 days)	NR	Eliminate the need for manual tracing, increase reproducibility, and save time when performing muscle ultrasound in the ICU	Doctors and nurses	Time spent on scans was significantly reduced from a median of 19.6 min (RIC 16.9–21.7) to 9.4 min (RIC 7.2–11.7) compared to using the AI tool (p < 0.001)	AI-assisted ultrasound improves efficiency and objectivity in the assessment of muscle loss in the ICU
Fuest et al.29	Machine learning model K-Means	Average duration of mobilization per day: Seriously ill and frail: 6 (2–27) Middle-aged: 13 (5–39) Young trauma: 20 (7–51) Old non-frail: 28 (9–67)	NR	Identify and group a set of patients according to their shared characteristics, to determine the effects of an intervention on these groups	Physiotherapists and ICU professionals	Early mobilization (<72 h) was the most significant factor for discharge in the young trauma group (OR: 10.0 [2.8–44.0], p < 0.001) and in the middle-aged group (OR: 3.0 (95% CI [1.5–6.0]), p < 0.001)	AI allows patients to be stratified according to their mobility tolerance, helping to individualize protocols and optimizing safety and efficacy

AI, artificial intelligence; CG, control group; CI, confidence interval; EMG, electromyography; F, female; ICU, intensive care unit; IG, intervention group; M, male; NR, does not report; OR, odds ratio; RFCSA, transverse area of the rectus femoris

Evaluation of Methodological Quality

The cohort studies and the four observational studies were rated as having good methodological quality, with average NOS scores of 8/9 for the cohort studies and 7/9 for the observational studies. The two nonrandomized quasi-experimental studies assessed using the JBI critical appraisal tool received scores of 6/9 and 8/9. According to the research team, these results indicate a low risk of bias. The methodological quality assessment of all included studies is presented in Supplementary Tables 2 to 4.

Discussion

The objective of this review was to describe the usefulness of AI in supporting the physical rehabilitation of patients admitted to the ICU. The included studies described four main categories of AI tools applied to physical rehabilitation in the ICU. Robot-assisted systems, such as VEMOTION®, enable uprighting and gait assistance, decreasing the physical burden on staff and promoting early mobilization.^24,25 Noninvasive mobility sensors continuously monitor patient posture and movement, providing quantitative data that exceeds the accuracy of conventional, usually subjective, clinical scales.^26,27 On the other hand, deep learning-based computer vision algorithms can automatically detect mobilization activities, quantifying variables such as the duration and number of staff involved, thus providing new insights into ICU workflow.¹⁴ Finally, RAIMUS software applied to ultrasound facilitates the automatic segmentation and measurement of the rectus femoris muscle, optimizing the time required per scan and improving the reproducibility of measurements.^28,29 Together, these tools highlight the diverse potential that AI offers to improve rehabilitation processes in the ICU.

The evidence suggests that AI has been implemented across various aspects of ICU rehabilitation, with its applicability focused on accelerating recovery while ensuring patient safety. Huebner et al.²⁴ and Warmbein et al.²⁵ emphasized the role of the VEMOTION® robotic system in facilitating early mobilization through verticalization and gait training. However, they did not report significant differences in hospital stay or MV duration when compared with conventional methods. Similar findings have been reported outside the ICU setting in patients with spinal cord injuries, where robotic exoskeletons have improved mobility. However, their long-term impact on functional recovery remains under discussion.^30–33

Other studies in our review, such as those by Reiter et al.²⁶ and Ma et al.,²⁷ demonstrated that noninvasive mobility sensors accurately detected motor activity in patients admitted to the ICU, strongly agreeing with traditional assessment scales. These tools facilitate continuous mobility evaluation and provide valuable data for standardizing early mobilization processes and protocols. In line with these findings, and although evidence on the use of such tools outside the ICU setting is limited, it has been shown that devices equipped with motion sensor technology in hospitalized patients enable accurate, real-time monitoring of their movements. These tools help optimize therapeutic interventions and contribute to more effective recovery by providing critical information on vital signs and the extent of postoperative mobility. As such, they are highly relevant for fall risk prevention and for identifying opportunities to initiate early rehabilitation, a role very similar to their utility within the ICU setting.^34–38 Although the use of these sensors continues to grow, authors like Sena et al.³⁹ have reported that they may present limitations in patients with reduced baseline mobility, suggesting the need to combine them with other assessment methods.

Another type of AI tool identified in our review was the RAIMUS software, an ultrasound-based evaluation program designed to automatically detect and measure the RF-CSA, markedly reducing scanning time and improving measurement reproducibility.²⁸ The evaluation of this variable has been recognized as a reliable method to support the diagnosis of ICU-AW, inform prognosis, and monitor musculoskeletal function in patients with critical illness.^40,41 This enables the identification of optimal timing for early and timely intervention, which is an essential aspect in complex patient settings where frequent muscle assessment plays a decisive role in clinical outcomes. However, studies like that by Katakis et al.⁴² conclude that automation of this process cannot fully replace clinical interpretation, as segmentation errors may lead to inappropriate therapeutic decisions.

The results of our review also highlight the use of technologies like machine learning algorithms, which have been employed to classify patients based on their response to rehabilitation and to optimize early mobilization strategies.²⁹ However, as noted in the literature, technical factors, such as the quality of the data used to train these algorithms, must be carefully considered, as they are critical to algorithm performance. Inadequate data quality can introduce bias in patient stratification.⁴³

Moreover, one of the key findings of this review is the observed differences among healthcare professionals in the use of AI tools during the rehabilitation process. Research indicates that physical therapists have been the primary users of AI tools in physical rehabilitation, whereas nursing staff have played a more prominent role in mobility monitoring and device management.²⁵ This distribution of roles aligns with studies that emphasize the need for a multidisciplinary approach to ICU rehabilitation to maximize the benefits of AI.²⁹ It is also consistent with the positions of major institutions such as the World Confederation for Physical Therapy (WCPT), which establishes that physical therapists collaborate in interdisciplinary rehabilitation programs that aim to prevent movement disorders or maintain/restore function and quality of life, responding to patient needs through a dynamic, practical approach, sometimes mediated by technological development.⁴⁴ This underscores that, in settings such as the ICU, a multidisciplinary approach combined with assistive technologies can strengthen the role of the physiotherapist and contribute to better patient outcomes.

Although most of the studies included in this review did not report any adverse events associated with the use of AI-mediated tools, Warmbein et al.²⁵ noted that some patients experienced pain during mobilization with the VEMOTION® robotic system, suggesting the need to adjust the intensity and duration of therapy. This finding is consistent with previous studies in robotic rehabilitation, which have reported discomfort in patients with lower tolerance to movement.^45,46 These results underscore the importance of ensuring adequate pain management in individuals undergoing therapy with such devices, as they are subjected to sustained pressure that often involves cyclic loading and unloading phases, which may be perceived as uncomfortable and lead to pain. In line with the above studies, have reported a positive impact of robot-assisted rehabilitation on several clinical outcomes, including spasticity, motor function, hand and finger function, and cognitive abilities, suggesting improvements in the quality of life of patients with a history of stroke. Furthermore, Chilura et al.^47–49 reported significant gains in functional recovery and disability reduction in a patient with ICU-AW following a 4-month intervention that combined robot-assisted physical therapy for upper and lower limbs with conventional physiotherapy. Although these scenarios differ from those in the ICU setting and the results are not directly comparable or generalizable, they demonstrate that the use of robotics integrated with AI systems has become part of clinical practice across various healthcare professions, intending to improve medium- and long-term patient outcomes. It is important to note that, as previously mentioned, the various devices and algorithms employed, although highly useful, require clear guidance and standardized rehabilitation protocols and procedures to function effectively. This dependency may limit the reproducibility of results across different clinical settings.⁴³

Challenges and Facilitators for the Use of AI

The included studies highlight several challenges and facilitators related to the implementation of AI tools in ICU rehabilitation. Key challenges were: The need for specialized training to operate robotic systems, patient discomfort or pain during mobilization with robotic devices, the requirement for standardized rehabilitation protocols, and potential bias related to data quality in machine learning algorithms.^14,24,25,29 In contrast, the most relevant enablers included: Objective and real-time monitoring of patient mobility, reproducibility and time savings in ultrasound-based muscle assessment, and AI-enabled patient stratification that allows for more individualized rehabilitation interventions.^14,26–29 These findings highlight both practical opportunities and barriers that healthcare professionals and researchers should consider when designing future AI-assisted rehabilitation studies.

The main strength of our review lies in the comprehensive and systematic database search, which ensured a thorough mapping of the existing literature on the topic. Additionally, the majority of the included studies demonstrated high methodological quality. To our knowledge, this is the first study to describe the application of AI in supporting physical rehabilitation in the ICU. It provides a foundation for future research and encourages clinical rehabilitation teams to incorporate these tools into patient assessment and treatment, guided by well-defined institutional protocols.

Limitations

The main limitations of this study include the limited availability of current evidence on the applicability and utility of AI to support rehabilitation in the ICU, the heterogeneity of study designs, small sample sizes in some studies, and the absence of randomized controlled trials. These factors hindered pooled analysis, limited the interpretation of results, and reduced the ability to draw definitive conclusions. Furthermore, some studies may have introduced selection bias by predominantly including patients who were more likely to benefit from AI.

Furthermore, none of the included studies evaluated the cost-effectiveness of AI-based interventions. This remains a critical gap, as implementation costs can represent an obstacle to widespread adoption in resource-limited healthcare systems.^16,17 Furthermore, all identified studies included adult populations aged 31 to 78 years, with no evidence available for pediatric ICU settings. This limits the generalizability of our findings to children and adolescents and highlights the need for future research in pediatric intensive care.

Conclusion

AI emerges as a valuable tool to support and guide the rehabilitation process in patients with critical illness. Technologies, such as the VEMOTION robotic assistance system, noninvasive mobility sensors, the RAIMUS software for assessing peripheral muscle function, and computer vision algorithms, have shown promising results for improving patient outcomes, optimizing resource utilization, and reducing the workload on healthcare personnel. Future research should focus on evaluating the effectiveness of these technologies through rigorous clinical trials and developing strategies to minimize potential adverse effects. Additionally, it is essential to promote the integration of AI tools into clinical rehabilitation protocols within an evidence-based medicine framework.

Supplementary Materials

The supplementary tables are available on the journal website www.ijccm.org.

Orcid

Harold A Payán-Salcedo https://orcid.org/0000-0002-5492-1214

Aura M Castro Aguilera https://orcid.org/0009-0002-8708-1503

María F Salinas Batioja https://orcid.org/0009-0009-1880-0966

Laura M Castillo Diaz https://orcid.org/0009-0000-9072-8895