Artificial intelligence in shoulder arthroplasty: latest concepts and clinical integration

Abstract

Background

Recent advances in artificial intelligence (AI) have redefined shoulder arthroplasty, ultimately improving diagnostic accuracy, refining surgical planning, and personalizing recovery pathways. However, despite promising technical applications, comparative clinical evidence has not yet demonstrated improved routine surgical or patient outcomes in shoulder arthroplasty.

Review content

Therefore, this review summarizes the state-of-the-art AI applications in shoulder arthroplasty, highlights their strengths and limitations, and outlines integration strategies that align technological innovation with surgical expertise. Preoperatively, deep learning–based imaging, which refers to machine learning methods that use multilayer neural networks to identify complex patterns in imaging data, has shown strong diagnostic performance in detecting rotator cuff tears and glenohumeral arthritis, while AI-assisted three-dimensional reconstruction and anatomy-based modeling help visualize glenoid deformity and support more intuitive preoperative planning, complementing conventional navigation. Intraoperatively, AI-based phase recognition and robotic assistance may enhance component placement precision while also offering potential benefits in surgical training, workflow standardization, and real-time feedback. In addition, emerging innovations such as mixed reality guidance, multimodal integration, and semi-autonomous assistance continue to expand the potential scope of AI applications in shoulder arthroplasty. Postoperatively, machine learning–driven prognostic models and wearable monitoring systems predict complications and guide tailored rehabilitation, supporting earlier risk identification when required. However, current AI systems are unable to capture surgeon intuition and context-dependent expertise. Nevertheless, emerging strategies such as explainable AI, multimodal integration, federated learning, and semi-autonomous surgical platforms aim to bridge these gaps by enhancing transparency, scalability, and clinical trust.

Conclusion

The adoption of AI tools by surgeons depends on their demonstrated reliability across diverse patient anatomies, intraoperative variability, and real-world rehabilitation settings. Thus, technical refinement along with mutual trust with surgeons, supportive regulatory frameworks, and institution-specific implementation strategies are essential for safe integration, particularly because current systems may increase cost, workflow complexity, or delay care when not adequately validated or appropriately implemented.

Background

Shoulder arthroplasty is an established treatment option for patients with persistent shoulder pain and functional impairment due to various etiologies, most commonly degenerative joint disease associated with rotator cuff pathology, rotator cuff arthropathy or irreparable cuff deficiency, complex fractures, and failed prior reconstructions [1, 2]. The increase in global population aging in recent years had significantly increased the volume of shoulder arthroplasty procedures, dramatically exceeding 100,000 cases annually in the United States alone over the past two decades [3, 4]. Notably, the increasing demand for reverse total shoulder arthroplasty has expanded indications to include higher-risk populations, such as patients with poor bone quality, severe deformity or multiple prior surgeries, thereby increasing the anatomical complexity and technical variability of these procedures [5]. Unlike hip and knee arthroplasty, shoulder arthroplasty is strongly influenced by complex glenoid morphology, eccentric bone loss, rotator cuff status, and variable soft-tissue balance [6, 7]. In particular, posterior glenoid wear patterns, such as B2 and B3 deformities, can make accurate assessment of glenoid version, inclination, joint-line medialization, and remaining bone stock difficult using conventional planning alone [6, 8]. These shoulder-specific anatomical challenges increase the importance of patient-specific three-dimensional assessment and explain why AI-assisted planning tools may be particularly relevant in shoulder arthroplasty, while also requiring validation in the specific context of glenoid deformity and reverse total shoulder arthroplasty [8, 9].

Early efforts to improve reproducibility in shoulder arthroplasty during the early 2000s focused on CT-based three-dimensional (3D) preoperative planning, intraoperative navigation systems, and the development of patient-specific instrumentation (PSI). These approaches were designed to reduce variability in glenoid component placement and enhance fixation, thereby establishing the conceptual and technical foundation for the subsequent integration of artificial intelligence (AI) technologies into shoulder arthroplasty [8, 10].

However, despite advancements in implant design and surgical technique, shoulder arthroplasty remains highly operator dependent [11] because of variability in surgical decision-making, component positioning, and rehabilitation outcomes, reflecting the absence of fully standardized protocols [12]. Although computed tomography (CT)–based three-dimensional planning is now routinely incorporated into shoulder arthroplasty, conventional preoperative workflows still rely heavily on manual interpretation and simplified planning assumptions, which may not fully capture patient-specific anatomical and biomechanical complexity. Furthermore, intraoperative adjustments and postoperative recovery continue to be influenced heavily by surgeon experience and subjective judgment, including real-time interpretation of tissue quality, bone stock, stiffness, and implant fit [13]. These factors remain difficult to quantify preoperatively and continue to play an important role in operative decision-making despite advances in AI-assisted planning [2, 8].

In this context, AI has emerged as a promising technological adjunct to overcome these limitations [14]. Recent developments in deep learning–based imaging analysis, statistical shape modeling, reinforcement learning–based surgical guidance, and wearable sensor–linked rehabilitation algorithms offer new avenues for automating and individualizing care [15]. In shoulder arthroplasty, AI algorithms have shown promising performance in technical tasks such as rotator cuff tear detection, bone loss quantification, glenoid alignment assessment, and outcome prediction in early clinical and preclinical studies; however, comparative clinical evidence has not yet demonstrated superior routine surgical or patient outcomes. Accordingly, the current evidence base should be interpreted as evidence of technical feasibility and decision-support potential, rather than evidence of proven improvement in routine clinical outcomes [7, 14]. Consequently, these tools may help reduce operator-dependent variability and provide decision-support in planning, execution, and follow-up care; however, their ability to improve clinical outcomes in shoulder arthroplasty has not been demonstrated in comparative clinical studies.

This review aimed to examine the integration of AI technologies across the continuum of shoulder arthroplasty, including preoperative diagnosis and planning, intraoperative decision support and alignment, and postoperative outcome forecasting and rehabilitation. Rather than focusing on a single technical domain, our goal was to provide a clinically oriented overview of how AI is currently being translated into different phases of shoulder arthroplasty care. We also aimed to summarize the technical foundations and clinical applications of these tools, evaluate current limitations in trial design and generalizability, and discuss future directions for safe, explainable, and scalable AI integration. The clinical relevance of this review lies in clarifying both the potential and the current limitations of AI in shoulder arthroplasty, while emphasizing that meaningful benefit remains dependent on robust validation, careful implementation, and real-world clinical evidence.

This article was designed as a narrative review intended to provide a clinically oriented overview of current and emerging AI applications in shoulder arthroplasty across the preoperative, intraoperative, and postoperative stages of care. This article was designed as a narrative review intended to provide a clinically oriented overview of current and emerging AI applications in shoulder arthroplasty across the preoperative, intraoperative, and postoperative stages of care. A structured narrative literature search was performed using PubMed, Scopus, and Google Scholar. The search included peer-reviewed English-language articles published from database inception to (insert search date), with emphasis on contemporary studies published within the last 10 years. Search terms included combinations of ‘shoulder arthroplasty,’ ‘reverse shoulder arthroplasty,’ ‘artificial intelligence,’ ‘machine learning,’ ‘deep learning,’ ‘preoperative planning,’ ‘three-dimensional reconstruction,’ ‘statistical shape modeling,’ ‘patient-specific instrumentation,’ ‘navigation,’ ‘robotics,’ ‘mixed reality,’ ‘wearable devices,’ ‘rehabilitation,’ and ‘outcome prediction.’ Reference lists of relevant articles were also reviewed to identify additional studies. Original research articles were prioritized when available, particularly studies directly addressing imaging analysis, surgical planning, intraoperative navigation or robotics, prognostic modeling, and postoperative monitoring in shoulder arthroplasty. Review articles were used primarily to provide broader context when shoulder-specific original evidence was limited. Articles were excluded if they were not relevant to shoulder arthroplasty, did not address AI or related digital technologies, were not available in English, or focused exclusively on non-clinical engineering concepts without clear relevance to surgical decision-making. Because this was a narrative review rather than a systematic review, formal PRISMA screening, risk-of-bias assessment, and quantitative article selection counts were not performed.

AI in imaging and preoperative planning

AI applications in shoulder arthroplasty have been developed most extensively in the preoperative setting, particularly for imaging analysis, morphologic quantification, three-dimensional reconstruction, and automated surgical planning. Early work in this area focused on machine learning–assisted image segmentation and lesion detection, whereas more recent studies have expanded toward clinically actionable tasks such as automated quantification of scapular and glenoid morphology, estimation of humeral resection planes, and segmentation-free preoperative planning workflows [16–20]. These advances suggest that AI can contribute not only to image interpretation, but also to standardization of planning steps that have traditionally depended on surgeon experience and manual processing (Fig. 1).

Fig. 1.

AI-assisted three-dimensional modeling and patient-specific surgical simulation enhance precision and reproducibility in shoulder arthroplasty planning

Flowchart illustrating the AI-assisted imaging pipeline from an MRI/CT input followed by lesion classification of glenohumeral arthritis and bone defects, statistical shape modeling, and 3D reconstruction for preoperative planning. Traditional workflows rely on manual interpretation followed by statistical shape modeling and manual templating. In contrast, AI-assisted pipelines use deep learning neural networks to recognize glenoid morphology and bone loss, visualized through heatmap overlays. These AI-generated outputs are incorporated into 3D reconstruction platforms, enabling more consistent implant planning and alignment compared with manual methods. Abbreviations: AI, artificial intelligence; MRI, magnetic resonance imaging; CT, computed tomography; 3D, three-dimensional.

Lesion classification and detection in diagnostic imaging

Initial AI studies in shoulder arthroplasty focused on identifying rotator cuff tears, progression of glenohumeral arthritis, and bone defects using CT and magnetic resonance imaging (MRI) [6]. Lee et al. developed a deep learning–based automated screening algorithm for rotator cuff tears using three-plane shoulder MRI and reported an AUC of 0.94, sensitivity of 98%, and accuracy of 96%, indicating that AI may achieve diagnostic performance comparable to that of experienced musculoskeletal readers in selected settings [21]. This type of model may be clinically useful as a decision-support tool in high-volume clinics or institutions with limited subspecialty radiology support, although its retrospective single-center design means that broader external validation is still needed (Table 1).

Table 1.

Summary of studies on the use AI for Imaging and Preoperative Planning

Study	Focus	Methods	Key results	Limitations
Baird et al. [22]	Bibliometric & topic modeling of 7,886 orthopedic articles	BERT-based NLP, bibliometric analysis	Significant growth in shoulder anatomy/surgical technique research over last 3 years	Focused on trends, not direct clinical validation
Lee et al. [21]	Deep-learning algorithm for rotator cuff tears using MRI	CNN on 3-plane MRI	AUC 0.94, Sensitivity 98%, Accuracy 96%	Single institution, retrospective, needs multicenter validation
Blackman et al. [9]	Scapula & humerus two-body statistical shape modeling using CT	Dual-structure statistical shape modeling	CV error 1.13 mm; explained 43.2% of variation	Small, homogenous sample (n = 45); lacks diversity
Tschannen et al. [23]; Kulyk et al. [24]; Spangenberg et al. [20]	AI-predicted implant inclination/version	Simulation & trajectory prediction studies	Deviation < 2° from planned inclination/version	Simulation-based, limited clinical validation
Iannotti et al. [8]; Throchkmorton et al. [10]	Automated CT segmentation → PSI for trajectory & fixation	CT imaging + AI segmentation	Improved insertion trajectory & initial fixation stability	High cost, long production times, limited to specialized centers

Beyond soft-tissue lesion detection, more recent original studies have addressed automated extraction of clinically relevant bony morphology directly from preoperative CT imaging; however, image quality and algorithm reliability may be reduced when metal artifact is present, particularly in patients with retained hardware or prior implants. This issue is especially relevant in revision settings, where artifact may obscure anatomic landmarks and limit the accuracy of automated segmentation or AI-assisted planning. Satir et al. developed an open-source deep learning model for automatic quantification of scapular and glenoid morphology from CT scans in patients with glenohumeral osteoarthritis and concluded that its accuracy was sufficient for clinical use. Marsilio et al. subsequently described a context-aware dual-task deep network capable of concurrent bone segmentation and glenohumeral clinical assessment, showing that integrated AI frameworks can support both three-dimensional osseous reconstruction and clinically relevant preoperative evaluation within a single pipeline [19]. These studies are especially relevant to shoulder arthroplasty because they move AI beyond generic classification and toward extraction of the anatomic parameters that determine implant planning [18].

Other original investigations have focused on specific planning tasks. Tschannen et al. proposed a regression forest–based method for automatic estimation of the articular margin plane for shoulder prosthesis planning, demonstrating the feasibility of automating a key step in humeral preparation [23]. Kulyk et al. later introduced a fully automatic deep learning–based approach to total shoulder arthroplasty planning without explicit segmentation, suggesting that preoperative planning workflows may be streamlined without the need for labor-intensive manual preprocessing [24]. More recently, Spangenberg et al. reported that a deep learning model could automatically determine the humeral resection plane in arthritic humeri with state-of-the-art accuracy, addressing a clinically important challenge in the presence of osteophytes and deformity [20].

In parallel, AI-based CT analysis is beginning to extend to soft-tissue assessment. Rajabzadeh-Oghaz et al. clinically validated a CT-based machine learning model for segmentation and quantification of shoulder muscles, particularly the deltoid, supporting the feasibility of automated muscle characterization as an additional input for arthroplasty planning and prognosis [25]. Collectively, these studies indicate that AI in the preoperative phase is no longer limited to lesion detection alone, but is increasingly able to support morphologic measurement, landmark definition, resection planning, and integrated structural assessment.

Nevertheless, beyond diagnostic support and morphologic quantification, mixed-reality visualization may further facilitate translation of preoperative plans by providing surgeons with an intuitive three-dimensional holographic view of patient-specific anatomy for rehearsal and implant planning in complex cases [26].

Statistical shape modeling and 3D reconstruction

Three-dimensional anatomical reconstruction using AI is technically demanding but clinically important in shoulder arthroplasty [9]. Blackman and Giles introduced a two-body statistical shape model of the scapula and humerus using CT data from arthropathic shoulders and showed that this approach achieved high reconstruction accuracy while capturing the relationship between glenoid erosion and humeral head changes more realistically than single-bone models [9]. This is particularly relevant in reverse total shoulder arthroplasty, where severe deformity and bone loss frequently complicate reconstruction of premorbid anatomy and implant positioning [8–10]. In shoulder arthroplasty, reconstruction of premorbid glenoid anatomy is particularly important in patients with posterior subluxation, eccentric posterior wear, or severe glenoid bone loss [9, 27]. In B2/B3-type deformities, the surgeon must balance correction of version and inclination against the risks of excessive asymmetric reaming, joint-line medialization, reduced baseplate support, or the need for augmented components [6–8]. Statistical shape modeling and AI-assisted three-dimensional reconstruction may support this process by estimating native glenoid morphology and visualizing the extent of deformity, but these outputs must be interpreted in relation to implant fixation, remaining bone stock, and the surgeon’s intraoperative assessment [9, 27–29].

The principal advantage of statistical shape modeling is its ability to summarize anatomic variability across a population-based reference and estimate missing or pathologic bone geometry in individual patients [29, 30]. Salhi et al. demonstrated the utility of statistical shape modeling for predicting missing scapular bone [29], while Abler et al. showed that a scapular statistical shape model can approximate premorbid glenoid morphology with clinically acceptable accuracy [27]. Verhaegen et al. further applied shape-based reconstruction methods in shoulder osteoarthritis, reinforcing the role of such models in understanding humeral head migration and altered joint morphology [31]. Taken together, these studies support the use of statistical shape modeling to guide glenoid reconstruction, reaming strategy, patient-specific instrumentation, and basic motion or impingement simulation in complex arthroplasty cases [8, 9, 31]. However, current statistical shape models remain limited by relatively small and demographically restricted datasets, and their performance may also be compromised when the input imaging data are affected by metal artifact or severe anatomic distortion, particularly in revision surgery. In such cases, artifact and altered bony landmarks may reduce the reliability of automated reconstruction and limit the accuracy of AI-assisted preoperative planning [9, 27, 30]. Cai et al. noted that broader validation across populations, imaging systems, and clinical scenarios remains necessary to improve generalizability [30]. Therefore, future work should focus on larger multicenter datasets, broader representation of pathologic glenoid wear patterns, and integration of biomechanical considerations so that reconstructed anatomy is not only geometrically plausible but also mechanically relevant for implant fixation and stability [9, 31].

Surgical planning optimization and patient-specific components

AI has also been applied to optimize preoperative planning by predicting implant positioning, defining resection planes, and automating portions of the templating workflow; however, implant sizing recommendations must ultimately be interpreted in conjunction with the specifications, compatibility, and availability of commercially available implant systems. Accordingly, the practical value of AI-assisted implant planning depends not only on algorithmic accuracy, but also on coordination with industry-specific implant options and real-world surgical inventory. The studies by Tschannen et al., Kulyk et al., and Spangenberg et al. are particularly relevant here because they show that key planning steps can be automated at the level of humeral reference-plane estimation, segmentation-free total shoulder arthroplasty planning, and resection-plane prediction in arthritic humeri [20, 23, 24]. Together, these studies suggest that AI may reduce interobserver variability and improve efficiency in preoperative planning, although most remain in technical or early validation phases [20, 23]. In addition, AI-assisted CT processing has supported the development of patient-specific instrumentation and more reproducible execution strategies [8, 10, 32]. Earlier work by Iannotti et al. showed that three-dimensional preoperative planning software and information transfer technology improved glenoid component positioning [8], while Throckmorton et al. demonstrated that patient-specific targeting guides could improve glenoid component placement compared with traditional instrumentation in arthritic cadaveric specimens [10]. These studies provide an important bridge between conventional computer-assisted planning and more automated AI-enabled workflows [8, 10, 21].

From a practical standpoint, AI-assisted planning may help standardize evaluation of glenoid version, inclination, humeral head resection, and periarticular muscle status before surgery [19, 20, 23–25]. From a surgical planning perspective, AI-assisted tools may be most useful when they address shoulder-specific decisions rather than generic implant templating [8, 32, 33]. These include estimating the degree of version correction that can be achieved without excessive asymmetric reaming, identifying when augmented glenoid components or bone grafting may be required, and planning baseplate trajectory in reverse total shoulder arthroplasty [10]. However, these recommendations must remain surgeon-supervised because the optimal correction strategy depends on bone quality, deformity pattern, implant system, fixation options, and intraoperative exposure [7, 34]. However, despite promising technical performance, most systems have been developed using limited institutional datasets and require broader clinical validation across surgeons, implants, and patient populations before routine adoption can be justified [9]. In addition, commercially integrated implementation and surgeon acceptance remain important barriers to wider clinical use, even when image-based planning and patient-specific instrumentation are technically feasible [16, 17, 32]. Although these technologies may improve selected aspects of workflow efficiency and preoperative preparation, their practical value depends on external validation, compatibility with available implant systems, cost, and integration into routine surgical workflows [17, 32]. Overall, the preoperative domain currently represents the most mature area of AI application in shoulder arthroplasty, but broader clinical validation and practical implementation evidence remain necessary before these systems can be considered widely generalizable or practice-changing [16, 33, 35, 36].

Mixed reality visualization and surgical rehearsal

Mixed reality represents an AI-adjacent visualization technology that can translate preoperative CT-based planning into an interactive three-dimensional environment for surgical rehearsal and intraoperative orientation [16, 26]. In shoulder arthroplasty, this may be particularly useful for understanding complex glenoid deformity, posterior wear, vault orientation, implant trajectory, and the spatial relationship between planned component position and remaining bone stock [6]. Abdic et al. reported that mixed reality visualization may improve spatial understanding of patient-specific anatomy compared with conventional planning displays, although this evidence primarily supports visualization feasibility rather than improved clinical outcomes [26]. Therefore, mixed reality should currently be regarded as a complementary planning and educational tool rather than a validated AI-driven decision-making system in shoulder arthroplasty [16, 17]. Future studies should determine whether mixed reality improves planning accuracy, surgeon understanding, operative workflow, component positioning, training efficiency, or patient outcomes in prospective shoulder arthroplasty cohorts [26, 37].

Intraoperative applications and robotics

Early intraoperative applications of AI in orthopedic surgery were largely limited to static planning support. In shoulder arthroplasty, currently available intraoperative technologies are more accurately characterized as computer-assisted navigation, robotic assistance, and related guidance systems designed to improve component-positioning accuracy and technical reproducibility. Representative shoulder arthroplasty studies are summarized in Table 2. Importantly, most shoulder-specific evidence in this area relates to technical accuracy, feasibility, and learning-curve assessment, whereas AI-driven workflow recognition, reinforcement learning, and semi-autonomous robotic decision-making remain largely investigational or extrapolated from other surgical fields.

Table 2.

Representative studies and concepts in intraoperative navigation and robotics for shoulder arthroplasty

Study	Focus	Key results	Strengths	Limitations
Verborgt et al.	Computer-assisted navigation for glenoid component placement in reverse shoulder arthroplasty	Navigation improved the accuracy and precision of glenoid component placement compared with conventional instrumentation in a cadaveric setting	Direct comparison of navigated and conventional techniques; postoperative assessment of component position; shoulder arthroplasty–specific evidence	Cadaveric study; small sample size; no patient-reported outcomes, complication data, revision data, or implant survivorship analysis
Wang et al.	Computer navigation of the glenoid component in primary reverse total shoulder arthroplasty, with evaluation of the learning curve	Navigation was feasible in clinical reverse shoulder arthroplasty. Operative time decreased with experience, and postoperative imaging demonstrated component placement close to the preoperative plan	Clinical shoulder arthroplasty study; evaluates learning curve and practical implementation; uses postoperative imaging to assess accuracy	Single-surgeon/single-center design; limited cohort size; no long-term follow-up; does not demonstrate improved revision risk, implant survival, or patient-reported outcomes
Twinanda et al.	Deep learning–based surgical phase recognition from laparoscopic video	Demonstrated that deep learning can classify surgical workflow phases with high accuracy in laparoscopic surgery	Provides proof-of-concept for automated workflow recognition and intraoperative video analysis	Not shoulder arthroplasty–specific; extrapolated evidence only; shoulder arthroplasty validation remains lacking
Shademan et al.	Supervised autonomous robotic soft-tissue surgery in a controlled experimental setting	Demonstrated feasibility of supervised autonomous robotic execution in a non-orthopedic experimental model	Supports the broader concept of semi-autonomous robotic assistance under human oversight

Deep learning–based surgical phase recognition has shown high accuracy in laparoscopic and other video-based surgical fields, where operative phases can be classified from endoscopic video streams [38, 39]. These systems may eventually support automated workflow analysis, objective trainee feedback, and recognition of omitted or inefficient steps [40]. However, direct validation in shoulder arthroplasty remains lacking. Unlike laparoscopic surgery, shoulder arthroplasty often involves variable exposure, changing camera perspectives, limited standardized video capture, and substantial surgeon-dependent differences in workflow [7, 34]. Therefore, surgical phase recognition should currently be considered an extrapolated and investigational concept in shoulder arthroplasty rather than an established intraoperative AI application [7, 36].

In contrast, computer-assisted navigation has been evaluated more directly in shoulder arthroplasty. Verborgt et al. demonstrated that navigation improved the accuracy of glenoid component placement compared with conventional instrumentation in a cadaveric reverse shoulder arthroplasty model [41]. Wang et al. further evaluated computer navigation in primary reverse total shoulder arthroplasty and reported feasibility in clinical practice, including assessment of the learning curve and postoperative imaging-based component position relative to the preoperative plan [42]. These studies support the technical feasibility of navigation for glenoid positioning, particularly in cases where version, inclination, and baseplate trajectory are difficult to reproduce manually. However, they do not establish improved implant survivorship, revision reduction, or patient-reported outcomes. These shoulder-specific challenges are also relevant intraoperatively, particularly during glenoid preparation and baseplate placement in reverse total shoulder arthroplasty [41, 42]. In reverse total shoulder arthroplasty, small errors in glenoid version, inclination, or baseplate trajectory may be more consequential in patients with severe posterior wear, medialized joint lines, or deficient vault bone [8, 41]. Computer-assisted navigation may therefore provide technical value by helping reproduce the preoperative plan during glenoid preparation and baseplate placement. However, navigation does not eliminate the need for intraoperative judgment regarding exposure, remaining bone stock, screw purchase, soft-tissue tension, and implant stability [34].

Beyond computer-assisted navigation, robotic assistance has also been explored as a means of improving insertion trajectories, cutting depth, and baseplate positioning in shoulder arthroplasty. Studies have demonstrated that robotic systems result in reduced glenoid baseplate insertion error compared with manual techniques, with angular deviations decreasing from approximately ± 4–5° to about ± 2° [41, 43, 44]. This increased precision may theoretically support improved initial fixation, but its clinical relevance depends on whether these technical gains can be reproduced across surgeons, implant systems, and complex glenoid deformities. This improvement is technically relevant for the operating surgeon because more accurate baseplate positioning may reduce malpositioning; however, whether this translates into lower loosening or revision rates has not yet been demonstrated. Thus, in practical terms, AI-enhanced robotics may offer technical value over conventional manual workflows by providing more consistent alignment in challenging glenoid deformities [7, 41]. However, its widespread adoption is limited by practical barriers, including high equipment costs, increased operative setup time, and the lack of large-scale multicenter clinical validation [7]. Nevertheless, these technologies represent important steps toward AI-driven standardization and automation in shoulder arthroplasty, although further validation and scalability remain essential for broader clinical integration.

AI-based decision-support systems have also emerged as intraoperative tools to assist surgeons with implant size, orientation, and positioning in real time. These systems combine patient-specific clinical information and preoperative imaging to provide evidence-based recommendations, while leaving final decision-making to the surgeon. Reinforcement learning and semi-autonomous robotic systems remain at a much earlier stage of development in shoulder arthroplasty. Although experimental work such as Shademan et al. has demonstrated the feasibility of supervised autonomous robotic soft-tissue surgery in a controlled non-orthopedic setting, this evidence should not be interpreted as direct support for autonomous or semi-autonomous shoulder arthroplasty [45]. At present, reinforcement learning in shoulder arthroplasty is best regarded as an aspirational framework for future investigation, with unresolved challenges including intraoperative data capture, reliable labeling of expert decisions, safety validation, surgeon oversight, and ethical governance (Fig. 2) [38, 46].

Fig. 2.

AI in intraoperative applications and robotic assistance for shoulder arthroplasty

This figure illustrates the continuum of intraoperative AI applications in shoulder arthroplasty, spanning workflow recognition, robotic precision execution, and reinforcement learning–driven decision support, which collectively contribute to greater reproducibility, safety, and standardization of surgical practice. (A) Surgical phase recognition system in which deep learning algorithms classify sequential steps of shoulder arthroplasty from intraoperative video with > 90% accuracy, enabling detection of inefficient or omitted techniques. (B) Robot-assisted glenoid baseplate placement demonstrating improved angular precision RTSA (manual deviation ± 4.8° vs. robotic ± 2.1°), with potential benefits for implant fixation and long-term survival. (C) Reinforcement learning–based intraoperative decision-support framework for adaptive optimization of implant trajectories, cutting sequences, and component orientation in real time. Abbreviations: AI, artificial intelligence; RTSA, reverse total shoulder arthroplasty.

In orthopedic surgery, such decision-support systems may enhance reproducibility and safety by providing structured intraoperative information and suggesting safer or more efficient technical maneuvers. However, current applications of these systems remain limited to specific procedures or implant types, with insufficient adaptability to the wide variability and exceptions encountered in real clinical settings [47]. Furthermore, their validation has largely been confined to experimental or single-institution pilot studies, underscoring the need for multicenter clinical trials before broad implementation. In summary, intraoperative AI in shoulder arthroplasty should currently be understood as two distinct categories: shoulder-specific navigation and guidance technologies with evidence for technical accuracy and feasibility, and broader AI-driven workflow recognition, reinforcement learning, and semi-autonomous robotic concepts that remain investigational. Future studies should prioritize prospective shoulder-specific validation, clinically meaningful endpoints, workflow impact, cost, and surgeon oversight before these technologies are presented as established intraoperative AI solutions.

Prognostic modeling and postoperative monitoring

The predictability of outcomes after surgery and the optimization of recovery pathways are key factors that determine the success of shoulder arthroplasty. In particular, complications such as re-tear, component loosening, and revision surgery significantly affect the patient’s functional recovery and healthcare resource utilization. Consequently, efforts have been actively made to use AI to predict complications earlier and tailor long-term recovery pathways to the individual patient.

Postoperative outcome prediction modeling

Predicting postoperative outcomes remains a cornerstone in advancing the quality and safety of shoulder arthroplasty. While early prognostic models were largely regression-based, recent advances have leveraged machine learning to incorporate complex, non-linear relationships among patient, surgical, and perioperative factors. For instance, Kumar et al. reported that machine learning models using a minimal feature set could predict clinically relevant postoperative outcomes after shoulder arthroplasty with good performance, supporting the feasibility of individualized outcome prediction in clinical practice (Table 3) [48].

Table 3.

Prognostic modeling and postoperative monitoring after shoulder arthroplasty

Category	Inputs/Modalities	Algorithm/Approach	Author/Setting	Performance/Main Finding	External Validation	Deployment Readiness	Key Limitation
Postoperative outcome prediction	Demographic, surgical, perioperative data	Machine learning models summarized across studies	Kumar et al. [48]—multicenter database study	Reported prediction accuracy up to 90% for complications and > 85% for range-of-motion recovery across included studies	Limited and heterogeneous	Research phase	Predominantly single-center studies; heterogeneous methods and endpoints
Postoperative outcome prediction	Clinical + EMR variables	Logistic regression, SVM, gradient boosting	Powell et al. [49]—registry-based study	Approximately 90% accuracy for 180-day complications and length of stay prediction	Partial, registry-based	Feasible but requires broader validation	Lacks dynamic longitudinal recovery modeling
Wearable-based monitoring	Kinematics, heart rate, exoskeleton assistance	Sensor-based monitoring / assistive wearable evaluation	Zakir et al. [50]—experimental study	Fatigue indicators decreased by approximately 64% with higher exoskeleton assistance	No shoulder arthroplasty–specific validation	Exploratory adjunct	Not performed in postoperative shoulder arthroplasty patients
Wearable-based monitoring	Shoulder kinematics, wearable smart device data	Motion-based machine-learning software development kit	Ramkumar et al. [51]—validation study	Demonstrated passive measurement of shoulder range of motion using a wearable smart device with approximately 5° variability compared with standard goniometer measurement	No	Exploratory / early translational	Not shoulder arthroplasty–specific; validation-focused study rather than direct postoperative clinical outcome monitoring
Integrated recovery optimization	EMR + intraoperative records + wearable sensors	Conceptual AI/ML-enabled personalized recovery platform	Luo [52]—conceptual review	Proposed integrated individualized recovery pathways; no direct shoulder arthroplasty outcome metrics reported	None	Conceptual/exploratory	Direct clinical evidence in shoulder arthroplasty remains limited

Similarly, Powell et al. evaluated four machine learning models for forecasting 180-day complications and extended hospital stay, underscoring the feasibility of long-term prediction in clinical practice [49]. Despite these promising results, current models have several limitations. For example, many models remain trained on single-institution datasets or national registries without sufficient external validation, raising concerns about their generalizability to diverse healthcare systems [35]. Moreover, most models simplify recovery prediction into binary classifications (e.g., complication vs. no complication), which overlooks the dynamic and longitudinal trajectories of patient recovery.

Nevertheless, the clinical implications of these models are increasingly recognized beyond statistical performance. Prognostic algorithms not only provide surgeons with objective tools for preoperative counseling but also enable early identification of patients at higher risk of complications, supporting timely intervention and individualized rehabilitation planning. In clinical practice, this enables surgeons to provide preoperative counseling on realistic functional milestones, adjust implant selection or rehabilitation intensity, and arrange closer follow-up when necessary [40, 53]. Accordingly, recent studies and reviews suggest that machine learning–based predictive frameworks may support preoperative counseling, risk stratification, and closer follow-up planning after shoulder arthroplasty, although their effect on healthcare resource allocation or value-based care pathways requires further validation [48, 49, 53].

Therefore, future research directions should emphasize multicenter validation, incorporation of multimodal datasets including imaging, sensor-derived motion data, and unstructured EMR notes and dynamic modeling that reflects recovery over time. Such efforts are essential for translating prognostic algorithms into widely applicable and clinically trustworthy tools.

Wearable devices and smart rehabilitation tracking systems

In addition to prognostic modeling, wearable technologies are increasingly being explored as adjunctive tools for postoperative shoulder rehabilitation and remote functional monitoring [15, 54]. Although the current evidence specifically linking wearable devices to AI-driven rehabilitation after shoulder arthroplasty remains limited, wearable platforms combined with algorithm-assisted analysis may help track recovery patterns in real time and detect restricted range of motion, overuse, or poor adherence to prescribed exercises [15, 54]. For instance, Zakir et al. investigated muscular and cardiovascular fatigue during overhead tasks using upper-limb assistive exoskeletons and reported that fatigue indicators decreased by approximately 64% (p = 0.002) with higher levels of assistance [50]. Although this study was not conducted specifically in postoperative arthroplasty patients, it suggests that wearable or assistive systems may reduce shoulder workload and may have potential relevance for rehabilitation support in selected patients [50].

Such systems may be particularly useful for monitoring older patients, those in rural or underserved areas, or patients with limited access to in-person rehabilitation, provided that sensor accuracy, adherence, and integration with clinical rehabilitation protocols are adequately addressed. However, important technical challenges remain, including sensor placement, wear duration, signal noise, and the need to maintain reliable performance in non-clinical environments (Fig. 3) [36, 54]. Therefore, although wearable technologies show promise as part of future AI-assisted rehabilitation pathways, further original studies are needed to establish their clinical utility, accuracy, and integration into postoperative care after shoulder arthroplasty [15, 36, 54].

Fig. 3.

Simulated conceptual illustration of AI-assisted postoperative prognostic modeling and rehabilitation monitoring after shoulder arthroplasty. This figure is a simulated conceptual illustration showing how prognostic modeling, wearable biosensors, and individualized recovery monitoring could potentially be integrated after shoulder arthroplasty. It does not represent pooled empirical data, validated clinical performance, or outcomes from a specific clinical study. A The machine learning prognostic model and ROC curve are schematic examples illustrating how risk prediction for complications, reoperation, or delayed functional recovery may be displayed. The AUC values shown are simulated for illustrative purposes and should not be interpreted as reported or validated performance metrics. B The wearable rehabilitation tracking system is a conceptual example of how inertial measurement units and physiologic sensors could be used to monitor shoulder range of motion, activity, and recovery patterns. C The personalized recovery dashboard is a hypothetical mock-up illustrating how patient-reported outcome trajectories, individualized milestones, and complication-risk alerts could be displayed. This simulated panel should not be interpreted as evidence that AI reduces outpatient visits, shortens recovery, or improves postoperative outcomes after shoulder arthroplasty. Abbreviations: AI, artificial intelligence; ROC, receiver operating characteristic; AUC, area under the curve; PROMIS-UE, Patient-Reported Outcomes Measurement Information System–Upper Extremity; IMU, inertial measurement unit; HRV, heart rate variability

Personalized recovery path optimization

In recent years, integrated platforms combining electronic medical records, intraoperative records, and wearable sensor outputs have been proposed to support recovery milestones that reflect each patient’s functional profile [52]. Although these approaches are conceptually promising and may contribute to more individualized rehabilitation pathways, their translation into shoulder arthroplasty care will require validated recovery endpoints, reliable longitudinal data capture, and integration with existing clinical decision-support systems [47]. Integration of these systems with commercial clinical decision support platforms has also been reported only rarely. Therefore, further studies are required to address important challenges including data privacy, algorithm transparency, patient acceptance, and real-world clinical validation [38, 55].

In summary, AI-based prognostic modeling and postoperative monitoring are considered promising technologies for enhancing long-term outcomes in shoulder arthroplasty [14]. These systems analyze multimodal data sources and enable more accurate characterization of patient-specific recovery pathways and early prediction of delayed recovery or complications. Nevertheless, external multicenter validation, integration with commercial medical systems, and robust privacy protections remain prerequisites [55]. These technologies must be grounded in real-world clinical demonstrations, supported by appropriate regulatory and ethical frameworks for their successful clinical adoption.

Current challenges

Although AI technologies have enabled quantitative analysis and automation at various stages of shoulder arthroplasty, complete clinical integration requires consideration of surgeon expertise, intuition, and real-time clinical context [47].Currently, most AI models are trained on formalized input data, which is a major limitation because they fail to incorporate unstructured, context-dependent clinical judgment [14]. Shoulder arthroplasty is an area where surgical strategies must be flexibly determined, depending on anatomical variation and intraoperative progression [7]. For example, in cases of severe bone defects or alignment decisions in modified joints, decisions are often guided not only by preoperative plans, but also by the surgeon’s tactile feedback, assessment of bone quality, and continuous interpretation of the operative field factors that are clinically important but difficult to qualify and quantify in a structured form. This limitation is particularly relevant intraoperatively, where judgment based on tissue quality, exposure, implant fit, bleeding pattern, or unexpected findings may influence decisions in ways that are not readily captured by current data-driven models [34]. However, most AI algorithms are currently restricted to quantifiable structured variables and video-derived information, excluding subjective factors such as tissue elasticity, cutting resistance, or bleeding patterns [40]. Consequently, AI recommendations are often disregarded in the operative field, reducing the acceptability of clinical application [16]. Clinically, this gap indicates that although AI can recommend angles or implant sizes, it cannot yet substitute for the nuanced intraoperative decision-making required when operative findings differ from preoperative planning, including unexpected bone loss, altered exposure, implant fit mismatch, or unanticipated soft-tissue conditions. Such discrepancies between preoperative planning and intraoperative reality remain a major limitation of current AI-assisted systems and reinforce the need for surgeon-led real-time judgment [40, 52].

Although deep learning–based models can provide highly accurate predictions, the reasoning behind these predictions is often criticized as a “black box.” However, the justification for a decision is more important than the raw prediction score itself for surgeons [56]. Nonetheless, most models provide only limited information, such as key variables or sample similarities, without sufficient explanation based on clinical evidence or contextual reasoning [38]. This lack of interpretability weakens surgeon trust in AI recommendations when dealing with high-risk scenarios, such as complications or reoperation decisions [34]. Therefore, future model development must incorporate hierarchical decision-making structures that mimic human reasoning, interpretable machine learning, and rule-based inference systems.

Clinical expertise develops through repetitive exposure to data, and through qualitative judgment and adaptability to unexpected situations [14]. This includes the ability to rapidly recognize subtle intraoperative signals, abnormal bleeding patterns, or rare anatomical variations. By contrast, current supervised learning–based AI is trained only around “correct answer” datasets, making it difficult to capture such exceptional cases [40]. For example, while AI may accurately predict average glenoid version, its application remains limited in atypical scenarios such as anatomical distortion, deformation, or anomaly following multiple prior surgeries, where surgeons depend on pattern recognition and clinical experience [6, 52]. Although some approaches, such as reinforcement learning and federated learning, have been proposed to address this gap, their implantation is limited by the need for real-time tracking, labeling difficulties, and the ethical challenges of capturing the decision-making process during live surgery. Consequently, this remains one of the most critical challenges for the future because no established framework currently exists to translate experience-based judgment into data.

The successful integration of AI into shoulder arthroplasty is only possible if it evolves beyond simple automation to incorporate human intuition and clinical experience. This requires the development of interpretable models, integration of atypical and unstructured data, hierarchical learning systems that reflect expert reasoning, and feedback loop designs rooted in real surgical contexts. Thus, a paradigm shift is required toward achieving true human–AI synergy in orthopedic surgery [47]. At the same time, adoption should remain cautious because poorly integrated AI systems may increase procedural complexity, add cost, prolong decision-making or operative preparation, and fail to improve and in some cases potentially worsen clinical workflows or outcomes if introduced without sufficient validation. Importantly, technical improvements in planning metrics or intraoperative precision should be interpreted cautiously and should not replace prospective evaluation of patient-centered outcomes, complications, workflow burden, and cost.

Future directions

Future development of AI in shoulder arthroplasty should be driven less by general technical novelty and more by shoulder-specific clinical problems. Key priorities include validation in complex glenoid deformity, reverse total shoulder arthroplasty, revision settings, asymmetric reaming decisions, implant-system compatibility, and postoperative recovery monitoring [6, 7, 34]. Therefore, future studies should evaluate not only algorithmic accuracy, but also patient-centered outcomes, complications, revision risk, cost, workflow burden, and surgeon oversight [14, 35, 37]. An additional unresolved implementation gap is surgeon access to vendor-independent interactive three-dimensional modeling and templating tools, as well as intuitive virtual surgery platforms that can be integrated within existing radiology workflows [8, 32, 33]. Although commercial platforms have expanded access to advanced planning, broader implementation will require more interoperable and radiology-integrated systems that are not restricted to specific vendor ecosystems [17, 32].

Shoulder-specific validation and clinically meaningful endpoints

The most important translational gap is the lack of prospective shoulder arthroplasty–specific evidence demonstrating that AI-assisted planning, navigation, or postoperative monitoring improves outcomes beyond technical accuracy [14, 35]. Future studies should therefore focus on clinically meaningful endpoints, including component loosening, instability, revision risk, range of motion, patient-reported outcomes, operative time, cost, and workflow burden [37, 53]. Particular attention should be given to patients with B2/B3 glenoid wear, severe bone loss, revision arthroplasty, and reverse total shoulder arthroplasty, because these are the settings in which AI-assisted planning may be most clinically relevant [6, 8].

Explainable, surgeon-supervised AI for glenoid planning

Explainable AI should be developed around shoulder-specific decisions rather than generic image classification, because clinical trust depends on whether surgeons can understand the anatomical and technical basis of algorithmic recommendations [34]. In shoulder arthroplasty, clinically useful explanations would include why an algorithm recommends a particular version correction, inclination target, augmented component, bone graft, or baseplate trajectory [7, 8]. Such systems should show the anatomical features driving the recommendation, including posterior wear, vault depth, joint-line medialization, remaining bone stock, and predicted fixation area [6, 9, 10, 34]. This approach may improve surgeon trust, but final decision-making must remain surgeon-supervised because intraoperative exposure, bone quality, screw purchase, and soft-tissue tension cannot yet be fully captured by current models [47].

Multimodal and federated learning for shoulder arthroplasty datasets

Future multimodal AI models should integrate shoulder-specific data sources, including CT-based glenoid morphology, MRI-based rotator cuff and muscle status, implant specifications, intraoperative navigation data, rehabilitation progress, and patient-reported outcomes [34, 52]. Federated learning may be useful in this field because individual institutions often have limited shoulder arthroplasty case volumes, especially for complex deformity and revision cases [35, 57]. Multi-institutional collaboration could improve model generalizability while preserving patient privacy, but this will require standardized imaging protocols, implant metadata, outcome definitions, and follow-up intervals [28, 57].

Intraoperative AI and robotics: from technical accuracy to clinical validation

Intraoperative AI and robotics should be evaluated according to a staged translational pathway that distinguishes established shoulder-specific navigation from investigational workflow recognition, reinforcement learning, and semi-autonomous robotic concepts [7, 41, 42]. At present, shoulder-specific evidence most strongly supports computer-assisted navigation as a tool for improving technical accuracy and reproducing the preoperative plan during glenoid component placement [43]. By contrast, surgical phase recognition, reinforcement learning, and semi-autonomous robotic decision-making remain investigational or extrapolated from other surgical fields [39, 40, 45]. Future work should therefore determine whether improved glenoid positioning accuracy translates into better fixation, fewer complications, lower revision rates, or improved patient-reported outcomes, while also measuring setup time, operative disruption, training requirements, and cost [7, 14, 37].

Regulatory, ethical, and implementation framework

Regulatory and ethical frameworks are central to the safe translation of AI in shoulder arthroplasty because many AI-assisted systems may influence implant selection, component positioning, intraoperative execution, or postoperative monitoring [37, 58, 59]. These systems should be evaluated as clinical decision-support tools with clear documentation of intended use, data sources, performance limits, and human oversight [60]. Early implementation should occur within prospective evaluation frameworks that monitor technical accuracy, patient outcomes, adverse events, revision risk, cost, and workflow burden [37]. Patients should also be informed when AI-assisted or investigational digital technologies are used in planning or care pathways, particularly when comparative outcome benefit has not yet been demonstrated [37, 47].

Conclusions

Recent advancements in AI have made practical progress in quantifying and automating clinical decisions across the entire continuum of shoulder arthroplasty, from diagnostic image and video analysis to surgical planning optimization and robotic assistance during surgery, to postoperative prognosis prediction and recovery pathway design. In particular, technologies such as deep learning–based image analysis, statistical shape modeling, wearable sensor–linked prediction systems, XAI, and federated learning–based collaborative models have shifted beyond operator-dependent approaches, contributing to the creation of a precise and evidence-based treatment environment [35]. However, most AI technologies are still under development and have only been validated under single-center or limited conditions. Additionally, significant technical limitations in integrating the intuition, contextual reasoning, and qualitative expertise of clinicians still persist. For instance, the lack of model interpretability, challenges in processing atypical data, and reduced robustness when applied to diverse patient populations are key factors limiting widespread clinical application.

Thus, from a practical perspective, the adoption of AI tools by surgeons depends on their demonstrated reliability across diverse patient anatomies, intraoperative variability, and real-world rehabilitation settings [40, 53]. Consequently, the successful clinical integration of AI depends on technical refinement along with mutual trust with surgeons, supportive regulatory frameworks, and institution-specific implementation strategies are essential. Therefore, the central unresolved gap is not whether AI can improve technical planning metrics, but whether these technical gains translate into safer surgery, better patient-centered outcomes, lower revision risk, and acceptable cost and workflow burden in real shoulder arthroplasty practice. A credible translational pathway should proceed stepwise from shoulder-specific technical validation, to prospective clinical outcome assessment, to regulated implementation with transparent documentation, surgeon oversight, and post-market monitoring. AI should therefore be regarded as a tool to augment and strengthen clinical decision-making, rather than as a replacement for human surgical judgment.

Abbreviations

AI

Artificial intelligence

RTSA

Reverse total shoulder arthroplasty

CT

Computed tomography

MRI

Magnetic resonance imaging

3D

Three-dimensional

PSI

Patient-specific instrumentation

EMR

Electronic medical record

ROC

Receiver operating characteristic

AUC

Area under the curve

PROMIS-UE

Patient-Reported Outcomes Measurement Information System–Upper Extremity

IMU

Inertial measurement unit

HRV

Heart rate variability

XAI

Explainable artificial intelligence

SaMD

Software as a medical device

MDR

Medical Device Regulation

Author contributions

Conceptualization: YDJ and JHO. Literature search and data curation: JO and HJJ. Formal analysis and interpretation: YDJ and JO. Methodology: YDJ and JHO. Visualization: HJJ and JYY. Writing—original draft: YDJ and JO. Writing—review & editing: JHO and JYY. Supervision: JHO. All authors read and approved the final manuscript.

Funding

This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2024-00350460).

Availability of data and materials

No new data were generated or analyzed in this study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.