An overview of artificial intelligence and machine learning in shoulder surgery

Introduction to ML

ML is a form of AI that focuses on the development of algorithms that learn patterns from data and then make predictions or decisions without the need for explicit programming. Unlike traditional rule-based programming (in which humans issue step-by-step instructions to computers), ML models use computational techniques to identify relationships among data and improve their understanding of such relationships over time to generalize new inputs. ML extracts useful insights from large datasets and has been shown to aid in decision-making in the fields of medicine, finance, engineering, and natural language processing (NLP) [1].

Learning Paradigms of ML

In general, ML uses three learning paradigms: supervised, unsupervised, and reinforcement learning (RL). These paradigms differ in methods of models learning from the data and the problems they solve [2]. Supervised learning is the most widely used approach. Models are trained using labeled data, and each input is paired with a known output. The algorithm learns by minimizing the difference between predictions and actual outcomes, effectively handling classification and regression tasks [3,4].

In contrast, unsupervised learning uses unlabeled data to reveal hidden patterns and/or structures when there are no predefined answers [4]. Such learning is commonly used for clustering (e.g., segmentation of customer data during targeted marketing) and dimensionality reduction (e.g., compression of high-dimensional genomic data). Principal component analysis is commonly used for unsupervised learning and reduces data complexity while preserving important features. K-means clustering groups data points on the basis of their similarity, revealing underlying distributions in large datasets [5]. Generative modeling is a form of unsupervised learning that has become very popular in recent years. ML systems create new data samples based only on training data.

Generative Modeling and Neural Networks

Recent advances in generative modeling include generative neural networks, which can be divided into generative adversarial networks and variational autoencoders [6]. These architectures generate new data that closely resemble those of a given dataset. Generative adversarial networks, first introduced by Ian Goodfellow in 2014, comprise two neural networks—a generator and a discriminator—that mutually compete [7]. The generator creates synthetic data, and the discriminator determines whether the data are real or fake. Over time, this adversarial process forces the generator to create increasingly realistic outputs, which have been used in fields ranging from medical image synthesis to drug discovery, deepfake technology, and art. In contrast, variational autoencoders use probabilistic encoding to learn latent data representations and then engage in highly effective image denoising, anomaly detection, and three-dimensional (3D) reconstruction [8].

RL is an even more advanced paradigm in which an agent learns how to make decisions by interacting with an environment while receiving feedback in the form of rewards or penalties [9,10]. Unlike supervised learning, which requires large labeled datasets, RL optimizes decision-making via trial and error and finds applications in robotics, autonomous driving, and gaming AI, in which the systems must dynamically adapt to new and complex scenarios [10]. RL has also aided the development of AlphaGo and AlphaFold, aiding strategic decision-making and prediction of protein structures, respectively [11].

Key Algorithms and Techniques in ML

ML models are built using a variety of algorithms, each with unique strengths and weaknesses. A schematic representation of the various ML algorithms is shown in Fig. 1, illustrating their underlying models and computational structures [12]. Linear regression and logistic regression are the foundational algorithms that predict numerical outcomes and binary classifications, respectively. More sophisticated techniques include decision trees, support vector machines, and ensemble learning methods such as random forest and gradient boosting (e.g., XGBoost, LightGBM) [13,14]. All of these combine multiple weak learners to improve predictive performance. One of the recent examples of ML studies compared logistic regression with other multiple advanced ML algorithms for retear prediction after arthroscopic rotator cuff repairs [15]. In recent years, deep learning, a subset of ML inspired by the structure of the human brain, has revolutionized fields such as computer vision, speech recognition, and NLP.

Fig. 1.

Schematic representation of machine learning algorithms. It illustrates the structural frameworks of different machine learning models, including supervised learning, unsupervised learning, and deep learning techniques.

Deep learning models, particularly ANNs and CNNs, excel in the analysis of complex, high-dimensional data such as images, audio, and text [14]. For example, CNNs have transformed image recognition tasks using hierarchical layers to detect patterns like edges, shapes, and textures, enabling breakthroughs in medical imaging, facial recognition, and autonomous navigation. Multiple studies have used CNNs to evaluate diagnostic images such as magnetic resonance imaging (MRI) or computed tomography (CT), like measuring fatty infiltration in MRI [16]. Other deep learning architectures, including recurrent neural networks and advanced variants termed long short-term memory networks, handle sequential data, rendering them ideal for applications in speech synthesis, machine translation, and time-series forecasting [17]. In addition, transformer-based models such as the generative pretrained transformer have revolutionized text generation, language understanding, and chatbot applications, demonstrating the power of generative neural networks when processing large-scale natural language data [18].

Explainable AI and Model Interpretability

In addition to traditional techniques, explainable AI (XAI) has garnered significant attention for ML models, particularly deep learning models, as they have become more complex and less easily interpretable. Methods including Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations help to explain model predictions by revealing the hierarchy of factors and how they contribute to the model results [19]. This increases trust and transparency, especially of high-stakes applications in healthcare, finance, and legal decision-making. It is increasingly possible to interpret the operation of ML models, which is essential to ensure fairness, reliability, and ethical deployment.

Despite such remarkable capabilities, ML faces several challenges, including data quality, bias, overfitting, and computational requirements [20,21]. Large datasets are often required to train robust models. Poor-quality data can yield misleading predictions. Overfitting, whereby a model memorizes training data rather than generalizing the data to accommodate new inputs, can reduce real-world performance. Addressing these challenges requires careful feature selection, regularization techniques, and rigorous cross-validation. Recent advances in federated learning and edge AI have mitigated privacy concerns and reduced computational bottlenecks by enabling ML training across decentralized networks without sharing sensitive data.

Predicting Clinical Outcomes after Shoulder Arthroplasty

One of the most impactful applications of ML in shoulder arthroplasty is the prediction of clinical outcomes, including ROM, pain reduction, functional improvement, and patient satisfaction. ML models trained on vast datasets of preoperative demographic details, radiographic assessments, and patient comorbidities predict the above-mentioned variables with remarkable accuracy. Algorithms including XGBoost and wide and deep neural networks have been used to predict active abduction, forward flexion, and external rotation after TSA and rTSA. These models achieve mean absolute errors ranging from 10° to 21°, surpassing the predictive reliability of traditional regression models [23,24]. Furthermore, ML achieves high accuracy in identifying patients who will experience minimal clinically important differences in functional scores such as the American Shoulder and Elbow Surgeons (ASES) score, Constant-Murley score, and visual analog scale pain score [25]. Such predictive tools help set realistic postoperative expectations for patients and help surgeons in terms of patient selection and surgical decision-making.

ML also plays crucial roles in predicting complications and stratifying patients according to risk. While shoulder arthroplasty has a high success rate, postoperative complications such as infections, dislocations, and implant failures can significantly compromise recovery. ML models that use logistic regression and decision trees and ensemble learning methods such as random forest and various gradient boosting machines predict 30-day postoperative complications with acceptable accuracy [26]. This allows early interventions, enabling surgeons to adjust surgical plans, optimize perioperative antibiotic regimens, and tailor follow-up strategies for high-risk individuals [27-29]. Additionally, deep learning techniques, including CNNs and NLP, have been used to analyze electronic medical records, operative notes, and imaging data to identify signs of periprosthetic joint infection before infection becomes clinically evident [30].

AI-Assisted Implant Selection, Surgical Planning, and Cost Prediction

Other significant advances facilitated by ML include optimization of implant selection and surgical planning. Traditionally, implant choice and positioning relied heavily on the experience of the surgeon and static templates, respectively. ML algorithms facilitate a more data-driven approach, using preoperative CT scans, patient biomechanics, and historical surgical outcomes to suggest the most appropriate type of prosthesis, implant size, and positioning. Some systems integrate 3D imaging with RL techniques to ensure that the implant configurations guarantee long-term joint stability. While the direct application of RL in shoulder surgery is still emerging, its effectiveness in dynamically optimizing clinical outcomes has been demonstrated in related orthopedic contexts. For instance, recent research has used RL to enhance cost-efficiency and adaptive clinical trial designs in knee osteoarthritis to dynamically monitor patients, maximize informative data collection, and minimize overall healthcare costs [23]. CNN-based models are now employed for real-time intraoperative navigation, enhancing glenoid component placement, screw positioning, and soft tissue balancing. Such AI-assisted surgical techniques have been integrated into robotic-assisted arthroplasty systems, improving precision and potentially reducing revision rates [27].

ML is also assisting the health-related financial economy. Predictive models are now being used to estimate the length of hospital stay, readmission risk, and perioperative costs. ML algorithms consider patient comorbidities, surgical complexity, and discharge requirements and achieve accuracy exceeding 85% when forecasting whether a patient will require prolonged hospitalization or a non-home discharge [26]. Additionally, cost estimation models that incorporate ML are being developed to predict perioperative expenditures of implant, anesthesia, rehabilitation needs, and post-discharge care [31,32].

Future Directions in ML for Shoulder Arthroplasty

As ML continues to evolve, further applications in the field of shoulder arthroplasty are likely. The next frontier is the development of personalized predictive models that integrate genomics, lifestyle factors, and patient-reported outcomes to generate highly individualized treatment recommendations. Federated learning approaches effectively enable the sharing of multi-institutional data without compromising patient privacy, significantly reducing biases related to patient diversity and single-institution limitations. Despite these strengths, federated learning does not entirely eliminate biases that may arise from differences in clinical protocols, data collection methods, or patient management strategies across participating institutions. Thus, standardizing data collection procedures and establishing rigorous validation practices are essential to enhance model reliability and clinical applicability. These approaches will further enhance model robustness and generalizability. Additionally, advances in AI-integrated robotics and smart implants—such as sensor-embedded prosthetics that provide adaptive feedback, aiding continuous performance monitoring—may revolutionize arthroplasty [33]. The continued advances in AI and ML are fundamentally reshaping shoulder arthroplasty planning, execution, and evaluation. By enabling more accurate risk assessments, personalized treatment plans, and cost-effective healthcare strategies, ML is not simply augmenting surgical decision-making, it is redefining the standard of care in modern orthopedic practice.

ML in RCTs

RCT is a common cause of shoulder dysfunction in both athletes and aging populations. Clinical evaluation and imaging techniques such as MRI and ultrasound have been the traditional gold standard techniques for diagnosing RCTs. ML has become a transformative tool for diagnosis, prognosis, and treatment planning. Advanced ML models enable precise tear characterization and predict reparability and required postoperative monitoring, enhancing both clinical efficiency and patient outcomes [34].

Diagnostic Applications of ML

One of the most immediate and impactful applications of ML in the context of RCT is automated diagnosis via medical imaging. Deep learning algorithms, particularly CNNs, have been trained to detect full- and partial-thickness tears on MRI and ultrasound scans with remarkable accuracy. The area under the curve (AUC) exceeds 0.90, and ML detects tears as well as or better than human radiologists [34].

One significant advance was automated segmentation of the rotator cuff muscles on MRI images. Traditionally, muscle quality evaluation involved subjective grading systems such as the Goutallier classification, and the results were inconsistent among observers. However, recent ML models allow fully automated segmentation of rotator cuff muscles, extracting quantitative metrics such as the extent of fatty infiltration and muscle atrophy with Dice similarity coefficients that exceed 0.92 [35]. Such models offer an objective and reproducible assessment of muscle degeneration, which is crucial for predicting surgical reparability and setting postoperative expectations. Ultrasound-based ML models have also gained traction [36,37]. These models allow early RCT detection, especially in outpatient settings where MRI may not be readily available [37]. In some studies, ML-enhanced ultrasound scan evaluation achieved an accuracy greater than 88% in identifying RCTs, rendering ultrasound screening a viable alternative to MRI [30].

Predicting Rotator Cuff Reparability and Treatment Planning

While RCT must be reliably detected, a more critical clinical question is whether the tear is reparable or not. ML models now predict RCT reparability by integrating preoperative MRI data, tear size, and the extent of fatty infiltration and muscle atrophy. One study used an elastic-net logistic regression model to distinguish reparable and irreparable tears with an AUC of 0.85, outperforming conventional scoring systems [38]. Such models are now being deployed as online prediction tools, allowing surgeons to input patient-specific parameters and review personalized reparability probabilities before choosing a surgical approach.

Other applications of ML during surgical planning include prediction of tendon healing and retear risk. A recent study used deep learning to predict the likelihood of rotator cuff retear after surgery based on intraoperative arthroscopic images [39]. By analyzing the integrity of repaired tendons, the models achieved a predictive accuracy of 91%, providing real-time intraoperative insights that aided the adjustment of surgical techniques and/or refinement of postoperative rehabilitation protocols.

Postoperative Monitoring and Outcome Prediction

ML is now being integrated into postoperative monitoring. One of the most promising applications is the use of both smartphone-based motion tracking and ML to assess shoulder function in real-time [40,41]. A recent study demonstrated that markerless motion capture using standard smartphone videos accurately assessed shoulder kinematics and predicted postoperative recovery trajectories. Such ML-powered motion analysis tools aid remote patient monitoring, reducing the need for frequent hospital visits and providing quantitative feedback on recovery. Patients perform simple predefined movement tasks, and the ML algorithm processes the data to detect motion asymmetry, ROM limitation, and compensatory movement patterns, which may indicate potential complications or delayed healing [42,43].

ML is now also used to predict postoperative functional outcomes, including the ASES score and ROM. By analyzing preoperative clinical factors, ML models can forecast which patients are most likely to achieve a meaningful improvement after surgery. Such models exhibit high predictive accuracy (AUCs >0.85), enabling clinicians to set realistic patient expectations and tailor the rehabilitation programs accordingly [44].

Future Directions and Challenges

While ML has demonstrated tremendous promise in terms of RCT diagnosis and management, several challenges remain, including data variability. ML models are often trained on data from single institutions and may not generalize across patient populations [34]. Future efforts should focus on multicenter collaborations and the use of federated learning approaches to develop more robust and generalizable models.

Additionally, model interpretability is of concern. Many deep learning models function as black boxes, hindering understanding of a specific prediction. XAI techniques such as SHAP are being increasingly used to enhance transparency and trust in ML-assisted decision-making. Another challenge is integration of ML models into clinical workflows. While several web-based prediction tools have been developed, widespread adoption requires seamless integration of ML with electronic medical records and surgical navigation systems. Advances in real-time AI assistance during arthroscopy could further enhance intraoperative decision-making, potentially improving surgical precision and long-term outcomes.

Application of ML in Trauma

Chung et al. [45] applied deep learning techniques to automate the detection and classification of proximal humerus fractures using plain radiographs . Their model, based on CNN, achieved superior performance compared with general physicians and orthopedists, particularly in complex 3- and 4-part fractures, demonstrating the potential to significantly enhance initial diagnostic accuracy and facilitate timely interventions. Furthermore, advanced ML methods have been used to assist preoperative planning of proximal humerus fractures. Jeon et al. [46] developed an AI-based virtual reduction model that segments and virtually reduces fracture fragments using CT imaging. Their system significantly improved reduction accuracy, operational speed, and overall reduction quality compared with manual methods, supporting surgeons in achieving anatomical alignment with greater efficiency and precision.

Application of ML in Shoulder Instability

Research has demonstrated the utility of ML in shoulder instability surgery. Predictive ML models, using demographic and clinical data, have been developed to forecast outcomes following surgical intervention for anterior shoulder instability, helping surgeons tailor their approaches to individual patient profiles, potentially improving postoperative outcomes and patient satisfaction [47].