This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic and Meta-Analysis (PRISMA) (Tricco et al., 2018) guidelines, ensuring methodological transparency and accuracy. A comprehensive literature search was performed using the MEDLINE/PubMed database. The search focused on identifying relevant literature pertaining to the use of AI in UE surgery. The search strategy was designed to capture all relevant studies published between November 2009 and April 2024. The electronic search strategy used was: (Artificial Intelligence OR Machine Learning OR Deep Learning) AND (Diagnosis OR Detection) AND (Hand Surgery OR Arm Surgery OR Elbow Surgery OR Shoulder Surgery).

Studies were included if they evaluated AI applications in UE surgery and were original research articles. Excluded studies included those unrelated to AI in UE surgery, review articles, letters to the editor, conference abstracts, and articles not published in English.

All database search results were imported into Rayyan, a systematic review management tool, where duplicates were automatically removed using a trained AI system, as described by Adu et al. (2024). Two independent reviewers then performed an initial screening of titles and abstracts to exclude studies that did not meet the eligibility criteria. Subsequently, full-text articles of potentially relevant studies were then reviewed independently by both reviewers. At any point, any disagreements regarding study inclusion were resolved through discussion, with the corresponding author serving as the final adjudicator in cases of unresolved discrepancies. Included studies were then sorted into categories based on the perceived primary focus of the paper. When study overlap between two categories occurred, discussion took place, and the studies were placed into their perceived primary category.

Following categorization, 11 of the 118 studies had overlap between two categories. Seven (Minelli et al., 2022; Ro et al., 2021; Alike et al., 2023; Lee et al., 2024; Gu et al., 2022; Kim et al., 2021; Ramkumar et al., 2018) of the studies overlapped between the Image Analysis and Measurement categories. Two (Kluck et al., 2023; Lu et al., 2021) of the studies overlapped between Image Analysis and Surgical Outcome Prediction. One study (Lee et al., 2018) overlapped between Image Analysis and Intraoperative Aid. One study (Cheng et al., 2023) overlapped between Intraoperative Aid and Clinical Decision Support Tool (Figures 3, 4).

Similar to a prior review on hand surgery, the most common application of AI in UE surgery was automated image analysis (Keller et al., 2023), accounting for 42 articles (Anttila et al., 2023; Chung et al., 2018; Dipnall et al., 2022; Droppelmann et al., 2022; Guermazi et al., 2022; Guo et al., 2023; Hahn et al., 2022; Minelli et al., 2022; Ro et al., 2021; Yi et al., 2020; Anttila et al., 2022; Feuerriegel et al., 2023; Feuerriegel et al., 2024; Grauhan et al., 2022; Kang et al., 2021; Kim et al., 2022; Shinohara et al., 2023; Wei et al., 2022; Yang et al., 2024; Yoon et al., 2023; Zech et al., 2024; Zech et al., 2023; Alike et al., 2023; Alike et al., 2023; Benhenneda et al., 2023; Jopling et al., 2021; Keller et al., 2023; Kuok et al., 2020; Lee et al., 2024; Lee et al., 2023; Mert et al., 2024; Ni et al., 2024; Oeding et al., 2024; Shinohara et al., 2022; Suzuki et al., 2022; Anderson et al., 2023; Cirillo et al., 2019; Georgeanu et al., 2022; Jeon et al., 2023; Li and Ji, 2021; Cirillo et al., 2021; Yoon and Chung, 2021). These studies focused on AI-driven interpretation of radiographs, magnetic resonance imaging (MRI), ultrasound, and arthroscopic images, with radiographs being the most frequently analyzed modality.

Assessing the implementation of AI in examining radiographs accounted for 24 articles (Anttila et al., 2023; Chung et al., 2018; Dipnall et al., 2022; Guermazi et al., 2022; Minelli et al., 2022; Yi et al., 2020; Anttila et al., 2022; Grauhan et al., 2022; Kang et al., 2021; Wei et al., 2022; Yang et al., 2024; Yoon et al., 2023; Zech et al., 2024; Zech et al., 2023; Alike et al., 2023; Alike et al., 2023; Jopling et al., 2021; Keller et al., 2023; Lee et al., 2024; Mert et al., 2024; Suzuki et al., 2022; Anderson et al., 2023; Jeon et al., 2023; Yoon and Chung, 2021). AI models show promising capability by quickly and accurately detecting fractures (clavicle, arm, elbow, wrist, hand), measuring critical shoulder angle, identifying shoulder arthroplasty models, and detecting conditions such as enchondromas, joint dislocations, rotator cuff tendon tears, and scapholunate ligament ruptures.

Six studies (Chung et al., 2018; Guo et al., 2023; Zech et al., 2023; Mert et al., 2024; Ni et al., 2024; Suzuki et al., 2022) directly compared AI performance to human clinicians in image analysis, showing that AI matched or outperformed human readers in diagnostic accuracy and speed. One study demonstrated that an AI model achieved an accuracy of 99.3%, a sensitivity of 98.7%, and a specificity of 100% in detecting distal radius fractures, surpassing the performance of three hand orthopedic surgeons (Suzuki et al., 2022). In detecting proximal humerus fractures, AI also outperformed general physicians and non-specialized orthopedists, particularly in complex 3- and 4-part fractures (Chung et al., 2018). AI models integrating deep visual features with clinical data improved diagnostic accuracy for supraspinatus/infraspinatus tendon complex (SITC) injuries, significantly benefiting junior physicians with limited experience (Alike et al., 2023).

A separate study showed that the diagnostic accuracy of an AI algorithm on dorsopalmar radiography regarding scapholunate ligament integrity was close to that of the experienced human reader (e.g., differentiation of Geissler’s stages ≤ 2 versus > 2 with a sensitivity of 74% and a specificity of 78% compared to 77 and 80%) with a correlation coefficient of 0.81 (p < 0.01) (Keller et al., 2023). When AI and humans’ ability to analyze radiographs were directly compared to each other in terms of accuracy or speed, we did not identify any articles that showed humans significantly outperforming AI. Table 1 shows the results of each study that directly compared the image analysis performance between AI models and human readers.

Additionally, several studies (Guermazi et al., 2022; Yoon et al., 2023; Zech et al., 2024; Alike et al., 2023; Anderson et al., 2023) evaluated AI-assisted human image analysis and found that AI augmentation improved clinician accuracy. In a retrospective study of fracture detection, AI-assisted readings increased sensitivity by 10.4% (75.2% vs. 64.8%), while maintaining specificity and reducing average reading time by 6.3 s per case (Guermazi et al., 2022). One study showed AI improves fracture detection among radiology and orthopedic residents in both pediatric and adult patients (Zech et al., 2024). Additionally, this study shows that AI enhances the specificity, sensitivity, and accuracy of physicians diagnosing supraspinatus/infraspinatus tendon complex injuries (Alike et al., 2023). Furthermore, AI assistance was shown to improve physician diagnostic sensitivity and specificity as well as interobserver agreement for the diagnosis of occult scaphoid fractures (Yoon et al., 2023). Similar findings were shown in several specialties, such as orthopedics, emergency medicine, radiology, and primary care, where the fracture miss rate was significantly reduced when aided by AI (Anderson et al., 2023).

A total of 24 articles (Allen et al., 2024; Biron et al., 2020; Digumarthi et al., 2024; Giladi et al., 2023; Gowd et al., 2019; Gowd et al., 2022; Hoogendam et al., 2022; Karnuta et al., 2020; Kausch et al., 2020; King et al., 2023; Kluck et al., 2023; Kumar et al., 2021; Kumar et al., 2020; Kumar et al., 2022; Li et al., 2023; Lu et al., 2022; Lu et al., 2021; Mclendon, 2021; Oeding et al., 2023; Polce et al., 2021; Rajabzadeh-Oghaz et al., 2024; Roche et al., 2021; Shinohara et al., 2024; Simmons et al., 2023; Vassalou et al., 2022) investigated AI’s ability to predict surgical outcomes in UE surgery. These studies focused on rotator cuff arthropathy, carpal tunnel syndrome, and calcific tendonitis, with total shoulder arthroplasty (TSA) being the most frequently analyzed procedure. Among these, 10 studies specifically assessed AI’s ability to predict patient outcomes following anatomic (ASA) or reverse (RSA) total shoulder arthroplasty. All articles except one were retrospective and tested a variety of language learning models (LLMs) with different input variables.

AI models demonstrated high predictive accuracy in estimating postoperative outcomes, such as improvements in shoulder function, patient satisfaction, and complication risk. The predictive variables analyzed included patient history/demographics, pain and functionality scores, physical exam findings, imaging data (X-ray, CT), and laboratory values. Multiple studies showed that machine learning models could achieve AUC values between 0.71 and 0.94, effectively predicting postoperative range of motion (ROM), risk of infection, and the likelihood of requiring revision surgery. One study demonstrated 92.9% accuracy (AUC 0.875) in predicting multiple clinical outcomes after TSA using a limited set of 19 preoperative variables, minimizing the need for extensive data input (Kumar et al., 2020).

For each of the 10 studies involving total shoulder arthroplasty patients, Table 2 details the input variables used, data set size, predictive task, and predictive ability.

AI has also been applied to automated motion analysis and physical assessment in 16 articles (Burns et al., 2018; Darevsky et al., 2023; Darevsky, 2023; Dousty and Zariffa, 2021; Gauci et al., 2023; Gu et al., 2022; Ibara, 2023; Kim et al., 2021; Koyama et al., 2022; Koyama et al., 2021; Lee et al., 2016; Ramkumar et al., 2018; Rostamzadeh et al., 2024; Silver et al., 2006; Takigami et al., 2024; Tsukamoto et al., 2024; Tuan et al., 2022). These studies explored AI models designed to analyze videos or images of body movements including shoulder range of motion, hand gestures, grip strength, and thumb opposition.

Six studies (Darevsky et al., 2023; Gu et al., 2022; Koyama et al., 2021; Ramkumar et al., 2018; Takigami et al., 2024; Tsukamoto et al., 2024) utilized widely accessible devices, such as smartphones and smartwatches, to aid in automated physical examination. These AI models demonstrated high accuracy, exceeding 90% in classifying rotator cuff injuries and nerve dysfunction based on motion analysis. One study used AI-powered pose estimation to measure shoulder internal and external rotation, achieving a correlation coefficient of 0.971 and a mean absolute error of 5.778° compared to standard goniometric measurements (Takigami et al., 2024) (Table 3).

UE orthopedics also includes prosthetic devices, which play a significant role for many amputee patients, and optimizing the function and utility of these devices with AI is an emerging topic of research. AI has played a key role in enhancing prosthetic limb control, particularly through surface electromyography (sEMG)-based myocontrol. Among the 16 studies (Atzori et al., 2014; Atzori et al., 2016; Castellini et al., 2009; Edwards et al., 2016; Hahne et al., 2017; Hwang et al., 2017; Jiang et al., 2020; Malešević et al., 2021; Mastinu et al., 2020; Nowak et al., 2023; Olsson et al., 2019; Osborn et al., 2021; Patel et al., 2017; Schmalfuss et al., 2018; Wang et al., 2022; Wang et al., 2020) in this category, many focused on improving real-time prosthesis functionality through AI-driven motor learning and predictive feedback systems (Table 4).

The first of these articles was published in 2009, and since then, interest in this field has increased significantly (Figure 5). In fact, this was the earliest article included in this review, showing that prosthetics was one of the first areas of interest to implement AI in the UE.

Movements performed by prostheses are performed in an “on/off” fashion, thus rendering coordinated movements with a set amount of force by particular muscles difficult. To overcome this, many prosthesis designs have aimed at incorporating electromyography (EMG) data to allow for more fine-tuned functionality. This is given further power when such input data is processed by way of a machine learning algorithm that can provide real-time feedback and updates as well as learn for future use. AI-driven pattern recognition algorithms have enabled fine-tuned, adaptive myoelectric control, allowing upper extremity amputees to achieve more coordinated, natural movement. Some studies incorporated real-time ultrasound feedback to improve AI-based prosthesis control, achieving accuracy comparable to electromyography-based models (Wang et al., 2020). Others demonstrated that machine learning-enhanced myoelectric control systems could significantly reduce reaction time and improve grip precision in prosthetic hand users (Nowak et al., 2023; Osborn et al., 2021; Patel et al., 2017).

Twelve studies (Bernard et al., 2022; Bockhacker et al., 2020; Cheng et al., 2023; Eslamian et al., 2016; Eslamian et al., 2020; Hein et al., 2021; Kuthiala et al., 2022; Lee et al., 2018; Li et al., 2021; Shafiei et al., 2021; Sühn et al., 2023; Suh et al., 2011) investigated AI’s intraoperative applications, including robotic-assisted surgery, real-time bacterial identification, and automated instrument tracking. Most of these studies are lab-based, with no proof-of-concept in actual surgeries. One study showed that AI-based bacterial identification systems detected osteomyelitis-causing pathogens within five hours, significantly faster and in a less labor-intensive manner than traditional microbial cultures (Bernard et al., 2022). Similarly, another study demonstrated that AI-assisted intraoperative soft-tissue sarcoma classification achieved an accuracy above 85%, outperforming the traditional gold standard of H&E staining frozen sections, which often delays completion of the surgical procedure (Li et al., 2021).

AI-enhanced robotic surgery was explored in three studies (Eslamian et al., 2016; Eslamian et al., 2020; Sühn et al., 2023), showing that autonomous AI-controlled surgical cameras improved visualization, reduced unnecessary movements, and enhanced procedural efficiency and flow. This method was found to be superior to manual camera movement by the surgeon or a trained camera operator. Such technology additionally keeps the surgical instruments in view and avoids unnecessary movement of the camera, preventing inadequate visualization and distraction to the surgeon (Eslamian et al., 2016; Eslamian et al., 2020).

The direct tactile assessment of surface textures during palpation is an essential component of open surgery that is impeded in minimally invasive and robot-assisted surgery. A data generation framework proved accurate (>96%) in using vibro-acoustic sensing to differentiate materials during minimally invasive and robot-assisted surgery. This technology could provide valuable information during procedures such as a total joint replacement or arthroscopy, in which the osteoarthritic cartilage could be identified and graded to help the surgeon plan and make intraoperative decisions (Sühn et al., 2023).

Other intraoperative uses for AI included automated surgeon distraction monitoring (Shafiei et al., 2021), real-time detection of peripherally inserted central catheter (PICC) tips (Lee et al., 2018), segmenting arm venous images (Kuthiala et al., 2022), and gesture-controlled sterile navigation systems. One study evaluating AI-assisted touchless image viewing in the operating room, predicted the hand gestures of eight surgeons with an average of 6.5 years of experience, reaching a 98.94% accuracy in executing the correct task (Bockhacker et al., 2020).

AI was utilized as a clinical decision support tool (CDST) in six articles (Bulstra et al., 2022; Daher et al., 2023; Jagiella-Lodise et al., 2024; Rigamonti et al., 2021; Simmons et al., 2022; Yamamoto et al., 2024), meaning they were used in some degree to aid clinical decision-making but did not fall under any of the above categories. These studies focused on diagnostic guidance, treatment planning, and patient education.

Two studies evaluated ChatGPT’s diagnostic capabilities in UE conditions. One study found that ChatGPT correctly diagnosed and recommended appropriate management for 93 and 83% of shoulder and elbow cases, respectively (Daher et al., 2023). Another study assessed ChatGPT’s ability to answer common patient questions related to hand and wrist pathologies, with responses receiving an accuracy rating of 4.83 out of 6 (Jagiella-Lodise et al., 2024).

Another study tested the ability of an AI program to predict scaphoid fractures given elements of a patient’s demographics, history, and physical exam findings without being provided imaging (Bulstra et al., 2022). This machine learning algorithm achieved an area under the receiver operating characteristic curve of 0.77 when predicting the probability of a scaphoid fracture for a retrospective patient cohort. Although accurate, this performance does not exceed that of experienced physicians, who have shown a negative predictive value of up to 96% when predicting scaphoid fractures using a Clinical Scaphoid Score, without the aid of imaging (Pham, 2025). Additionally, this program was able to recommend advanced imaging for patients with a ≥ 10% risk of fracture, yielding 100% sensitivity, 38% specificity, and would have reduced the number of patients undergoing advanced imaging by 36% without missing a fracture.

Another study evaluated how a CDST would help surgeons plan preoperatively whether to perform an anatomic or reverse total shoulder arthroplasty for a patient with osteoarthritis. While this tool did not necessarily direct their decision, it improved their confidence in their own chosen decision (Simmons et al., 2022). Finally, one study discussed the ability of an AI model to analyze gait characteristics from in-shoe wearable monitors to predict distal radius fracture risks (Yamamoto et al., 2024).

These studies are outlined in Table 5.

Risk of bias was assessed using the QUADAS-2 tool for diagnostic accuracy studies and the PROBAST tool for prediction model studies. Among the studies evaluated with QUADAS-2 (a total of 28), 15 were judged to have a high overall risk of bias, 10 had a low risk, and 3 had an unclear risk. For studies assessed with PROBAST (a total of 90), 53 demonstrated a high overall risk of bias and 37 had a low risk. No studies in the PROBAST group were rated as having an unclear risk of bias. These assessments provide insight into the methodological quality and potential limitations of the included studies. Table 6 and Figure 6 show the results of the QUADAS-2 analysis, and Table 7 and Figure 7 show the results of the PROBAST analysis.