Artificial intelligence and machine learning in sports medicine: mapping clinical tasks and assessing clinical maturity - a scoping review Authors: Jakob Lindskog, Kristian Heder Ternell, Yinan Yu, Ida Lindman, Kristian Samuelsson, Eric Hamrin Senorski Published: 2026 ABSTRACT AbstractBackgroundArtificial intelligence (AI) and machine learning (ML) are rapidly transforming the medical field. The aim of this review was to outline the current scientific state of AI and ML application in sports medicine, evaluate the level of clinical validation and readiness for implementation, and identify key priorities to guide future advancements and implementation into injury risk assessment, diagnosis, rehabilitation and clinical decision-making in sport medicine.MethodsA scoping review was conducted with a literature search performed on February 5, 2026, using the MEDLINE, EMBASE and Web of Science databases which targeted AI or ML application on individuals within a sports medicine context.ResultsOf 8,677 studies, 97 studies were included. Most research covered orthopaedics (70.1%) and neurology (18.6%), where AI was applied for injury prediction, diagnostic image analysis, and recovery estimation. Predictive and estimation models were the dominant application (57.7%). Reported discriminative performance was frequently high. However, the majority of studies relied on retrospective datasets and internal validation. Calibration reporting was uncommon, and prospective workflow integration was rare, with a single study attempting an interventional prevention strategy. Substantial heterogeneity in modelling approaches, data inputs, and outcomes definitions was observed.ConclusionAlthough AI and ML applications in sports medicine frequently demonstrate strong within-sample performance, most remain in early-stage development. Currently, these tools should be viewed as supportive adjuncts rather than autonomous decision-making systems.Key TermsAI, Predictive modeling, Diagnostic imaging, Rehabilitation, Deep learning, Return to sportSupplementary InformationThe online version contains supplementary material available at 10.1186/s12911-026-03615-w. Introduction Artificial intelligence (AI) and machine learning (ML) are transforming healthcare by enabling machines to effectively analyze data, recognize patterns, and aid decision-making [1]. There is research that suggests that AI and ML might outperform humans within specific healthcare domains [2–4]. For example, ML models have, based on radiographic analysis, identified patients at risk for poor outcomes after unicompartmental knee replacement with greater accuracy than surgeons [2]. Similarly, chatbots powered by large language models (LLMs) have in one study surpassed physicians in diagnostic accuracy [3]. Furthermore, AI has reduced the miss rate for colorectal neoplasia detection by half [4]. These examples highlight the potential and increasing role of AI and ML in improving healthcare outcomes. Within sports medicine, AI and ML applications have attracted considerable interest. For example, Hu et al. [5] demonstrated that the use of a convolutional neural network (CNN) could detect anterior cruciate ligament (ACL) injury through magnetic resonance imaging (MRI) with an accuracy of 96.5%. Furthermore, Allen et al. [6] reported that a decision tree model could discriminate between early, typical and delayed recovery after sports-related concussion (SRC). Moreover, an extreme gradient boosting (XGBoost) model was used to predict level of match participation in football athletes after Achilles tendon rupture [7]. In addition, a recent scoping review found that ML models have been widely applied to support ACL injury prediction, rehabilitation monitoring and return to sport (RTS) decision-making [8]. These studies showcase the potential of AI and ML in sports medicine. However, the literature remains fragmented, with limited understanding of how applications vary across fields and whether current work has progressed beyond retrospective analyses toward prospective or interventional use. Furthermore, it remains unclear whether existing AI and ML applications have reached sufficient developmental and clinical maturity to inform decision-making in sports medicine. The aim of this review was to outline the current scientific state of AI and ML application in sports medicine, evaluate the level of clinical validation and readiness for implementation, and identify key priorities to guide future advancements and implementation into injury risk assessment, diagnosis, rehabilitation and clinical decision-making in sport medicine. Method Protocol and registration As this was conducted as a scoping review with the purpose to map a rapidly growing field a formal review protocol was not considered needed. Eligibility criteria To be included in this review, papers needed to be written in English, published year 2000 or later, and were required to report on the use of AI or ML within the context of sports medicine. Peer-reviewed papers on all levels of evidence in accordance with the Oxford classification were included. All empirical study designs, including quantitative, qualitative, mixed-methods, case studies, and pilot studies, were considered with no restrictions with regards to cohort size or patient characteristics. Reviews, conference abstracts, commentaries (editorials, opinion pieces), system proposals (frameworks, protocols, datasets), articles without full-text availability, and pre-prints were excluded. System proposals were excluded, as the scope of this review was limited to studies that investigated the application of AI/ML within sports medicine. Information sources This study was conducted and presented in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) checklist [9]. A scoping review approach was performed due to the heterogeneity of AI methods, outcomes, and study designs, which made a quantitative synthesis and detailed risk of bias assessment less feasible. A literature search was planned in collaboration with and was executed out on 2026–02-05 by medical university librarians, with expertise in electronic database searching. The search strategies were peer reviewed by another senior medical university librarian prior to execution in accordance with the PRESS Checklist [10]. No additional manual searching of references list or grey literature was conducted after the primary database search. The literature search included the MEDLINE, EMBASE, and Web of Science databases, to provide comprehensive coverage of sports medicine research relevant to the clinical scope of this study. Search The initial search strategy was developed to identify studies which involved the use of AI or ML technologies in the context of sports medicine. The search targeted two core elements. First, the population: Individuals within a sports medicine context (athletes, physically active individuals or patients with or at risk of sports related injury). Second, the concept: Utilization of AI- or ML-based technologies, tools, or platforms intended to support the prevention, diagnosis, rehabilitation, or RTS decision-making within sports medicine. The broad search strategy was adopted to comprehensively map the diversity of AI and ML applications in sports medicine. The complete search strategy can be found in the Supplementary information. Selection of sources of evidence The screening process was performed with the Rayyan reference management platform (rayyan.ai) [11]. Screening of studies was performed by two independent reviewers (JL and KHT) starting with title and abstract screening followed by full-text evaluation. At all stages, any differences were resolved by mutual agreement. For title and abstract screening, agreement between the authors was assessed with Cohen’s Kappa coefficient. Inter-rater reliability for full-text screening was not formally quantified; disagreements were resolved through discussion and consensus between reviewers. Data charting process Draft data charting tables were developed in Microsoft Excel (Version 16, Microsoft Corporation, Remond, WA, USA) to record extracted data from the included articles, guided by the scoping review research statement. The first two authors (JL and KHT) performed the data charting process. Data items Study characteristics extracted included: bibliographic details (author(s), year of publication, and scientific journal), study characteristics (medical field, and country of origin), field of population (sport, sample size, age, study participant sex), AI-/ML application, data sources and input variables, output variables (outcomes), AI-/ML model(s) used, and key findings. Medical field classification was based on the primary clinical focus of each study. Studies were categorized as orthopaedics when the AI or ML application addressed musculoskeletal pathology, surgical outcomes, or rehabilitation related to orthopaedic conditions. The AI applications were categorized into four mutually exclusive application groups based on the primary aim of the AI component for each study. This categorization was not based on a single established or universally accepted classification framework; however, was pragmatically developed for the purpose of synthesizing the heterogeneous literature. Categories were defined a priori as follows: 1) predictive and estimation models, in which algorithms were primarily used to predict outcomes, risks, or clinical parameters; 2) classification and pattern recognition models, which focused on classification, clustering, feature extraction, or anomaly detection; 3) diagnostic and detection models, which aimed to identify diseases, conditions, or abnormalities; and 4) reasoning and summarization assistance models, where LLMs were central. When a study could plausibly fit more than one category, it was assigned to the group that best reflected the dominant role of the AI-/ML component. Critical appraisal of individual sources of evidence Due to the exploratory and descriptive nature of scoping reviews, and the heterogeneity of included studies, risk of bias and study quality assessments were not feasible and did not influence scoping review outcomes. Synthesis of results The results from the included studies were synthesized descriptively with a narrative approach, supported by tables and figures where appropriate [9]. Data charted from each study was grouped and summarized in accordance with key themes. These included: medical fields studied, AI-/ML applications, years and countries of publication, study populations, AI-/ML models used in general, with more detailed subgroup analyses performed for orthopaedics and neurology due to the higher number of included studies in these fields, and lastly, the overall model development stage. No meta-analysis or quantitative pooling was conducted due to the heterogeneity of study designs, AI-/ML approaches, and reported outcomes. Instead, findings were mapped to highlight the breadth of research activity, common areas of application, and gaps in the literature. Results In total, 8,677 records were identified, of which 97 were included (Fig. 1). For title and abstract screening, the Cohen’s Kappa coefficient was calculated to 0.877, which suggests near perfect agreement. Of the included studies, the medical fields represented were as follows: orthopaedics (n = 68, 70.1%), neurology (n = 18, 18.6%), radiology (n = 3, 3.1%), cardiology and cardiopulmonary (n = 2, 2.1%), nephrology (n = 1, 1.0%), odontology (n = 1, 1.0%), endocrinology (n = 1, 1.0%) and various (self-reported participation-restricting injuries [non-diagnosis-specific]; sports rehabilitation and digital health) (n = 3, 3.1%) (Table 1). Within orthopaedics, AI was primarily applied for injury prediction, outcome estimation, and rehabilitation monitoring, particularly concerning lower-extremity and ACL-related injuries. In neurology, models predominantly focused on SRC management, including classification of SRC severity and prediction of recovery duration. Radiology studies used AI for automated image optimization, detection of ligament injuries and bone marrow lesions, while the remaining medical fields involved isolated applications in injury-risk estimation (cardiopulmonary and cardiology), acute physiological responses (nephrology), dental injury prediction (odontology), and low bone mineral density (endocrinology). PRISMA flowchart for the inclusion process. AI = Artificial intelligence, ML = Machine learning, n = number Table 1Overview of included studiesAuthorYearJournalCountryMedical fieldField of populationSample size, nAge, mean years or rangeSex, males %AI applicationAI outcomeModel(s)Key findingsData usedAbasi et al.[12]2025BioData MininIranCardiopulmonaryElite football256NRNRPredictive & Estimation modelsPrediction of reinjury riskSVM, CatBoost, RF, XGBoostCatBoost; Acc: 0.9138, F1: 0.9148; SVM: AUC: 0.9725Cardiopulmonary dataAllen et al.[6]2023Journal of Neurosurgery: Pediatrics.USANeurologyPediatric athletes49315.768Predictive & Estimation modelsPrediction of early (≤14), typical (15–27), delayed (≥28) recovery time (days) from SRCDTAUC: 0.80, Youden: 0.44. Sen: >0.90 (Classified early recovery)Demographics, post-SRC symptom scales, time-to-clinic presentation, concussion history, presence of defined symptom clustersAoyag et al.[13]2021SpineJapanOrthopaedicsJunior high-school athletes22313.572Diagnostic & Detection modelsDistinguish lumbar spondylolysis from non-specific low back painCARTSen: 0.64, Spec: 0.92, AUC: 0.79Demographics, school grades, symptom onset time, history of lower-back pain, pre-existing conditions and anthropometryAyala et al.[14]2019International Journal of Sports MedicineSpainOrthopaedicsProfessional football96NR100Predictive & Estimation modelsPrediction of risk factors of hamstring injuryDTAUC: 0.837, Sen: 0.778, Spec: 0.838Preseason: psychological, neuromuscular, and demographical dataBazarian et al.[15]2021JAMA Network OpenUSANeurologyAthletes58019.554Classification & Pattern recognitionClassification of SRC based on EEGGenetic algorithmSen: 0.860, Spec 0.708, NPV: 0.901, PPV: 0.620, AUC: 0.89EEG, cognitive tests, symptom inventoriesBergeron et al.[16]2019Medicine & Science in Sports & ExerciseUSANeurologyHigh-school football2004NRNRPredictive & Estimation modelsEstimation of symptom resolvment after SRCNB, SVM, 5-nearest neighbours, DT, RF, MLP, radial basis function networkNB and RF with 100 or 500 trees: AUC: 0.656–0.742Symptom and recovery dataBriand et al.[17]2022Frontiers in Sports & Active LivingCanadaVarious (orthopaedics and neurology)Short-track speed skaters11210Predictive & Estimation modelsPrediction of injuryRFSen: 0.5, Spec 0.7Longitudinal: training load, physiological, neuromuscular, psychological well-being, heart rate variability and injury history dataCalderon-Diaz et al.[18]2023SensorsChileOrthopaedicsProfessional football110NR100Predictive & Estimation modelsPrediction of muscle injuryDT, discriminant methods, NB, SVM, KNN, NN, XGBoostXGBoost: Prec: 78%.Biomechanical and muscle performance dataCao et al.[19]2008IEEE Transactions on Neural Systems & Rehabilitation EngineeringUSANeurologyRugby and American football612044Classification & Pattern recognitionClassification of residual functional deficitSVMAcc: 77.1%, Sen: 80.0%, Spec: 75.0%EEGCastellanos et al.[20]2021Sports MedicineUSANeurologyUS Military cadettes156821965Predictive & Estimation modelsPrediction of SRC riskSVMAUC: 0.73Baseline demographic, clinical, cognitive and behavioral dataChen et al.[21]2022Computational & Mathematical Methods in MedicineChinaOrthopaedicsBasketball9352076Classification & Pattern recognitionClassification of thoracolumbar vertebral fractures (ABC)Deeplearning: Faster RCNNAcc: 86.4%, Cohen’s kappa: 0.850CT imagesChu et al.[22]2022Annals of Physical & Rehabilitation MedicineUSANeurologyYouth athletes65513.7 (male), 14.0 (female)55Predictive & Estimation modelsPrediction of SRC recoveryCatBoost, DT, elastic net, RF, XGBoost, TabNetCatBoost: AUC: 0.8 (males), and 0.78 (females)Preinjury risk factors, injury severity measures, post-SRC functional and symptom dataDandrieux et al.[23]2025BMJ Open Sport & Exercise MedicineFranceVariousTrack and field1123462Predictive & Estimation modelsInvestigate association between injury risk estimation and injury burdenNegative binomial regressionAUC: 0.63Longitudinal: training activity, psychological state, sleep quality, and self-reported injury statusDe la Fuente et al.[24]2023Science & Medicine in FootballChileOrthopaedicsFootball2122.50Classification & Pattern recognitionClustering to determine risk profiles based on biomechanical propertiesumap3 clusters of biomechanical propertiesBiomechanical datade Leeuw et al.[25]2022European Journal of Sport ScienceNetherlandsOrthopaedicsVolleyball1027100Predictive & Estimation modelsPrediction of overuse injuriesSubgroup discoveryJump load was an important predictor for 70% of playersLongitudinal: training load, subjective wellness reports and overuse symptom questionnairesDiCesare et al.[26]2020Annals of Biomedical EngineeringUSANeurologyFootball20160Classification & Pattern recognitionClassification of sub-SRC impact exposureXGBoostAcc: 83.5%Wearable sensor data, video-verified head impact recordings and MRIsDiniz et al.[7]2022Knee Surgery, Sports Traumatology, ArthroscopyVariousOrthopaedicsFootball20928.3100Predictive & Estimation modelsPrediction of level of match participationClustering, XGBoostXGBoost: AUC: 0.81Match participation and performance dataDiniz et al.[27]2024Knee Surgery, Sports Traumatology, ArthroscopyVariousOrthopaedicsFootball23626.6100Predictive & Estimation modelsCross-validation of identified ACL injury mentionsOpenAI’sGPT-4o miniSen: 88.4%, Spec: 99.3%Publicly available textual and database dataElkin et al.[28]2018Applied Clinical InformaticsUSAOrthopaedicsMixed4694450Diagnostic & Detection modelsDiagnosis of knee injuryBayesian and heuristic modelSpecificity-based Bayesian model significantly outperformed heuristic modelPatient-reported questionnaireEskofier et al.[29]2012Computer Methods in Biomechanics & Biomedical EngineeringCanadaOrthopaedicsRunners8041.1 (male), 36.0 (female)50Classification & Pattern recognitionClassification of participants with or without patellofemoral pain syndromeAdaBoostAcc: 100%.Biomechanical dataEvans et al.[30]2024PLoS ONEUKOrthopaedicsRugby3620.7100Classification & Pattern recognitionClassification of non-contact lower limb injuries risk factorsBayesian pattern recognition and assessed by means of: NB, J48 DT, SVM, KNNAUC 0.76 (severe non-contact lower limb), 0.70 (non-contact lower limb), and 0.71 (non-contact ankle)Longitudinal: training load, performance test results, musculoskeletal screening metrics and injury historyFarhadian et al.[31]2020BMC Sports Science, Medicine and RehabilitationIranOdontologyPediatric athletes35611.3 (injured), 10.6 (uninjured)54Predictive & Estimation modelsPrediction of dental injuryRFAcc: 89.3%Demographic and behavioral dataFerris et al.[32]2021American Journal of Sports MedicineUSANeurologyCollegiate athletes38819.963Diagnostic & Detection modelsDiagnosis of SRCAdaBoostIncreased diagnostic accuracy by 4.4% to AUC: 0.848, and increased Sen by 9%Multimodal concussion assessment dataFreitas et al.[33]2025PLoS ONEPortugalOrthopaedicsProfessional football3426.3100Predictive & Estimation modelsPrediction of non-contact injuries in footballersSVM, Feedforward NN, AdaBoostSVM: Acc: 74%, Sen: 71%, Spec: 74%Wearable GPS dataGarcia et al.[34]2019Journal of NeurotraumaUSANeurologyAthletes and military2456119.358Predictive & Estimation modelsPrediction of SRC levelsCARTSen: 91.07% to 97.40%Concussion assessment and demographic dataGaudet et al.[35]2019Journal of Science & Medicine in SportCanadaOrthopaedicsSwimmers and handball3421.70Classification & Pattern recognitionClustering to determine shoulder injury based on subjective outcomesK-mean clusteringSen: 86%, Spec: 100%, diagnostic OR: 229.67 (KJOC). Sen: 86%, Spec: 37%, diagnostic OR: 3.53 (CKQUEST)Functional performance and self-reported clinical assessment dataGiorgino et al.[36]2024DiagnosticsItalyOrthopaedicsNANANANAReasoning & summarization assistanceUse of an LLM for patient educationGoogle Bard & ChatGPT-3.5Both models show good promise in patient educationText-based conversation responsesGirard et al.[37]2025Knee Surgery, Sports Traumatology, ArthroscopyCanadaOrthopaedicsAdolescents with and without ACL injury13415.3 (ACL injured),13.8 (controls)30 (ACL injured)44 (controls)Classification & Pattern recognitionClassification of ACL injury statusDTEntire group: Acc 0.675, Sen 0.70, Spec 0.65, F1 0.684; Females: Acc 0.769; Males: Acc 0.533Biomechanical dataGoggins et al.[38]2022International Journal of Sports MedicineUKOrthopaedicsElite pathway cricket1718.20Predictive & Estimation modelsPrediction of injuryDT, RFDT: AUC: 0.66. RF: AUC: 0.72Longitudinal: training load and performance monitoring dataGultekin et al.[39]2025Knee Surgery, Sports Traumatology, ArthroscopyTurkeyOrthopaedicsNANANANAReasoning & Summarization modelsEvaluation of LLM-generated ACL surgery patient education responsesChatGPT-4o, DeepSeek R1Both high accuracy (3.9/4) and consistency (4/4); ChatGPT more comprehensive (4.0 vs 3.2, p < 0.001); DeepSeek clearer (3.9 vs 3.0, p < 0.001) and more readable (FKGL 8.9 vs 14.2; FRES 61.3 vs 32.7)Text-based conversation responsesGuo et al.[40]2025PeerJChinaOrthopaedicsCollegiate basketball10420.4100Predictive & Estimation modelsPrediction of ACL injury incidenceRF, SVM, XGBoost, LRRF: AUC 0.80; Accuracy 0.962; XGBoost AUC 0.79; Logistic regression AUC 0.76; SVM AUC 0.66Demographic, injury history, biomechanical and EMG dataHecksteden et al.[41]2023Science and Medicine in FootballGermanyOrthopaedicsProfessional football8824.6100Predictive & Estimation modelsForecasting non-contact time-loss injuriesGBoost, LRGBoost: CV AUC 0.61; Test AUC 0.62; without screening data AUC 0.56; without upsampling AUC 0.48Physical performance, clinical, injury history and daily training, recovery and exposure dataHenriquez et al.[42]2020Frontiers in Sports & Active LivingUSAOrthopaedicsStudent athletes12219.659Predictive & Estimation modelsPrediction of musculoskeletal injuryRFAcc: 79%Biomechanical, physical performance, demographic and injury history dataHopkingson et al.[43]2022European Journal of Sport ScienceVarious EuropeanOrthopaedicsRugby246NRNRClassification & Pattern recognitionClassification of injurious or no-injurious tacklesRFAcc: 0.919, Sen: 0.995, Spec: 0.525Video-derived tackle characteristicsHsu et al.[44]2022Journal of Human KineticsVariousNephrologyUltramarathon runners224495Predictive & Estimation modelsPrediction of acute kidney injurySVMSen: 90%, Spec 100%Baseline psychological, biochemical, and body composition dataHu et al.[5]2025Scientific ReportsCroatiaRadiologyMixed3064NRNRDiagnostic & Detection modelsDetection of ACL injuryCNN + modified political optimizerAcc: 96.496%, Sen: 99.767%, Spec: 98.557%MRIHuang et al.[45]2022Frontiers in PhysiologyChinaOrthopaedicsYouth basketball1616.60Predictive & Estimation modelsPrediction of lower extremity non-contact injuryFusion model, XGBoost, RFFusion model: Prec: 0.9932, recall: 0.9976, F2: 0.9967 (non-injured). Prec: 0.9317, recall: 0.9167, F2: 0.9171 (minimal LE NC). Prec: 0.9000, recall: 0.9000, F2: 0.9000 (mild LE NC)Longitudinal: training load, perceived well-being, psychological responses, physical performance metrics, and injury historyHuang et al.[46]2023Frontiers in PhysiologyChinaOrthopaedicsYouth basketball17150Predictive & Estimation modelsPrediction of lower limb non-contact injuryCost-sensitive NNAUC: 0.8590, Prec: 0.6360, recall: 0.8700, F2: 0.7980, Brier: 0.1020Physical fitness, physiological data: performance metrics, biochemical markers, physiological responses, and perceived exertionHwang et al.[47]2025Orthopaedic Journal of Sports MedicineSouth KoreaOrthopaedicsAthletes1132767Predictive & Estimation modelsPrediction of subjective function, symptoms, and psychological readinessGBoost, SVM, LR, DT, RFGBoost: AUC: 0.844, F1: 0.889 (Successful recovery of PASS, IKCD). RF: AUC: 0.835, F1:0.732 (PASS ACL-RSI)Isokinetic muscle strength and y-balance test results and patient reported outcomesHwang et al.[48]2024Digital HealthSouth KoreaOrthopaedicsAthletes1023074Predictive & Estimation modelsPrediction of return to sport after ACL reconstructionRF, GBoostRF: AUC: 0.952 (single leg hop), and 0.949 (Tegner activity scale). GBoost: AUC: 0.868 (single leg vertical hop)Physical performance data: balance, and isokinetic muscle strengthJacob et al.[49]2022Scientific ReportsIcelandNeurologyElite athletes5438.40Classification & Pattern recognitionClassification of SRCRF, GBoost, AdaBoost, SVM, MLPSVM: Acc: 95.5%EEG, EMG; heart rate, and center of pressure and concussion assessment scale (SCAT5)Jauhiainen et al.[50]2022American Journal of Sports MedicineVariousOrthopaedicsElite football and handball791210Predictive & Estimation modelsPrediction of ACL injurySVM linear and with imbalance handling, RF, L2-regularized LRLinear SVM: AUC: 0.63Preseason biomechanical and physical performance dataJauhiainen et al.[51]2020International Journal of Sports MedicineFinlandOrthopaedicsYouth basketball and floorball31416.0 (male), 15.4 (female)45Predictive & Estimation modelsPrediction of injury riskRFAUC: 0.63Baseline biomechanical and physical performance data and anthropometricsJia et al.[52]2022Computational Intelligence & NeuroscienceChinaOrthopaedicsGymnasts12615.30Classification & Pattern recognitionIdentification of injury through imagesFuzzy pattern recognition. NN.Identified injury situation through imagesImage data and biomechanical force analysisKarbalaie et al.[53]2026Journal of Sports SciencesSwedenOrthopaedicsMixed: Patients with ACL-R10725.2 (ACLR), 22.4 (controls)36Classification & Pattern recognitionClassification of high versus low fear of re-injuryCNN, LRCNN: Acc 75.6%, F1 0.6, MCC 0.52; 8.6% higher Acc compared to LRBiomechanical dataKolodziej et al.[54]2023Scandinavian Journal of Medicine & Science in SportsGermanyOrthopaedicsYouth elite football5617.2100Predictive & Estimation modelsPrediction of lower extremity injury riskLASSO. Leave-One-OutLASSO: AUC: 0.63, Sen: 35%, Spec: 79%Biomechanical, neuromuscular and postural control dataKunze et al.[55]2021Journal of Bone & Joint SurgeryUSAOrthopaedicsAthletes11183032Predictive & Estimation modelsPrediction of functional ImprovementENPLR, stochastic GBoost, RF, AdaBoost, NN, SVMENPLR: AUC: 0.77, intercept: 0.7, slope: 1.22, Brier: 0.14Clinical, demographic and radiographic registry dataKunze et al.[56]2021Orthopaedic Journal of Sports MedicineUSAOrthopaedicsMixed: Patients with ACL-R4422952Predictive & Estimation modelsPrediction of clinically meaningful improvement after ACL reconstructionStochastic GBoost, RF, NN, SVM, AdaBoost, ENPLRENPLR: AUC: 0.82, intercept: 0.10, slope: 1.15, Brier: 0.068Clinical and surgical registry dataLipps Lene et al.[57]2024Journal of Experimental OrthopaedicsFranceOrthopaedicsAthletes9621.9 (male), 21.1 (female)64Diagnostic & Detection modelsIdentification of participants with earlier knee injuryDT, MLP, XGBoostDT and MPL: AUC: 0.94, Acc: 0.95, Prec: 1.0, Recall: 0.88, F1: 0.93Biomechanical and psychological dataLópez-Valenciano et al.[58]2018Medicine & Science in Sports & ExerciseSpainOrthopaedicsProfessional football and handball132NR100Predictive & Estimation modelsPrediction of muscle injuryC4.5 DT, SimpleCart, ADTree, RandomTreeADTree: AUC: 0.747, Sen: 65.9%, Spec: 79.1%Preseason demographic, psychological and neuromuscular dataLövdal et al.[59]2021International Journal of Sports Physiology & PerformanceNetherlandsOrthopaedicsHigh-level middle- and long-distance runners74NR64Predictive & Estimation modelsPrediction of injuryXGBoostAUC: 0.724 (day), and 0.678 (week)Longitudinal training load data (GPS and subjective training feedback)Lu et al.[60]2022Orthopaedic Journal of Sports MedicineUSAOrthopaedicsElite basketball210326100Predictive & Estimation modelsPrediction of lower extremity muscle strain/injuryXGBoost, RF, NN, SVM, elastic net penalized LR, generalized LRXGBoost AUC: 0.840Longitudinal player performance and historical injury dataMartínez-Gramage et al.[61]2020SensorsSpainOrthopaedicsTriathletes1914.653Predictive & Estimation modelsPrediction of running injuryRFAUC: 0.8, Sen: 0.6, Spec: 0.8, NPV 0.7, Matthews correlation coefficient 0.4Biomechanical, neuromuscular, and injury incidence dataMaxin et al.[62]2024DiagnosticsUSANeurologyCollegiate football9320100Diagnostics & Detection modelsDiagnosis of acute SRCRF, KNN, SVM, LR (SMOTE)Post-SMOTE RF: Acc 91%, Sen 98%, Spec 86%, AUC 0.91, F1 0.92Smartphone-based quantitative pupillometryMcBee et al.[63]2024JMIR medical educationUSAVariousNANANANAReasoning & summarization assistanceLLM for interdisciplinary panel discussion on sports medicineChatGPT-4Reasonably pointed to various benefits such as 24/7 support, personalized advice, automated tracking, and remindersText-based conversation dataMurray et al.[64]2024Sports & HealthUSANeurologyStudent athletes4092056Classification & Pattern recognitionClassification participants with or without SRCLRSingle-task tests were slower in patients with SRCBiomechanical and behavioral performance data including cognitive task response ratesNechita et al.[65]2025DiagnosticsRomaniaCardiologyYouth athletes3127 to 17NRDiagnostic & Detection modelsDetection of cardiovascular injury riskRF, CNNRF: Acc: 97.87%, Sen: 75%, Spec: 98.3%, Prec: 98%Physiological ECG dataNolte et al.[66]2025Journal of Sports SciencesGermanyOrthopaedicsMixed: Patients with/without ACL injury54922.2 (male), 23.0 (female)67Predictive & Estimation modelsPrediction of participants being ACL-injured or notRFAUC: 0.90 (male), AUC: 0.92 (female)Isokinetic strength test dataNonnenmacher et al.[67]2025Bone & Joint OpenGermanyOrthopaedicsAthletes with periacetabular osteotomy23531.917Predictive & Estimation modelsPrediction of early RTS at 3 and 6 months after surgeryLR, Conditional inference treeEarly RTS associated with surgical approach, sport frequency, psychological factors, and pain; delayed RTS with male sex and older age.Preoperative demographic and patient-reported questionnaire dataNose-Ogura et al.[68]2025Physician and SportsmedicineJapanEndocrinologyAthletes61420.9 (development), 19.6 (validation)0Predictive & Estimation modelsPrediction of low bone mineral densityLASSODevelopment AUC 0.89; Validation AUC 0.74; Sensitivity 0.83; NPV 0.85Preoperative questionnaire and dual-energy X-ray absorptiometry dataOhlsen et al.[69]2025CureusUSAOrthopaedicsNANANANAReasoning & Summarization assistance modelsAgreement of LLM recommendations with clinical guidelines for ACL and meniscal injuriesChatGPT-4o, Gemini 2.5 ProChatGPT: 82% agreement, Gemini: 73% agreement; no significant difference between modelsText-based conversation dataOliver et al.[70]2020Journal of Science & Medicine in SportEnglandOrthopaedicsYouth elite football35514.3100Predictive & Estimation modelsPrediction of non-contact lower extremity injuryMultivariate LR, supervised learning DTDT: AUC: 0.663, Sen: 55.6%, Spec: 74.2%Preseason neuromuscular screening data and anthropometric measuresOzbek et al.[71]2025ArthroscopyTurkey / USAOrthopaedicsNANANANAReasoning & Summarization assistance modelsQuality assessment of LLM responses to hip arthroscopy patient questionsChatGPT 4.020/25 rated “excellent”; 5/25 “satisfactory”Text-based conversation dataPérez-Contreras et al.[72]2025Applied Sciences-BaselChileOrthopaedicsProfessional football4122.356Predictive & Estimation modelsPrediction of non-contact muscle injury riskLR, DT, KNN, RF, GBoost, NNKNN: Acc 87%, AUC 0.87; Gradient Boosting: Acc 84%, AUC 0.90; Logistic Regression AUC 0.50Preseason biomechanical and training load dataPiłka et al.[73]2023SensorsPolandOrthopaedicsFootball3624100Predictive & Estimation modelsPrediction of football injuryXGBoostPrec: 92.4%, recall: 96.5%, F1: 94.4%Training and match load dataQuinn et al.[74]2024ArthroscopyUSAOrthopaedicsNANANANAReasoning & summarization assistanceLLM to test quality of information with regard to ACL reconstructionChatGPT-4, GeminiChatGPT-4 and Gemini: Overall good ability to generate accurate and relevant responsesText-based conversation dataRossi et al.[75]2023Sport Sciences for HealthItalyOrthopaedicsElite football1824.7100Predictive & Estimation modelsPrediction of non-contact injury riskDT, GBoost, k-mean clusterAcc increased to 63% (15% improvement) after blood profile was added to workload-only modelsGPS-derived external workload and blood biomarker dataRichter et al.[76]2023Sports BiomechanicsNorwayOrthopaedicsElite football and handball822210Predictive & Estimation modelsPrediction of participants with previous-/future-/no ACL injuryDT, RF, discriminant analysis, NB, KNN, SVM, LR, NNCluster of models: Average AUC 0.62, Sen: 0.59, Spec: 0.58Biomechanical dataRobinson et al.[77]2022American Journal of Physical Medicine & RehabilitationUSANeurologyAthletes2732152Predictive & Estimation modelsPrediction of prolonged recovery after SRCDTAcc: 0.7636, Sen: 0.6429, Spec: 0.8889, PPV: 0.8571, NPV: 0.7059Symptom evaluation data (SCAT5)Rommers et al.[78]2020Medicine & Science in Sports & ExerciseBelgiumOrthopaedicsYouth elite football73411.7100Predictive & Estimation modelsPrediction of musculoskeletal injuryXGBoostAcc: 85%, Prec: 85%, recall: 85%Preseason anthropometric, motor coordination and physical performance dataRuddy et al.[79]2018Medicine & Science in Sports & ExerciseAustraliaOrthopaedicsAustralian football36223.2 (2103 season), 25.0 (2015 season)10Predictive & Estimation modelsPrediction of hamstring injuryNB, LR, RF, SVM, NNMedian of all 5 models: AUC: 0.58 (2013 season), 0.57 (2015 season)Preseason demographic, injury history and strength test dataRuiz-Pérez et al.[80]2021Frontiers in PsychologySpainOrthopaedicsElite Futsal13922.552Predictive & Estimation modelsPrediction of soft tissue injuryC4.5, Alternating DT, SVM with SMO, KNN, Instance-Based LearningVarious models: AUC: 0.701 to 0.767Preseason psychological and neuromuscular dataSaghafi et al.[81]2018Proceedings of SPIEUSANeurologyYouth and high-school football1229 to 18N/AClassification & Pattern recognitionClassification of white matter changes after head impact exposureCNNAUC: 85.71%, F1: 83.33%Neuroimaging and biomechanical dataSaglam et al.[82]2025BMC Medical Informatics & Decision MakingTurkeyOrthopaedicsNANANANAReasoning & Summarization assistance modelsComparison of GPT’s in clinical decision-makingGPT-4, GPT-3.5GPT-4 significantly outperformed GPT-3.5 (p < 0.001; Cohen’s d = 1.42); higher treatment and rehabilitation suitability (p < 0.001)Text-based conversation dataSchulc et al.[83]2024Orthopaedic Journal of Sports MedicineUSAOrthopaedicsProfessional athletes with ACL injury129NRNRDiagnostic & Detection modelsIdentification of ACL injury through video analysisRecurrent NNAUC: 0.88, F1: 0.63Video-derived biomechanical dataShibata et al.[84]2019Journal of Orthopaedic ScienceJapanOrthopaedicsPatients with ACL-R38623.753Predictive & Estimation modelsPrediction of quadriceps strength recovery 6 months after ACL-RDT, Stepwise multiple linear regressionPreoperative QSI, age, and pre-injury Tegner score predicted 6-month QSI; decision tree correctly classified 46.8% of casesPreoperative isokinetic quadriceps strength, demographic, clinical and intraoperative finding dataSong et al.[85]2022Wireless Communications & Mobile ComputingChinaOrthopaedicsTrack and field1219.5N/APredictive & Estimation modelsEvaluation of rehabilitation effectiveness using AI and virtual reality-assisted trainingProbabilistic NN, SVMAI+virtual reality group achieved > 96% physical function recovery; overall rehabilitation score 93.79 vs 82.38 (control)Physiological blood measures, functional, strength and speed assessment dataSparks et al.[86]2024JB & JS Open AccessUSAOrthopaedicsNANANANAReasoning & summarization assistanceLLM to investigate accuracy of patient education with regard to orthopaedic conditionsChatGPT-3.5Moderately accurate outputs for general inquiries. Lack in the quantity of information for risk factors and treatment options.Text-based conversation dataStirling et al.[87]2025Journal of Orthopaedic ResearchCanadaRadiologyPatients with ACL injury10033.632Classification & Pattern recognitionAutomated quantification of bone marrow lesion volume and association with pain outcomesCNNBone marrow lesions present in 95%; 96.1% volume reduction at 1 year (p < 0.001); baseline BML volume modestly associated with symptomsMRI and patient-reported questionnaire dataTamez-Peña et al.[88]2021Frontiers in NeurologyUnknownNeurologyStudent athletes12218.853Classification & Pattern recognitionClassification of SRCSVMSen: 0.80, Spec: 0.74Neuroimaging radiomic dataTedesco et al.[89]2020SensorsIrelandOrthopaedicsNon-elite rugby1226100Diagnostic & Detection modelsIdentification of gait patterns in participants with or without ACL injuryKNN, NB, SVM, GBoost, MLP, stackingMLP: Acc: 73.07; GBoost: Sen: 81.8%Inertial sensor dataThanjavur et al.[90]2021Frontiers in Human NeuroscienceCanadaNeurologyAdolescent athletes5813.4 (injured), 14.7 (uninjured)100Classification & Pattern recognitionClassification of SRCConcNet 2 and 3Acc: 94%, AUC: 0.971EEG dataTsilimigkras et al.[91]2024Journal of Sports Science & MedicineGreeceOrthopaedicsProfessional football25N/A100Predictive & Estimation modelsPrediction of muscle injury riskSVMAcc: 0.78, Sen: 0.73, Spec: 0.85Physiological and mechanical workload dataUsami et al.[92]2024Knee Surgery, Sports Traumatology, ArthroscopyJapanOrthopaedicsMixed: Patients with ACL-R38625.149Diagnostic & Detection modelsDetection of graft rupture and contralateral ACL injuryNNAUC: 0.81 (graft rupture), 0.74 (contralateral ACL injury)Clinical, demographic, and surgical medical record dataVallance et al.[93]2020Applied Sciences-BaselFranceOrthopaedicsElite football4029.4100Predictive & Estimation modelsPrediction of non-contact injury riskKNN, DT, RF, XGBoost, SVM, MLP, Linear discriminant analysis, LR, Ridge regression, NB1-month prediction: XGBoost AUC 0.97; 1-week prediction: questionnaires outperformed GPS data; internal load strongest short-term predictorGPS-derived external load, rating of perceived exertion and well-being questionnaire dataValle et al.[94]2022Sports MedicineSpainOrthopaedicsElite football7624.2100Predictive & Estimation modelsPrediction of recoveryLinear regression, RF, XGBoostXGBoost (days to recovery): Mean absolute error: 9.78884, Root mean squared error: 12.1450, R-squared: 0.4847Clinical and MRI dataVillarreal-Espinosa et al.[95]2024KneeUSAOrthopaedicsNANANANAReasoning & summarization assistanceLLM for patient education with regard to ACL surgeryChatGPT-45/10 responses completely accurate (by two reviewers), and 3/10 completely accurate (by at least one reviewer). Inter-rater reliability: weighted kappa: 0.57. 80% of responses were reproducible over timeText-based conversation dataWang et al.[96]2026Scientific ReportsVariousOrthopaedicsProfessional football31224.7100Predictive & Estimation modelsPrediction of non-contact lower extremity injuriesRF, SVM, GBoost, DNN, Ensemble modelEnsemble AUC 0.759Isokinetic strength, training load, injury history and biomechanical dataWeng et al.[97]2025Journal of Sports SciencesTaiwanOrthopaedicsVarious level baseball9818.0 (injured), 17.5 (uninjured)100Predictive & Estimation modelsPrediction of upper extremity injuryGIRD, LR, RF, CatBoost, SVMCatBoost: AUC: 0.66, Acc: 0.70Clinical and musculoskeletal dataYates et al.[98]2025BMJ Open Sport & Exercise MedicineEnglandNeurologyContact sport athletes37524.278Predictive & Estimation modelsPrediction of SRC recoveryRFAcc: 94.6%, Sen: 100%, Spec: 93.8%, PPV: 71.4%, NPV: 96.3%Clinical and MRI dataYe et al.[99]2023Frontiers in PhysiologyNetherlandsOrthopaedicsElite runners64N/A65Predictive & Estimation modelsPrediction of running injuryGASF-DCAE-DNNAUC 0.985, Gmean: 0.930, Sen: 0.997, Spec: 0.868. Test: AUC: 0.891, Gmean: 0.830, Sen: 0.816, Spec: 0.845Longitudinal training load, and physiological performance dataYüce et al.[100]2024CureusTurkeyOrthopaedicsNANANANAReasoning & summarization assistanceLLM for patient education with regard to sports surgeryChatGPT-4DISCERN: 44.75 points. Sports surgery-specific scoring: 13.3 pointsText-based conversation dataZhan et al.[101]2025ArthroscopyChinaOrthopaedicsMixed: Patients with MPFL-R218NRNRPredictive & Estimation modelsPrediction of clinical outcomes in patients with medial patello-femoral ligament reconstructionRF, LR, SVM, DT, implemented MLP, KNNVarious models: AUC: 0.760 to 0.969, and Acc: 76.8% to 95.2% (Subjective outcomes); AUC: 0.952, and Acc. 95.2% (Return to pre-injury sport); AUC: 0.756, and Acc: 75.4% (Return to pivoting sports); AUC: 0.943, and Acc: 94.9 (Recurrent instability)Clinical, demographic and radiographic dataZhan et al.[102]2023Journal of Sport & Health ScienceUSANeurologyMixed: Lab, MMA, American football, automobile, NASCAR3262NRNRClassification & Pattern recognitionClassification of head impact subtypesRFAcc: 96%Biomechanical data from head impact recordingsZhang et al.[103]2022Contrast Media & Molecular ImagingChinaRadiologyMixed: Patients with ACL injury903960Diagnostic & Detection modelsImage optimization to assess ACL integrityiDose4 Iterative Reconstruction Algorithm.Improved image qualityClinical CT dataZhu et al.[104]2026BMC Sports Science, Medicine and RehabilitationChinaOrthopaedicsPatients with ACL-R3031.9 (RTS group),36.9 (no-RTS group)73 (RTS group),60 (no RTS group)Classification & Pattern recognitionIdentification of urinary proteomic biomarkers associated with RTSLASSOAUC range 0.827–0.876Urinary proteomic, isokinetic strength, hop test, thigh circumference and patient-reported questionnaire dataZhu et al.[105]2026Journal of Clinical MedicineChinaOrthopaedicsPatients with ACL-R7930.989Predictive & Estimation modelsEvaluate effectiveness of a rehabilitation protocol incorporating an AI-based assessment and correction system on functional recoveryIntelligent Functional Movement and Physical Fitness Assessment System (ZD-200S-JG)Trial group showed significantly greater improvements in patient-reported outcomes and range of motion and rehabilitation adherenceNon-wearable three-dimensional motion capture, clinical and patient-reported questionnaire dataACL-RSI = Anterior cruciate ligament-return to sport after injury, Acc = Accuracy, AdaBoost = Adaptive boosting, AI = Artificial intelligence, AUC = Area under the receiver operating curve, CART = Regression tree analysis, CatBoost = Categorical boosting, CNN = convolutional neural networks, DT = Decision tree, CT = Computer tomography, ECG = electrocardiogram, EEG = electrocochleography, EMO = Electromyography, ENPLR = Elastic-net penalized logistic regression, F1 = F1-Score, F2 = F2-Score, GASF-DCAE-DNN = Gramian Angular Summation Field-Deep Convolutional Auto-Encoder-Deep Neural Network, GBoost = Gradient boosting, GPS = Global positioning system, GPT = Generative pre-trained transformer, IKCD = International Knee Documentation Committee, KNN = K-nearest neighbor, LASSO = Least Absolute Shrinkage and Selection Operator, LE = Lower extremity, LR = Logistic regression, MLP = Multilayer perceptron, MRI = Magnetic resonance imaging, NA = Not applicable, NB = Naïve Bayes, NN = Neural networks, NPV = Negative predictive value, NR = Not reported, OR = Odds ratio, PASS = Patient acceptable symptom state, PPV = Positive predictive value, Prec = Precision, RF = Random forest, RTS = Return to sport, Sen = Sensitivity, Spec = Specificity, SRC = Sports-related concussion, SVM = Support Vector Machine, Youden = Youden index Overview of included studies OpenAI’s GPT-4o mini 15.3 (ACL injured), 13.8 (controls) 30 (ACL injured) 44 (controls) 31.9 (RTS group), 36.9 (no-RTS group) 73 (RTS group), 60 (no RTS group) ACL-RSI = Anterior cruciate ligament-return to sport after injury, Acc = Accuracy, AdaBoost = Adaptive boosting, AI = Artificial intelligence, AUC = Area under the receiver operating curve, CART = Regression tree analysis, CatBoost = Categorical boosting, CNN = convolutional neural networks, DT = Decision tree, CT = Computer tomography, ECG = electrocardiogram, EEG = electrocochleography, EMO = Electromyography, ENPLR = Elastic-net penalized logistic regression, F1 = F1-Score, F2 = F2-Score, GASF-DCAE-DNN = Gramian Angular Summation Field-Deep Convolutional Auto-Encoder-Deep Neural Network, GBoost = Gradient boosting, GPS = Global positioning system, GPT = Generative pre-trained transformer, IKCD = International Knee Documentation Committee, KNN = K-nearest neighbor, LASSO = Least Absolute Shrinkage and Selection Operator, LE = Lower extremity, LR = Logistic regression, MLP = Multilayer perceptron, MRI = Magnetic resonance imaging, NA = Not applicable, NB = Naïve Bayes, NN = Neural networks, NPV = Negative predictive value, NR = Not reported, OR = Odds ratio, PASS = Patient acceptable symptom state, PPV = Positive predictive value, Prec = Precision, RF = Random forest, RTS = Return to sport, Sen = Sensitivity, Spec = Specificity, SRC = Sports-related concussion, SVM = Support Vector Machine, Youden = Youden index AI applications The AI application categories of the included studies were as follows: predictive and estimation models (n = 56, 57.7%), classification and pattern recognition models (n = 20, 20.6%), diagnostic and detection models (n = 11, 11.3%), and reasoning and summarization assistance models (n = 10, 10.3%) (Table 1). Within orthopaedics, predictive models were most common (43/68, 63.2%), whereas classification models were most common within neurology (9/18, 50.0%). Across AI applications, predictive and estimation models were mainly employed to forecast injury risk (n = 34), or recovery-related outcomes (n = 15), including RTS probability and functional improvement after orthopaedic injury/surgery. Classification and pattern-recognition models (n = 17) were primarily used to distinguish between injured and uninjured states, classify SRC or gait patterns, and identify biomechanical risk clusters. Diagnostic and detection models (n = 8) were mainly applied for image- or video-based injury identification, such as ACL or lumbar spine pathology. Reasoning and summarization assistance studies (n = 10) exclusively investigated LLM models (ChatGPT, Gemini, Bard, DeepSeek) for patient education and information quality assessment. Annual distribution and countries of published studies Only two studies were identified before 2018, while the number increased thereafter, and peaked in 2025 (20 publications, 20.6%) (Fig. 2 and Table 1). Of all included studies, 86.6% were published between 2020 and 2026. Of the included studies, most studies were conducted in the USA (n = 26, 26.8%), followed by China (n = 10, 10.3%), Canada (n = 6, 6.2%), Spain (n = 5, 5.2%), Japan, Germany, and Turkey (n = 4, 4.1%, respectively), Chile, France, and Netherlands (n = 3, 3.1%, respectively), Italy, South Korea, England, United Kingdom, and Iran (n = 2, 2.1%, respectively), and Australia, Ireland, Norway, Poland, Finland, Iceland, Croatia, Sweden, Romania, Belgium, Portugal, Taiwan, and Greece each had one (1.0%) contribution, while six (6.2%) studies were conducted across multiple countries, and one (1.0%) study had unknown origin. Annual distribution of published studies. For year 2026, only publications up until 5 of February were included Fields of population The most common specific fields of population were football (n = 26), followed by basketball (n = 6), rugby (n = 4), handball (n = 4), and runner-related populations (n = 4). Studies which included mixed athlete populations (n = 22) and a combination of athletes and non-athletes (n = 10) were also common. The distribution of study population across sport categories and levels of participation is presented in Fig. 3. Data environments differed by sport: football and running studies predominantly analyzed global positioning system (GPS)-based workload and accelerometry, whereas basketball, handball, and volleyball more often used laboratory biomechanical tests, balance/fatigue assessments, or physical performance analysis. Rugby studies included video-derived tackle characteristics. Sample sizes ranged from small elite squads (n = 10) to large youth/registry cohorts (n = 24,561), with prospective designs typical for load-monitoring studies and cross-sectional designs were common for imaging/electroencephalogram (EEG) tasks. Across sports, participant sex distribution varied across studies, and no studies specifically included para-sport populations. While many studies integrated multidimensional inputs that combined load, wellness/psychological, and clinical variables, relatively few integrated truly multimodal data that combined different modalities such as imaging, senosor-based signals, or video [15, 26, 40, 44, 49, 68, 75, 85, 87, 101, 104, 105]. Distribution of study populations across included studies. The inner ring represents sport, and the outer ring level of participation. n = number AI models The five most included models used overall were random forest (RF) (n = 29), support vector machine (SVM) (n = 27), decision tree (DT) (n = 20), neural network (NN) (n = 15), and XGBoost (n = 12) (Table 2). Among the three most common population groups, RF models were most frequently used in studies which involved general athlete populations (9/20, 45.0%) and mixed populations (4/7, 57.1%). In studies focused on football, SVM and XGBoost models were the most commonly used, each applied in 10 out of 25 studies (40.0%). Beyond these core algorithms, categorical boosting (CatBoost), adaptive boosting (AdaBoost), Light gradient boosting machine, and elastic-net penalized regression were frequent ensemble or regularized variations, and 14 studies explored newer architectures such as deep-learning models (TabNet, ConcNet, CNNs, feedforward NNs, recurrent NNs [RNNs], Faster region-based CNN [Faster R-CNN]) for imaging or signal-based classification. Model selection generally reflected data type: tree-based ensembles and SVMs dominated structured tabular datasets (e.g., sensor, GPS, clinical, or performance data), while CNNs and RNNs were used for imaging and video, and task-specific networks such as ConcNet for EEG. Unsupervised clustering (umap, k-means, subgroup discovery) appeared occasionally for identifying biomechanical or risk profiles. In total, 42 studies incorporated explainable AI methods (feature importance or Shapley additive explanation values). Ten studies investigated LLMs (ChatGPT, Gemini, Bard, DeepSeek) for patient education rather than prediction. Across studies, reported model performance was generally favorable, with accuracy ranging between 63 and 100%, area under the receiver operating characteristics curve (AUC) between 0.57–0.985, sensitivity between 35 and 100%, precision between 63.6%-100%, and F1-score between 0.6–0.944 depending on study design and outcome definition; in studies that compared approaches, ML models generally demonstrated higher discriminative performance than traditional regression methods [40, 53, 70, 72]. However, the model performance metrics were predominantly derived from internal cross-validation (n = 68), or no validation (n = 15) procedures, while external/temporal validation and reporting of calibration metrics were rare (n = 4), which underscores that while AI model performance was often strong, methodological maturity and reproducibility remain limited across sports medicine applications. Table 2Heatmap of AI-/ML model(s) used in each included study* Heatmap of AI-/ML model(s) used in each included study* Studies in orthopaedics Within orthopaedics (n = 68), AI was applied to three main areas: injury prediction, diagnostic imaging/video analysis, and outcome estimation after ligament reconstruction (Table 1). Common inputs included isokinetic knee strength, hop and balance tests, jump biomechanics, GPS-derived loads, clinical/radiographic fields, and MRI/computer tomography (CT) or match-video frames. Most studies (43/68, 63,2%) implemented predictive/estimation models, which typically used preseason screening (neuromuscular tests and anthropometrics), training-load/GPS and wellness logs, or clinical/surgical registry data to forecast ACL injury/reinjury, muscle strain, or RTS and functional recovery. Discrimination was generally poor to excellent (AUC 0.61–0.97, accuracy 65–98%), with tree-based ensembles (RF, XGBoost, gradient boosting) frequently achieved AUC ≥ 0.80 [7, 40, 43, 47, 60, 61, 78] and, in a few ACL-specific outcome models, approached 0.90–0.95 [48, 66]. Diagnostic/detection studies (9%) focused on imaging and video, for example, CNN or RNN models to identify ACL injury from online match footage or to distinguish lumbar spondylolysis from non-specific low-back pain, reported accuracies of 73–96%. Beyond imaging, several works modeled post-operative trajectories (e.g., graft rupture risk, subjective function, and psychological readiness), drawing on isokinetic strength, hop tests, y-balance tests and patient reported outcomes. A small subset evaluated LLMs for orthopaedic patient education (e.g., ACL reconstruction and sports surgery information quality), which represented methodological exploration rather than prediction. Taken together, the orthopaedic literature reports strong within-sample performance on structured, tabular data and promising results for image/video-based detection, however, external/temporal validation and diverse cohorts (including female athletes) remain limited, which constrained generalizability. Studies in neurology In neurology-focused research (n = 18), nearly all studies (17/18, 94.4%) examined SRC, that used models to predict recovery time or classify concussion presence/severity (Table 1). Inputs were predominantly clinical and neurocognitive assessments (e.g., sport concussion assessment tool 3/5, vestibular ocular motor screening, symptom inventories, prior SRC, and time-to-clinic), sensor-based biomechanics (head-impact kinematics, and dual-task gait/behavioral performance), and physiological signals including EEG (resting-state and task-based) and radiomics from MRI/diffusion tensor imaging, where several studies combined these into multimodal feature sets [15, 26, 32, 49, 81, 98]. Reported performance was moderate to excellent, with AUC 0.70–0.96 and accuracy 75–96%. For recovery-time prediction, decision-tree and boosting approaches achieved AUC ~0.80 and sensitivity > 0.90 for early-recovery classification in pediatric and youth cohorts [6, 22]. For SRC classification, SVM/RF/boosting models commonly reached AUC ≥ 0.80 [32, 49, 62], while deep-learning on EEG (e.g., ConcNet) reported the highest accuracy (94%) and AUC (0.971) [90]. Overall, neurology applications show strong within-sample discrimination across both predictive and classification use-cases and growing interest in multimodal modeling. However, external/temporal validation, standardized feature sets, and explainability remain limited, which constrained generalizability beyond single-site cohorts. Model validation and translational stage Across all included studies, the majority relied on retrospective datasets or observational prospective and internal validation procedures, most commonly cross-validation or train-test splits. Only four studies employed external datasets to assess generalizability, and reporting of calibration metrics was uncommon [15, 68, 76, 96]. Prospective implementation within preventive or decision-making workflows was rare, with only one study attempting a feedback-based interventional approach [23]. Two studies evaluated AI-integrated rehabilitation systems within structured training programs, which included one randomized controlled trial in postoperative ACL rehabilitation [105] and one application which combined AI and virtual reality for athlete rehabilitation training [85]. Discussion This scoping review provides a synthesis of AI applications across sports medicine, and demonstrates that, although methodological development has accelerated in recent years, most studies remain in an early developmental stage. Artificial intelligence has been widely applied for injury prediction, diagnostic imaging, and recovery estimation across diverse athletic and clinical populations, mostly within orthopaedics and neurology. Despite frequently high reported performance metrics, the literature is characterized by substantial heterogeneity in model selection, data modalities, outcome definitions, and validation procedures. Most studies relied on retrospective or observational prospective datasets and internal validation methods, whereas external or temporal validation and prospective prevention or intervention frameworks were rare. Consequently, the current evidence base does not yet support routine clinical integration of AI-driven decision tools in sports medicine. In practice, AI models should currently be interpreted as adjunct tools to support, rather than replace, clinician judgement in injury risk assessment, diagnosis, and rehabilitation planning. Although many models demonstrated strong discriminative performance, which often achieved AUC values ≥ 0.80, these findings must be interpreted in the context of important methodological limitations. High within-sample accuracy suggests that AI can effectively identify patterns associated with injury risk, recovery trajectories, or RTS potential. However, the vast majority of studies performed validation within the same dataset, typically through internal cross-validation or train-test splits. Only four studies employed external datasets to assess generalizability [15, 68, 76, 96]. This represents a critical methodological limitation, as internal validation tends to overestimate model performance and fails to account for differences in population characteristics, data collection methods, or sporting environments [106]. In addition, the prediction timeframes across included studies varied, with short-term models typically based on repeated within-subject measurements (e.g., training load or wellness data) and longer-term models relying on baseline variables, which introduces heterogeneity in model design and clinical applicability. Without robust external or temporal validation, the true predictive value and clinical reliability of these models remain uncertain. Accordingly, the literature reflects predominantly early-stage model development, with limited progression toward external validation, prospective integration, or demonstrated impact on clinical decision-making. Future studies should prioritize multi-center external validation across teams, seasons, and demographic groups, to evaluate whether reported performance translates into meaningful clinical utility. Clinically, the uncertainty of findings limits the use of AI models for individual-level decision-making, such as identification of athletes at high risk of reinjury or determination of readiness for RTS, where reliable generalization across populations is important. While AI-driven prediction models have been reported with strong retrospective accuracy, only Dandriex et al. [23] attempted to integrate predictions into a prospective, feedback-based prevention strategy, in which daily individualized feedback was provided to track-and-field athletes based on self-reported wellness data. However, adherence was low (average daily response rate of 37%), and no significant association with injury burden was observed, although a modest protective effect was suggested among participants with at least 9% response rate [23]. In addition to predictive frameworks, a small number of studies have integrated AI directly into structured rehabilitation programs. For example, one randomized controlled trial evaluated an AI-based assessment and correction system after ACL reconstruction [105], which demonstrated improvements in functional outcomes and rehabilitation adherence, while another study applied AI and virtual reality technology within athlete rehabilitation training [85]. These findings illustrate both the potential and the current limitations of AI-supported frameworks, namely, information alone does not yield benefit unless it is coupled with consistent athlete engagement, integration into clinical or training workflows, and actionable feedback mechanisms capable of influencing real-word decision-making, such as modification of training load, or informing RTS clearance. Sports injuries are inherently multifactorial, which arise from complex interactions between training load, biomechanics, physical fitness, psychological status, and contextual factors such as playing surface and competition demands [107, 108]. Consequently, models that rely solely on single wellness or workload metrics are unlikely to capture the full spectrum of injury risk. This highlights the need for multimodal data integration that combines these factors to better reflect the complexity of athletic performance and health [109]. Recent work demonstrates that combination of factors such as genetics, biomechanics, nutrition, and training load can improve injury prediction, although the approach was represented within a single tabular data modality rather than true multimodal frameworks [110]. In the context of sports medicine, models intended to inform RTS decisions or reinjury risk estimation must also be interpretable, to allow stakeholders to understand which variables drive predictions and how they align with established clinical reasoning [109], particularly in clinical settings where practitioners must justify decisions related to rehabilitation progression or RTS clearance based on transparent and clinically interpretable information. The growing use of high-performing yet opaque black-box models, such as DNNs, poses a barrier to practical implementation, as limited explainability can erode user confidence and hinder clinical decision-making [111]. Accordingly, development of AI frameworks that balance predictive strength with transparency will be crucial to support actionable and trustworthy decision support. Future research should extend beyond prediction to design and evaluate controlled, prospective, decision-driven systems that integrate real-time recommendations and test whether AI-supported interventions can meaningfully reduce injury incidence or improve recovery outcomes [112]. Overall, studies that evaluate the capabilities of LLMs in sports medicine suggest that these tools can generate generally accurate and informative responses, particularly for patient education. However, two key concerns were identified. Firstly, Sparks et al. [86] reported that LLM outputs often lacked sufficient detail with regard to risk factors and treatment options. Secondly, Villarreal-Espinosa et al. [95] found that two of ten responses about ACL surgery did not reach a very accurate rating, and that 80% responses were reproducible over time. In addition, LLMs are inherently susceptible to hallucinations, distribution shifts between training data and real-world use, and potential instability of output, all of which may further undermine their reliability in clinical athletic contexts [113–115]. These findings highlight the potential risk that patients may receive incomplete, inaccurate, or inconsistent information. Taken together, LLMs have serve as accessible adjunct educational tools, but current evidence does not support unsupervised clinical deployment. At present, their use should remain supervised by clinicians, particularly when addressing diagnosis, surgical decision-making, or RTS guidance. Future research should focus on benchmarking LLM performance against verified clinical standards and explore how these tools can be safely embedded into patient communication and rehabilitation pathways without compromising information accuracy. Future work should explore fine-tuning and retrieval-augmented generation to incorporate domain-specific knowledge and possibly improve clinical utility and reliability. Image and diagnostic AI applications showed promise in accuracy for identification of ligament injuries, bone fractures, and concussion-related imaging patterns. These approaches leverage automated feature extraction from radiological or video data, which reduces reliance on manual interpretation [116]. However, most studies were limited by small sample size, and lack of external or multi-center validation, thus restricting clinical applicability [117]. Furthermore, image-based models often function as black boxes, with limited explainability of which image features drive decisions, which underscores the need for interpretable visualization methods to support clinical implementation. In practice, AI applications may assist clinicians in improving diagnostic accuracy and reducing inter-observer variability, although current limitations restrict independent use in clinical decision-making. Across studies in which ML models were used alongside traditional regression approaches [40, 53, 70, 72], ML algorithms demonstrated higher discriminative performance. However, these studies were limited to within-sample or internally validated analyses, and formal benchmarking with strict comparisons between methods was rare. Consequently, although ML models may offer incremental gains in discrimination, it remains unclear whether these improvements translate into superior calibration, generalizability, or meaningful enhancement of clinical decision-making. Future research should therefore include structured comparative evaluations that assess not only discrimination but also calibration, interpretability, and net clinical benefit to determine whether increased algorithmic complexity provides practical advantages over established biostatistical methods. Across studies, substantial model heterogeneity was observed, with each investigation employing diverse algorithms, feature sets, and outcome definitions. This diversity underscores the exploratory nature and limits consensus on optimal model families or architectures for specific sports or data types [118]. In parallel, the generalizability of existing models remains uncertain, as most datasets were small, and single-site, with minimal or no inclusion of youth, or para-sport athletes and limited representation outside Europe and North American leagues [106]. These sampling biases restrict the broader applicability of reported findings. Furthermore, the validity of any AI model ultimately depends on the quality of data input [119]. Inconsistent data collection methods, and missing contextual variables can all undermine predictive accuracy, regardless of algorithmic sophistication [119]. Collectively, this methodological variability reinforces that the field remains in developmental phase, in which foundational issues of data standardization, and reporting transparency must be addressed before reliable large-scale implementation can be achieved. Addressing these foundational issues through standardized data handling, and transparent reporting will be important to build robust, and trustworthy AI systems in sports medicine. This review indicates that most current applications of AI and ML in sports medicine remain in an early developmental exploratory stage and should be interpreted cautiously in clinical settings. Consequently, these models should not yet be relied upon to independently guide clinical decision‑making, diagnosis, prognosis, or RTS recommendations. Instead, the current role of AI and ML models lies in supporting clinicians by synthesizing complex data patterns that may not be readily identifiable through traditional assessment alone. At present, many models rely on retrospective or prospective observational datasets, lack external validation, and have not undergone evaluation in real‑world clinical workflows, which limits their immediate clinical reliability and generalizability. Furthermore, the majority of the included studies focused on orthopaedics or neurology, therefore, the findings observed in these areas may not be directly applicable to other domains within sports medicine. The main contribution of this review was to provide clinicians and researchers with a clearer understanding of the current maturity and limitations of AI and ML technologies within the field. By synthesizing the available evidence, this work aims to support more informed, critical appraisal of algorithm‑generated outputs and to help clinicians recognize when such tools may complement rather than replace clinical reasoning. Furthermore, the review outlines priority areas in which rigorous model validation, prospective study designs, bias assessment, and implementation research are important. Addressing these gaps will be crucial to ensure that future AI‑ and ML‑based tools can transition safely and effectively from experimental settings to practical, patient‑centered clinical application. This study has several limitations. The included studies varied widely in sport, level of play, sample size, data modality, outcomes, and performance metrics which did not allow for a quantitative synthesis and limits comparison across models. Therefore, it was determined to synthesize narratively to emphasize patterns rather than pool effects. In addition, due to the heterogeneity of included studies, no risk of bias or quality assessment was performed. Moreover, the predominance of internal validation limits the generalizability of findings. Thus, results must be interpreted cautiously. Furthermore, publication bias might skew the picture of the accurate results of included studies, and true performance of AI models might not be captured. Given the rapid growth of this field, relevant studies may have been published after our search and were therefore not captured. Engineering-led studies focused solely on algorithmic development without application to human data in a sports medicine context were excluded; therefore, the technical foundation of AI methods may be underrepresented. To organize this heterogeneous literature, AI applications were pragmatically grouped into four mutually exclusive categories. While this approach facilitates synthesis, it simplifies the inherently multidimensional nature of AI systems, which often vary substantially in data sources, modelling approaches, and validation strategies within the same clinical task. Accordingly, the categories should be interpreted as heuristic groupings to support interpretation rather than rigid distinctions between fundamentally different AI methodologies. Given the large number of included studies, it was not feasible to verify the reproducibility of individual models or reported performance metrics; limited availability of code, data, and methodological detail further restricted independent validation and highlights the need for improved transparency in AI- and ML-based sports medicine research. Conclusion The use of AI applications in sports medicine demonstrate strong within-sample discriminative performance for injury risk, recovery, and diagnostic imaging, yet most remain limited to retrospective analysis with limited external validation and minimal evidence of clinical workflow integration. This review shows that the field is characterized by substantial methodological heterogeneity and limited progression toward prospective implementation. For clinicians within sports medicine, current AI tools should therefore be regarded as exploratory decision-support adjuncts rather than implementation-ready systems. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements The authors thank librarians Kajsa Magnusson and Ann Liljegren for valuable advice and performing of the literature searches. The authors also thank librarian Ida Stadig for her assistance in developing the AI-related search strategy. All librarians are affiliated with the Medical Library, Sahlgrenska University Hospital. Author contributions Jakob Lindskog drafted the initial version of the manuscript, performed screening and synthesis of data, has approved the final work for publication, and has agreed to be accountable for all aspects of the work. Kristian Heder Ternell has contributed majorly during the drafting of the manuscript, performed screening and synthesis of data, has approved the final work for publication, and has agreed to be accountable for all aspects of the work. Yinan Yu, Ida Lindman and Kristian Samuelsson have contributed during the interpretation of data, made meaningful contributions during the final stages of manuscript drafting, has approved the final work, and has agreed to be accountable for all aspects of the work. Eric Hamrin Senorski has contributed majorly during the drafting of the manuscript, analysis and interpretation of data, is responsible for the design concept, has approved the final work, and has agreed to be accountable for all aspects of the work. Funding Open access funding provided by University of Gothenburg. This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability This scoping review is based on previously published studies. The data-charting spreadsheet generated during the review is available from the corresponding author on reasonable request. Declarations Ethical approval Ethical approval was not required for this study as it involved the analysis of previously published literature Consent for publication Not applicable. Competing interests Author Kristian Samuelsson reports a relationship with Getinge AB that includes: board membership. No other competing interests to declare. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement This scoping review is based on previously published studies. The data-charting spreadsheet generated during the review is available from the corresponding author on reasonable request.