Journal DOI 10.68956/journal.human-performance-futures

Creative Commons source reading · Return-to-sport readiness

Predictive accuracy of machine learning and markerless gait analysis for return-to-sport following lower extremity injury: a systematic review and meta-analysis

Satija R; Dkhar D

Originally published in Front Sports Act Living · 2026-06-16

From the article

Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.

Abstract

AbstractBackground and objectiveAnterior cruciate ligament and lower extremity injuries impose a substantial burden in sports medicine, yet conventional assessments fail to capture the dynamic nature of injury risk and recovery. Machine learning and markerless gait analysis offer potential improvements in predicting return-to-sport readiness and re-injury risk, but their clinical reliability remains uncertain. This review evaluates the predictive accuracy and clinical applicability of these approaches.MethodsA comprehensive search of PubMed, Embase, Scopus, IEEE Xplore, and CINAHL was conducted up to March 2026. Studies involving adults with lower extremity injuries using machine learning or markerless systems and reporting predictive metrics were included. Pooled accuracy estimates were calculated using a fixed-effect model with…

Participant characteristics and study findings

Across the 11 included studies, sample sizes ranged from 12 to 500 participants, encompassing predominantly athletic and physically active populations with lower extremity injuries. Studies were stratified by injury type: ACL injury (7 studies), chronic ankle instability (2 studies), adductor strain (1 study), and general lower limb injuries (1 study). Outcome type was categorised as return-to-sport readiness (6 studies) or re-injury risk prediction (5 studies). Technology type was clearly distinguished: wearable sensors (9 studies) vs. markerless motion analysis (2 studies). Most participants were young adults, with several studies focusing on post-ACLR cohorts, while others included runners, soccer players, and individuals with chronic ankle instability. All studies employed observational or prediction model designs, with no…

The validation gap and its implications for clinical trust

A critical finding of this review is the near-total absence of external validation across included studies. Only one study attempted external validation by applying a previously developed algorithm to a new cohort, resulting in dramatic performance degradation (AUC dropping from 0.94 to 0.53–0.62) (2). This validation gap severely limits clinical trust and generalizability. Internal validation techniques varied widely: leave-one-subject-out cross-validation produced conservative accuracy estimates (0.66–0.73), while standard k-fold cross-validation inflated performance (0.91) due to data leakage (11). Nested cross-validation frameworks demonstrated better generalization but were rarely implemented (11). SMOTE oversampling improved F1-scores by 0.13 in small-sample studies but may introduce synthetic data artifacts (10, 23). Without…

Editorial reading guide

  • How were injury, recovery and exposure defined and verified?
  • Which criteria support progression, and what finding should stop or change the plan?