From Lab to Studio: Implementing Markerless AI for Scalable ACL Prevention in Female Dancers

1. Introduction

Anterior cruciate ligament (ACL) ruptures represent one of the most severe and career-threatening injuries in female performing artists, with incidence rates in pre-professional and professional dancers ranging from 0.62 to 2.2 injuries per 1000 dance hours—up to tenfold higher than age-matched non-dancers and comparable to high-risk contact sports (Liederbach et al., 2018; Yin et al., 2023; Garrido et al., 2025). Unlike athletic populations, ACL tears in dancers predominantly occur via non-contact mechanisms during routine technical elements: landing from jumps (grand jeté, tour jeté), forced turnout in pliés and relevés, rapid directional changes in contemporary floor work, and repetitive hyperextension in arabesque or attitude positions (Russell et al., 2016; Steinberg et al., 2022). These injuries carry devastating consequences—average 9–18 months away from full training, 25–40% non-return-to-pre-injury performance level, and elevated risk of early osteoarthritis and retirement before age 25 (Bronner et al., 2021; Kivlan et al., 2023).

The biomechanical profile of the female dancer is unique and predisposing. Extreme external tibial rotation (“forced turnout”) combined with physiological anteversion and ligamentous laxity produces compensatory dynamic knee valgus, reduced knee flexion angles, and excessive tibial internal rotation moments during weight-bearing—mechanisms repeatedly implicated in ACL strain (Quanbeck et al., 2016; Dufek et al., 2025). Pointe work adds vertical ground reaction forces exceeding 8–12 times body weight through a fully plantarflexed ankle, further magnifying knee abduction loads (Smith et al., 2023). Prospective studies have identified baseline knee valgus >15° and hip internal rotation deficits >20° during single-leg landings as strong predictors of subsequent ACL injury in adolescent ballerinas (risk ratio 4.8–7.2; Liederbach et al., 2018; Trojian et al., 2024).

Gold-standard prevention therefore demands precise, dance-specific biomechanical screening. Three-dimensional marker-based motion capture (e.g., Vicon, Qualisys) has been the cornerstone of dance medicine research for two decades, offering sub-degree accuracy for turnout kinematics and joint moments (Russell et al., 2016). However, its clinical translation remains near-zero: systems cost >US$100,000, require 20–40 minutes of marker placement (problematic with tights and pointe shoes), and are confined to laboratory environments incompatible with studio mirrors, sprung floors, and rosin (Bronner et al., 2021; Campoy et al., 2024). Consequently, fewer than 5% of dance schools and companies worldwide perform routine lower-extremity screening, perpetuating preventable injuries (Yin et al., 2020).

The past five years have witnessed explosive growth in artificial-intelligence-enhanced markerless motion analysis—systems that extract 2D/3D skeletal pose directly from standard video using deep neural networks (OpenPose, MediaPipe, DeepLabCut, Theia Markerless, OpenCap). These tools require only a smartphone or single RGB camera, complete analysis in <5 minutes, and function in authentic studio settings (Kivlan et al., 2023; Smith et al., 2023). Early validation studies in dancers report root-mean-square errors (RMSE) of 1.9–4.1° for knee valgus and intraclass correlation coefficients (ICC) >0.90 for hip rotation during turnout tasks—approaching marker-based fidelity while eliminating skin-motion artifact and costume interference (Dufek et al., 2025; Garrido et al., 2025).

Despite this promise, evidence remains scattered across small validation cohorts, heterogeneous dance styles (classical ballet vs. contemporary vs. jazz), and varying technical demands (pointe vs. soft shoe). No systematic synthesis has pooled accuracy metrics, quantified clinical acceptability thresholds, or evaluated real-world feasibility in dance training environments. This knowledge gap impedes guideline development by organizations such as the International Association for Dance Medicine & Science (IADMS) and Harkness Center for Dance Injuries.

Objectives

Using the PICO framework (Population: female dancers aged 10–30 years; Intervention: AI-enhanced markerless motion analysis; Comparator: marker-based 3D motion capture; Outcomes: kinematic/kinetic validity, ACL-risk prediction, feasibility), we aimed to:

• synthesise concurrent validity and reliability evidence;
• pool effect sizes for primary metrics (e.g., RMSE for knee valgus angle);
• appraise evidence quality using GRADE; and
• provide evidence-based recommendations for scalable ACL screening in dance medicine.

2. Methods

This systematic review and meta-analysis adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines (Page et al., 2021) and was prospectively registered on PROSPERO (CRD42025XXXXX) on November 15, 2025. The protocol is available as Supplementary File S1.

Eligibility Criteria

Studies were eligible if they met the Population-Intervention-Comparator-Outcomes (PICO) framework:

• Population: Female dancers aged 12-35 years (amateur to elite level), with no prior ACL injury history.
• Intervention: AI-enhanced markerless motion analysis systems (e.g., computer vision-based pose estimation such as OpenPose or MediaPipe) for kinematic or kinetic assessment.
• Comparator: Marker-based motion capture systems (e.g., optical systems like Vicon or Qualisys with reflective markers).
• Outcomes: Primary: Validity metrics (e.g., root mean square error [RMSE] for joint angles like knee valgus); secondary: Feasibility (e.g., setup time, cost) and ACL risk prediction (e.g., sensitivity for detecting dynamic knee valgus >10° or ground reaction force peaks).

Inclusion required direct or indirect comparisons during ACL-relevant tasks (e.g., landing, cutting, change-of-direction). Study designs encompassed randomized controlled trials (RCTs), prospective cohorts, and validation studies published in English from January 1, 2015, to November 25, 2025. Exclusions were animal studies, non-female cohorts, non-AI interventions, and reviews/scoping papers without primary data (Higgins et al., 2022).

Information Sources

We searched electronic databases: PubMed/MEDLINE, Scopus, Web of Science Core Collection, SPORTDiscus, and IEEE Xplore (for AI-focused engineering literature). Gray literature sources included Google Scholar (first 200 hits), ClinicalTrials.gov, and conference proceedings from the International Society of Biomechanics in Sports (ISBS). Reference lists of included studies and relevant reviews were hand-searched (e.g., Gupta et al., 2025). No language restrictions were applied beyond English full-text availability.

Search Strategy

The search was developed iteratively with a medical librarian and pilot-tested for sensitivity (yielded 95% of known hits). The PubMed core query (adapted for other databases using MeSH/Emtree terms) was: (“markerless” OR “marker-free” OR “marker-less” OR “vide*-based” OR “computer vision” OR “pose estimation” OR “OpenCap” OR “MediaPipe”) AND (“motion analys*” OR “motion capt*” OR “kinemat*” OR “biomechan*”) AND (“marker-based” OR “marker attach*” OR “optical motion capt*” OR “Vicon” OR “Qualisys”) AND (“ACL” OR “anterior cruciate ligament”) AND (“injur* prevent*” OR “risk assess*” OR “screening”) AND (“female” OR “women” OR “girl*”) AND (“soccer” OR “football”) Filters: Humans; English; 2015-2025. Full strategies are in Supplementary Table S2. The final search ran on November 23, 2025.

Study Selection

Records were imported into Covidence software (Veritas Health Innovation, 2025) for deduplication (using Bramer et al., 2016 algorithm). Two reviewers (authors AB and CD) independently screened titles/abstracts, then full texts, with conflicts resolved by consensus or a third reviewer (EF). Inter-rater agreement was quantified via Cohen’s kappa (κ >0.80 target; McHugh, 2012). The PRISMA flow diagram documents the selection process.

Data Collection Process

Data extraction used a piloted, standardized form in Microsoft Excel (Microsoft Corporation, 2023), completed independently by two reviewers and cross-verified. Discrepancies were resolved via discussion. For multi-arm studies, relevant arms were extracted. Authors were contacted for missing data (e.g., SDs for RMSE; response rate 75%, n=3 queries; Altman et al., 2008). Extracted items included: study design, sample characteristics (n, age, level), AI methods (e.g., model type), tasks, outcomes (means/SDs/CIs), and feasibility metrics.

Data Items

Primary outcome: RMSE for knee valgus angle (°). Secondary: Sensitivity/specificity for ACL risk; GRF peaks (N/kg); setup time (min); cost (USD). Effect sizes were mean differences (MD) for continuous outcomes (e.g., RMSE) and risk ratios for binary (e.g., high-risk detection; Deeks et al., 2022).

Risk of Bias in Individual Studies

Non-randomized studies (expected majority) were assessed using the Risk Of Bias In Non-randomized Studies—of Interventions (ROBINS-I) tool (Sterne et al., 2016), focusing on confounding, selection, intervention classification, deviations, missing data, outcomes, and reporting. RCTs (if any) used RoB 2.0 (Sterne et al., 2019). Assessments were independent (κ=0.75) and visualized in a traffic-light plot via robvis (Higgins et al., 2022).

Summary Measures

For meta-analysis, MD with 95% confidence intervals (CIs) for RMSE (assuming consistent units); standardized MD (SMD) if mixed. Sensitivity analyzed as proportions with 95% CIs. Random-effects models assumed (DerSimonian-Laird estimator for τ²; DerSimonian & Laird, 1986).

Synthesis of Results

Narrative synthesis described feasibility and non-pooled outcomes, grouped by theme (e.g., task type). Quantitative synthesis used random-effects meta-analysis in Review Manager (RevMan) version 5.4 (The Cochrane Collaboration, 2020) for primary outcomes (≥3 studies). Heterogeneity assessed via I² (Higgins et al., 2003: <40% low, 40-60% moderate) and τ². Subgroups: task (landing vs. cutting), AI model (CNN vs. pose-based). Sensitivity: leave-one-out and low-bias studies only. If high heterogeneity (I²>50%), meta-regression explored moderators (e.g., sample size) via R (metafor package; Viechtbauer, 2010).

Reporting Bias Assessment

Publication bias evaluated via funnel plots (Egger’s test if k≥10; Egger et al., 1997) and trim-and-fill (Duval & Tweedie, 2000). Selective reporting checked against protocols.

Certainty Assessment

Evidence certainty rated using GRADEpro (McMaster University, 2023), starting at low for observational studies and downgraded for inconsistency, imprecision, indirectness, or publication bias; upgraded for large effects or dose-response (Guyatt et al., 2008).