ACLGuard: Automated Screening of Sports-Related ACL Injury Risk Using Multimodal, Physics-Informed Machine Learning and 3D Pose Estimation

Abstract

ACL injury is a serious problem in modern sports medicine, but traditional biome- chanical screening is limited by the cost of laboratory motion capture and the sub- jectivity of clinical observation. Current deep learning methods for ACL diagnosis are also inherently reactive, based on post-injury MRI data, rather than preventative analysis of movement. In this thesis, we propose a novel computational framework, ACLGuard, for the automated screening of ACL injury risk from standard monocu- lar video. Our framework is based on the LESS assessment protocol, which does not require any specialist equipment. A data engineering pipeline was built to extract 703,640 paired 2D-3D tempo- ral sequences from the AthletePose3D dataset to address the domain gap between generic pose estimation benchmarks and high-velocity athletic motion directly. This domain-specific data was used to train a dual-stream Transformer-GCN architecture (MotionAGFormer). The core algorithmic contribution is the Kinematic-Weighted Loss Function (Lkw), which spatially weights the lower kinetic chain joints (hips, knees, and ankles) by a factor of ten, to guide the model to focus the optimization on those anatomical regions that are most predictive of ACL injury. On a held-out test set of 3,067 sequences, Lkw achieves a global MPJPE of 0.031 (normalized units) and PCK@0.05 of 76.3% – statistically equivalent to the standard MSE baseline – while also reducing per-joint error on 5 of 6 lower-chain joints.The Left Hip demonstrates the largest targeted improvement of 9.5%. A systematic ab- lation study over penalty weights w ∈ {1,5,10,20} confirms that w = 10 provides the optimal balance between global pose accuracy and lower-body clinical precision. Kinematic feature extraction from predicted 3D coordinates achieves a Pearson cor- relation of r = 0.88 against ground-truth knee flexion angles on a representative jump-landing sequence (MAE = 7.19◦), confirming the pipeline correctly identifies postures entering the ACL risk zone (θknee < 30◦). End-to-end validation on smart- phone video of three athletic subjects produced automated risk classifications without any manual intervention, demonstrating operational readiness as a scalable clinical screening tool.