Predicting ACL Injuries Using Machine Learning Models and Tibial Anatomical Predictors

Abstract

The tibial slope and the tibial depth are well-established risk factors for Anterior Cru- ciate Ligament (ACL) injury. As ML continues to progress, it has become an increasingly reliable tool for clinical screening and risk factor analysis. This thesis aims to develop and validate an explainable prognostic ML model to predict ACL injury outcomes from these Tibial Anatomical Feature (TAF), and identify the most predictive features among these parameters.

A dataset comprising Coronal Tibial Slope (CTS), Medial Tibial Slope (MTS), Lat- eral Tibial Slope (LTS), Medial Tibial Depth (MTD), and sex was constructed using MRI scans taken from 104 subjects (44 males: 22 injured, 22 uninjured; 60 females: 27 in- jured, 33 uninjured). Two distinct ML pipelines were developed: a self-developed pipeline (including K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF), XGBoost, CATBoost, Multi-Layer Perceptron (MLP), and TabNet) and an advanced AutoGluon pipeline (including XGBoost, LightGBM, CatBoost, TabPFN, TabM, TabICL, MITRA, and their weighted ensembles). Both were designed as end-to-end pipelines to pro- cess the dataset and output predictions with integrated feature importance explanations. Empirically, the AutoGluon Pipeline demonstrated superior performance and training-time efficiency. The recommended F2-tuned standard ensemble achieved an F2-score of 0.736 on the validation set. On the test set, it demonstrated a test balanced accuracy of 0.955, F1-score of 0.952, F2-score of 0.980, ROC AUC of 1.000, precision of 0.909, and recall of 1.000. A full-dataset model, the F2-tuned full-dataset ensemble refitted on the entire dataset for clinical deployment achieved a validation F2-score of 0.813. The global feature importance analyses performed via SHapley Additive exPlanations (SHAP), established the descending order of influences as MTD, LTS, MTS, CTS, and sex.

In summary, the study recommends two versions of the F2-tuned prognostic models, one being a standard ensemble model and the other a full-dataset ensemble. The former, which demonstrated moderately high predictive power, was designed for subsequent research comparison. The latter, without access to the original held-out test set, is constructed for maximum robustness and generalization in real-life clinical deployment. Global feature importance analyses elucidated from the standard ensemble decreased MTD along with increased LTS and MTS as most contributive features for ACL injury. These models serve as both feature attribution tools as well as clinical screening tools. These models are intended to be integrated into clinical practice as explainable machines to assist clinicians in predicting the likelihood of ACL injury.