Machine learning model identifies tibial anatomical variables as potential risk factors for anterior cruciate ligament injury

TL;DR

A machine learning model was developed to predict ACL injury risk based on tibial anatomy, identifying key anatomical features that could help clinicians assess injury likelihood.

Contribution

The study introduces a machine learning model validated for predicting ACL injury risk using tibial anatomical variables and identifies the most predictive features.

Findings

The default tested model achieved high performance metrics on the test set, including 100% ROC AUC and 100% recall.

Decreased medial tibial depth and increased medial/lateral tibial slopes were top predictors of ACL injury.

The full-dataset model showed improved validation F2-score compared to the default tested model.

Abstract

The tibial slope is a well‐known risk factor for anterior cruciate ligament (ACL) injury. As machine learning continues to progress, it has become an increasingly explored tool for clinical screening and risk factor analysis. This study aims to develop and validate a prognostic machine learning model to predict the outcome of ACL injury from tibial anatomic parameters and identify the most predictive features.

A pre‐published dataset of coronal, medial and lateral tibial slopes and medial tibial depth was constructed using magnetic resonance imaging scans taken from 104 subjects (44 males: 22 injured, 22 uninjured; 60 females: 27 injured, 33 uninjured). The dataset was split into train‐validation and test sets to ensure robust model evaluation. AutoGluon‐enabled machine learning models, including XGBoost, LightGBM, CatBoost, TabPFN, TabM, TabICL, MITRA and their weighted ensembles were trained and tuned with respect to the F2‐score across ten different random seeds. Two instances of the best‐performing model were developed: a default tested model (weighted ensemble from the default seed of 42) and a full‐dataset model (weighted ensemble retrained on the entire dataset). Global SHapley Additive exPlanations analysis was used to elucidate the most predictive features, and local SHapley Additive exPlanations analysis to provide interpretability for individual predictions.

The default tested model achieved a 73.60% validation F2‐score. On the test set, it demonstrated a 95.44% test balanced accuracy, 95.24% F1‐score, 98.04% F2‐score, 100% ROC AUC, 90.91% precision and 100% recall. The full‐dataset model achieved an 81.30% validation F2‐score. The relative importance of tibial anatomical features were identified.

Overall, the study presented two prognostic models with moderately high predictive power to identify subjects with high likelihood of ACL injury. Decreased medial tibial depth along with increased medial and lateral tibial slopes were reported as top predictors for ACL injury. These models can potentially be integrated into clinical practice to assist clinicians in predicting the likelihood of ACL injury, but require external validation.

Level III, case‐control study.

Figures