Application of Machine Learning for Predicting Functional Recovery in Patients with Traumatic Brain Injury

Abstract

Traumatic brain injury (TBI) continues to be a leading cause of morbidity globally, and the ability to predict functional recovery in this patient population is complex and multifactorial. Here, we establish a supervised ML framework that predicts TBI recovery outcomes to set novel performance benchmarks using a parameterization that both circumvents common fragmentation, imbalanced dataset, overfitting, and interpretability challenges. Training data through October 2023 moved us to strong models integrating heterogeneous clinical-neuroimaging-demographic data, filling gaps among existing prediction efforts. We use advanced techniques to address missing values and mitigate overfitting, allowing our models to be generalizable to different patients. In addition, this will be important for building new interpretability methods to make such predictions more interpretable, and, as a consequence, clinically actionable. The study also discusses how to handle imbalanced datasets, and how to make models that generalize to different patient populations. It thus aims to provide a reliable, ethical and effective tool for prediction of recovery outcomes in TBI, validated not only on multi-centre cohorts but also addressing ethical concerns, e.g. data privacy and security. Ultimately, this study envisages optimization of clinical decision by providing better long-term recovery prediction and thus paving the way for better personalized therapies for TBI patients in the future.