Explainable Machine Learning for Coma Outcome Prediction Based on Structural and Functional Brain MRI.
Authors
Affiliations (7)
Affiliations (7)
- Critical Care Unit, University Teaching Hospital of Purpan, Toulouse, France.
- Toulouse NeuroImaging Center, Toulouse University, Toulouse, France.
- Critical Care Unit, University Teaching Hospital of Rangueil, Toulouse, France.
- Neurocritical Care Unit, University Teaching Hospital of Purpan, Toulouse, France.
- Cognitive Science Group, Instituto de Investigaciones Psicológicas, Facultad de Psicología Universidad Nacional de Córdoba-CONICET, Córdoba, Argentina.
- Artificial and Natural Intelligence Toulouse Institute ANITI, Toulouse, France.
- Institute of Research in Informatics (IRIT) of Toulouse, Toulouse, France.
Abstract
Advanced MRI is recommended for the clinical evaluation of patients with coma. However, the implementation of these guidelines has been hindered by an inadequate identification of relevant markers among the vast amount of reported MRI-derived metrics. We developed and validated an innovative and explainable machine learning (ML) analytical pipeline to fill this critical knowledge gap. Prospective cross-sectional study. Three Intensive Critical Care Units affiliated to the University in Toulouse (France). Patients with coma (Glasgow Coma Scale score at the hospital admission ≤ 9) from either traumatic or anoxic origin. Patient's neurologic outcome was assessed at 3 months by using the Coma Recovery Scale-Revised. Whole-brain advanced structural MRI data and functional connectivity analysis of resting-state networks known to contribute to conscious processing. A specifically designed ensemble of explainable ML methods was applied and cross-validated. None. Overall, 64 patients with coma due to either traumatic (n = 26) or anoxic (n = 38) brain injuries were studied and compared with 55 controls. The median delay between ICU admission and MRI scan was 9 days (interquartile range, 6-16 d). At 3 months, 50% patients (32/64) had an unfavorable outcome. All the models showed valuable generalization capacities: coma diagnosis (mean accuracy, 0.934%), primary brain injury discrimination (mean accuracy, 0.762 %), and neurologic outcome prediction (mean accuracy, 0.824 %). A new ensemble of brain MRI-derived metrics was specifically related to coma state, its etiology, and the patient's potential for recovery at 3 months. The structural and functional integrity of mesocircuit and frontoparietal networks appeared to carry the most relevant information.