Deep white matter MRI predicts outcomes in coma of various etiologies: a cohort study

Background: A reliable outcome prognostication tool for patients in coma of various etiologies would facilitate ICU treatment by providing objective information to caregivers and patients' relatives. This study aimed to predict outcome based on supervised machine learning and magnetic resonance diffusion tensor imaging (DTI) metrics. Methods: In this multicenter international study, a training set of 531 patients not responding to simple orders at day 5 after coma onset underwent diffusion-weighted MRI between day 5 and 45. A classifier was developed using DTI metrics, patient age, and delay between admission and MRI as features. Unfavorable outcome (UFO) was defined as GOSE 1–4 at one year. Three prognosis areas were defined: a “red” zone (specificity for UFO above 95%), a “green” zone (specificity for favorable outcome, FO, above 90%), and a “no determination zone” (NDZ) for patients classified in neither the red or green zone. The classifier was validated on an external test set of 211 patients. Results: The training set included 531 patients (age 48 ± 16 years; MRI at 19 ± 8 days post-injury), with 75.9% GOSE 1–4 and 24.1% GOSE 5–8 at one year. Normalized DTI metrics were FA 0.82 ± 0.12 and MD 1.10 ± 0.13. The external test set (n = 211; age 47 ± 16 years; MRI at 21 ± 12 days) showed similar outcome distribution (75.4% GOSE 1–4, 24.6% GOSE 5–8) and DTI values (FA 0.83 ± 0.09, MD 1.07 ± 0.12). Both sets were comparable in age, sex, initial GCS, and outcome ratios. In the external test set, ROC AUC was 0.89. For UFO classification, specificity was 98.1%, PPV 99.1%, and sensitivity 68.6%. For FO classification, specificity was 95.0%, PPV 77.8%, and NPV 86.3% whereas 30.8% of the patients were in the NDZ. After excluding patients for whom life sustaining therapies were withdrawn (n = 104), specificity was 96.6% and 82.4% for UFO and FO classification, respectively. Conclusion: This classifier demonstrates a high specificity to predict coma outcome while patients are still in the ICU, irrespective of coma etiology. These results may assist practitioners in making informed decisions.