Predicting Long-Term Recovery of Consciousness in Prolonged Disorders of Consciousness Based on Coma Recovery Scale-Revised Subscores: Validation of a Machine Learning-Based Prognostic Index

Abstract

Prognosis of prolonged Disorders of Consciousness (pDoC) is influenced by patients’ clinical diagnosis and Coma Recovery Scale-Revised (CRS-R) total score. We compared the prognostic accuracy of a novel Consciousness Domain Index (CDI) with that of clinical diagnosis and CRS-R total score, for recovery of full consciousness at 6-, 12-, and 24-months post-injury. The CDI was obtained by a combination of the six CRS-R subscales via an unsupervised machine learning technique. We retrospectively analyzed data on 143 patients with pDoC (75 in Minimally Conscious State; 102 males; median age = 53 years; IQR = 35; time post-injury = 1–3 months) due to different etiologies enrolled in an International Brain Injury Association Disorders of Consciousness Special Interest Group (IBIA DoC-SIG) multicenter longitudinal study. Univariate and multivariate analyses were utilized to assess the association between outcomes and the CDI, compared to clinical diagnosis and CRS-R. The CDI, the clinical diagnosis, and the CRS-R total score were significantly associated with a good outcome at 6, 12 and 24 months. The CDI showed the highest univariate prediction accuracy and sensitivity, and regression models including the CDI provided the highest values of explained variance. A combined scoring system of the CRS-R subscales by unsupervised machine learning may improve clinical ability to predict recovery of consciousness in patients with pDoC.

Keywords

• disorders of consciousness
• coma recovery scale-revised
• prognosis
• rehabilitation
• machine learning