Consciousness-Domain Index: a data-driven clustering-based consciousness labeling

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

“…Prognostication of patients with pDoC remains a major challenge due to a lack of validated outcome predictors. This multicenter international study confirmed that classifying individual patients by means of a novel clinical index (i.e., the CDI) combining the CRS-R subs-cores improved prognostic accuracy compared to clinical diagnosis or to the CRS-R total score, as previously observed in a large cohort of patients with pDoC [19].…”

“…Importantly, the improvement of the prediction accuracy obtained by means of the CDI, compared to the clinical diagnosis, was found to be stable at approximately 5%, with a sensitivity 10% higher than for clinical diagnosis. This finding suggests that the prediction value of the CDI observed previously [19] was in fact not impacted by patient demographics and medical history but was strictly related to the behavioral findings. The CDI derives from a simultaneous multi-dimensional assessment of consciousness, balancing items derived from all CRS-R subscales.…”

“…The IBIA DoC-SIG database [14] was used for the external validation of the CDI, previously internally validated on a different database from an Italian prospective study (hence "reference database") [6,19]. For both studies, inclusion criteria were: (i) age ≥ 18 years; (ii) clinical diagnosis of VS/UWS or MCS, according to standard diagnostic criteria [4]; (iii) traumatic or non-traumatic (i.e., vascular or hypoxic-ischemic/anoxic) etiology; and (iv) time post-injury from 28 days to 3 months.…”

“…Patients from the reference database were used to derive cluster properties and to deploy the unsupervised model as internally validated in a previous methodological study [19]. In that study, the number of clusters (N clusters ) that best divided the cohort was 2 (yielding maximal Silhouette score).…”

“…Recently, a Consciousness Domain Index (CDI), computed by means of unsupervised machine learning through the combination of the CRS-R subscores, classified a large cohort of patients with pDoC into two clusters, where the visual and motor subscales were the most discriminating factors. The clustering predicted patients who showed consciousness recovery at 6 months post-injury with a higher accuracy than the clinical diagnosis [19]. The prognostic value of the CDI has only been tested in a single patient sample for predicting a short-term outcome (i.e., 6 months post-injury).…”