The prognosis of neurological outcomes in patients with prolonged Disorders of Consciousness (pDoC) has improved in the last decades. Currently, the level of consciousness at admission to post-acute rehabilitation is diagnosed by the Coma Recovery Scale-Revised (CRS-R) and this assessment is also part of the used prognostic markers. The consciousness disorder diagnosis is based on scores of single CRS-R sub-scales, each of which can independently assign or not a specific level of consciousness to a patient in a univariate fashion. In this work, a multidomain indicator of consciousness based on CRS-R sub-scales, the Consciousness-Domain-Index (CDI), was derived by unsupervised learning techniques. The CDI was computed and internally validated on one dataset (N=190) and then externally validated on another dataset (N=86). Then, the CDI effectiveness as a short-term prognostic marker was assessed by supervised Elastic-Net logistic regression. The prediction accuracy of the neurological prognosis was compared with models trained on the level of consciousness at admission based on clinical state assessments. CDI-based prediction of emergence from a pDoC improved the clinical assessment-based one by 5.3% and 3.7%, respectively for the two datasets. This result confirms that the data-driven assessment of consciousness levels based on multidimensional scoring of the CRS-R sub-scales improve short-term neurological prognosis with respect to the classical univariately-derived level of consciousness at admission.