A Deep Learning-Based Prognostic Classification Model for Patients with Prolonged Disorders of Consciousness

Abstract

Accurate prognostic assessment of patients with prolonged disorders of consciousness is critical for clinical decision-making. However, traditional behavioral scales and neuroimaging techniques are limited by subjective interpretation and low temporal resolution, which impede the dynamic characterization of consciousness fluctuations. To address these challenges, this study proposes a deep learning-based prognostic classification model integrating multiscale electroencephalogram features. First, geometric features including maximum radius, regional density, and dispersion were extracted from power spectrum-Poincaré scatter plots, while nonlinear dynamic features were constructed using sample entropy and multiscale entropy. Second, a temporal hybrid network enhanced by cross-attention mechanisms was designed to strengthen the modeling of feature interdependencies. Experimental validation on 8-channel electroencephalogram data from 15 patients demonstrated a classification accuracy of 90.83 percent and sensitivity of 94.73 percent, with significant performance improvements compared to random forest and support vector machine baselines.