Abstract: To address the critical technological bottleneck in mine emergency rescue—namely, the difficulty of rapidly and accurately quantifying the assessment of a trapped person’s vital state (especially the continuous changes in consciousness level) after localization—this paper studies and constructs a vital state assessment system based on the intelligent fusion of multimodal physiological information. The core of this research lies in abandoning the traditional paradigm of general feature transfer and instead designing a knowledge-driven native feature engineering framework specifically for the specific problem of “assessing the coma-low arousal continuum spectrum post-mine disaster”.The system synchronously acquires electroencephalogram (EEG), electrocardiogram (ECG), and Peripheral Capillary Oxygen Saturation (SpO₂) signals, for which depth-specific feature extraction algorithms are respectively designed: To overcome the constraint of spatial information loss in single-channel forehead EEG, a “deep temporal information mining” framework is proposed, integrating high-resolution time-frequency analysis, multi-scale nonlinear dynamics (such as multi-scale sample entropy, recurrence quantification analysis), and transient waveform morphology. For ECG and SpO₂, dedicated feature sets centered on quantifying autonomic nervous function gradients and oxygenation trends/load are constructed, respectively. Building on this, the paper further proposes a hierarchical conditional fusion system, which structurally encodes the clinical prior knowledge of “assessing vital signs first, then assessing the state of consciousness” into the network architecture, achieving a logical closed loop for feature extraction and fusion decision-making. In tests based on public datasets and a self-built “Arousable” dataset (totaling 2 500 samples), the system achieved an average assessment accuracy of 94.90% and a recall rate of 94.30% for the four states: “arousable”“mild coma”“deep coma” and“near death” .This performance significantly outperforms mainstream comparative models, providing an effective algorithm and system architecture for precise vital state assessment in extremely restricted environments.