Temporal EMS Protocol Prediction System via Progressive Clinical Feature Learning

Emergency Medical Technicians (EMTs) must select treatment protocols under severe time and cognitive constraints while patient information is still being collected. Existing Emergency Medical Service (EMS) prediction systems typically require complete records, leaving early response unsupported. We present a Temporal EMS Protocol Prediction System that progressively ranks protocols as evidence accumulates from dispatch through provider impressions. Progressive Clinical Feature Learning (PCFL) uses independently trained predictors at successive, capture-time-audited stages, restricting predictions to available information. Stage-wise Ensembles for Temporal Pooling (STEP) adaptively combines stage-wise and modality-specific predictions without additional patient information. The Dispatch–Protocol Confidence Score (DPCS) provides stage-dependent reliability estimates from empirical dispatch–protocol relationships. Staged-Evidence Prompting (SEP) evaluates large language models (LLMs) under matched evidence constraints, enabling record-level comparisons with structured predictors. Evaluation on the National Emergency Medical Services Information System (NEMSIS) dataset shows improving protocol-ranking accuracy as clinical evidence accumulates, with structured predictors consistently outperforming zero-shot LLMs under matched conditions. STEP further improves ranking performance, while DPCS provides an interpretable reliability signal with observed accuracy. These findings support modeling temporal EMS documentation structure to provide progressively informed decision support throughout patient care.