ESCORT ACS Dataset (Dialogue Edition): 500 Synthetic Prehospital Acute Coronary Syndrome Cases with Protocol-Deviation Annotations and Vital-Sign Timelines

This dataset contains 500 synthetic prehospital Acute Coronary Syndrome (ACS) cases in turn-by-turn dialogue format, with protocol-deviation annotations and time-stamped vital-sign timelines. It is designed for training and evaluating classifiers that detect EMS protocol violations, and for turn-taking analysis, speech/audio synthesis, and simulation. All cases are fully synthetic and contain no real patient or clinician data. Each case is a .docx transcript of a paramedic–patient encounter (roughly T+00:00 to T+20:00 from first medical contact) using PARAMEDIC / PATIENT / PARAMEDIC (thinking) / ACTION / MONITOR / RADIO roles, with [T+MM:SS] timestamps on every line. Above each transcript, a PROTOCOL ANNOTATION SUMMARY block records — for adverse cases — each violated checkpoint with its timestamp, the rule broken, and the observed clinical consequence; correct cases list the checkpoints passed. Inline [✗ DEVIATION] and [✓ CORRECT] markers label the exact lines, making the cases directly usable as supervised training data. Deviation cases are explicitly annotated as errors so they are clearly distinguishable from endorsed practice. The case set comprises 400 standard ACS cases (50 per region across GR, BE, DE, NL, FR, IT, SE, ES), 50 adverse-event cases, and 50 mimicker cases, scored against a catalogue of 22 protocol checkpoints. Each case has a companion .timeline.json replay script mapping clinical moments to scenario presets or custom vital-sign targets, and the included vital_signs_server_v4.py can auto-replay any timeline so a simulated monitor's vitals move in sync with the dialogue. A summary spreadsheet (ESCORT_ACS_Dataset_Summary.xlsx) aggregates all 500 cases across ~60 fields including demographics, risk factors, presenting vitals, ECG findings, culprit vessel, treatments and doses, complications, adverse-event type, mimicker, destination, and deviation counts. Regional EMS protocols are modelled from published guidance as of 2023–2024. Generation is reproducible via a fixed random seed. Known simplifications: the dialogue is synthetic and may read more formally than live radio traffic; vital-sign trajectories are linear-to-target interpolations that approximate clinical deterioration/recovery rather than high-fidelity biosignals; and deviation detection is checkpoint-based rather than behaviour-based (it catches missed or wrong actions, not tone, cognitive load, or team dynamics).