NLP call intake, MPDS/ProQA structured interrogation, and live audio ML dispatcher-assist for out-of-hospital emergency response
The first fifteen seconds of an emergency call are the noisiest, least structured, and arguably most important seconds in the entire chain of survival. Automatic speech recognition (ASR) converts narrowband telephone audio into text in real time, while a natural-language-processing layer extracts symptom entities, negations, and timing cues from whatever the caller manages to say — long before any formal protocol has begun.
Landline and cellular 911 audio is transmitted as narrowband telephony (roughly 300–3400 Hz, often 8 kHz sampling) — far lower fidelity than the audio modern speech models are usually trained on. A dispatch-grade ASR system has to cope with several compounding problems simultaneously:
• Background noise: traffic, wind, television, other people shouting in the room • Emotional speech: crying, screaming, hyperventilating callers distort normal phonetic patterns • Multiple talkers: a caller relaying information from someone else on scene ("she says he's not breathing") • Non-native accents and regional dialect variation across a jurisdiction's caller population • Line dropout and cellular handoff artifacts mid-sentence
Modern telephony-tuned acoustic models (often fine-tuned transformer-based ASR, similar in spirit to Whisper-class architectures but adapted to 8 kHz call-center audio) are typically retrained on large corpora of historical, de-identified 911 recordings so the model learns the specific acoustic and lexical distribution of emergency calls — full of interruptions, filler words, and incomplete sentences — rather than clean read speech.
Once a transcript stream exists, an NLP layer performs named-entity recognition tuned to emergency-medical vocabulary: chief-complaint keywords ("chest pain," "can't breathe," "unconscious"), temporal markers ("started twenty minutes ago," "just happened"), and — critically — negation and hedging. The sentence "he IS breathing" and "he is NOT breathing" differ by one token but represent opposite triage realities, so negation-scope detection is one of the hardest and highest-stakes NLP problems in this pipeline.
Lay callers rarely use clinical terminology. A caller describing a myocardial infarction might say "funny feeling in my chest" or "my arm feels weird" rather than naming a symptom cluster. Effective extraction systems maintain a mapping layer between colloquial phrasing and clinical concepts, informed by large annotated corpora of real call transcripts linked to confirmed patient outcomes.
The extracted entities are not used to replace the dispatcher's questioning — they are surfaced as a running summary alongside the structured interrogation described in Stage 2, giving the dispatcher (and any downstream AI-assist layer) a head start on likely acuity before the formal protocol resolves.
Caller panic measurably degrades both signal and content: distressed callers speak faster, omit subject pronouns, and skip details they assume are obvious. NLP systems built for this domain are evaluated specifically on noisy, incomplete, panic-inflected transcripts — not on clean benchmark speech — because that is the actual operating condition.
The Medical Priority Dispatch System (MPDS), created by Dr. Jeff Clawson in 1976 and maintained by the International Academies of Emergency Dispatch (IAED), is the most widely used protocol for converting an emergency caller's answers into a standardized severity code. ProQA, the software that operationalizes MPDS, walks the dispatcher through a scripted Case Entry, chief-complaint protocol selection, and a fixed sequence of Key Questions — resolving every call to a Determinant Code such as 9-E-1.
Before any specific protocol is chosen, ProQA forces every call through a small set of universal Case Entry questions: the address, a callback number, and — critically — whether the patient is conscious and breathing normally. These two Key Questions can immediately trigger the highest-priority Echo-level response and initiate Pre-Arrival Instructions before the dispatcher has even identified the full chief complaint, precisely because "not breathing" outranks almost every other consideration in the system.
Once Case Entry is complete, the dispatcher selects one of 30-plus Advanced MPDS (AMPDS) chief-complaint protocols — Protocol 6 Breathing Problems, Protocol 9 Cardiac or Respiratory Arrest, Protocol 10 Chest Pain, Protocol 17 Falls, Protocol 26 Sick Person, and so on — based on the caller's initial description. Protocol selection is deliberately conservative: when a call's presentation is ambiguous, ProQA is designed to route toward the protocol most likely to catch a life threat, accepting some over-triage in exchange for lower miss rates.
Each protocol contains a scripted branching sequence of Key Questions — yes/no or short-answer prompts read verbatim by the dispatcher — that narrows the presentation down to one of five Determinant Levels: Alpha (minor), Bravo (moderate), Charlie (serious, ALS should be considered), Delta (serious, ALS with lights-and-sirens), or Echo (immediately life-threatening). The resulting code, formatted as Protocol-Level-Suffix (e.g. 9-E-1 for a confirmed cardiac arrest under Protocol 9), tells the CAD (computer-aided dispatch) system exactly which unit type and response mode to send — without requiring the dispatcher to exercise unstructured clinical judgment under time pressure.
This structure is the backbone of every subsequent step in this simulation: NLP-extracted entities from Stage 1 can pre-populate likely answers, and the AI audio-assist described in Stage 3 runs in parallel with — never instead of — this scripted interrogation.
Once an Echo-level cardiac or respiratory arrest is identified, ProQA immediately transitions into scripted Pre-Arrival Instructions (PAIs), most importantly Dispatcher-Assisted CPR (DA-CPR): the dispatcher talks the caller through compression-only CPR in real time, often counting the compression rate aloud. Getting bystander CPR started even 60–90 seconds sooner measurably changes cardiac-arrest survival odds, which is why this hand-off from interrogation to instruction is deliberately built into the protocol rather than left to dispatcher discretion.
Every MPDS-certified center runs a formal Continuous Quality Improvement (CQI) program: a sample of calls is reviewed against the scripted protocol for compliance, and centers can earn Accredited Center of Excellence (ACE) status from IAED by sustaining high compliance and case-review scores over time.
A subtle but important design choice: the two Case Entry "is he breathing / is he conscious" questions can escalate a call to Echo-level before the dispatcher has even chosen a chief-complaint protocol. Structured interrogation is intentionally front-loaded toward catching the deadliest presentations first.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| Alpha | |||
| Bravo | |||
| Charlie | |||
| Delta | |||
| Echo |
Copenhagen Emergency Medical Services, working with the AI company Corti, deployed a real-time machine-learning listener that runs continuously alongside ProQA, analyzing raw call audio — not just the transcript — for acoustic patterns associated with cardiac arrest, most notably agonal (gasping) breathing, and flags likely out-of-hospital cardiac arrest (OHCA) to the dispatcher even when the caller never uses expected words like "not breathing."
Unlike a keyword spotter, the audio-assist model ingests acoustic features directly from the call — prosody, pitch irregularity, breathing cadence, and the characteristic irregular low-frequency vocalization of agonal gasping — alongside the evolving transcript. Agonal breathing is a critical target because it is present in a large share of early cardiac arrest calls but is frequently mistaken by callers and even trained dispatchers for the patient "still breathing," since gasping sounds like breathing to an untrained or panicked ear.
The classifier, trained on a large historical archive of recorded emergency calls linked to confirmed outcomes (via regional cardiac arrest registries), outputs a continuously updating probability score. When that score crosses a threshold, the system surfaces a discreet on-screen alert to the dispatcher — it never interrupts the ProQA script, removes dispatcher control, or auto-dispatches units on its own.
The most rigorous evaluation of this technology is a large randomized controlled trial run across Copenhagen's dispatch centers (Blomberg et al., JAMA Network Open, 2021), comparing dispatcher performance with and without the AI system live on the call. The topline result is more nuanced than the technology's early promise suggested: sensitivity for OHCA recognition was numerically higher in the AI-assisted arm (roughly 84% versus 82% in the control arm) but the overall trial did not demonstrate a statistically significant system-wide improvement across its full study period, partly because ProQA-trained dispatchers already catch a large majority of arrests through the Case Entry questions alone.
This is an important, deliberately unglamorous finding: AI audio-assist is best understood as a second set of ears that helps most in edge cases — fatigued dispatchers, unusual presentations, ambiguous caller language, high call volumes — rather than a guaranteed uniform accuracy jump. Later deployments and subgroup analyses have continued to refine where the tool adds the most value.
Similar audio- and language-model-based decision-support research is underway for other time-critical presentations — stroke symptom detection from speech patterns, sepsis-risk flagging from combined vitals-adjacent language cues, and opioid-overdose recognition from caller description patterns — though most of these remain in earlier research or limited-pilot stages compared to the multi-year, RCT-validated cardiac arrest use case.
The system is explicitly built as advisory decision support, not autonomous triage: alerts require dispatcher acknowledgment, the underlying ProQA interrogation continues unchanged, and no dispatch tier is altered by the AI without a human in the loop.
Every triage system, human or algorithmic, sits on a sensitivity/specificity tradeoff. Under-triage sends a lower-priority response to what is actually a life-threatening emergency, delaying lifesaving care. Over-triage sends a lights-and-sirens Code 3 response to a call that did not need it, consuming scarce ambulance hours and elevating crash risk. MPDS is deliberately tuned to accept more of the latter in order to minimize the former.
Under-triage happens when a genuinely critical call — an evolving cardiac arrest, a stroke, a severe airway obstruction — is coded to a lower determinant level, resulting in a slower, non-Code-3 response or a BLS-only unit reaching a patient who needed ALS immediately. Because the true acuity of the call is unknown at coding time, under-triage is invisible until the crew arrives and finds a patient far sicker than dispatched for.
Cardiac arrest survival falls roughly 7–10% for every minute that passes without CPR and defibrillation, so even a few extra minutes lost to under-triage can be the difference between a walking discharge and a fatality. This is precisely why false negatives are treated as the far more dangerous error class in dispatch system design, even though they occur less often than false positives.
Over-triage — dispatching lights-and-sirens for a call that turns out to be non-emergent — carries its own well-documented costs. Emergency driving under lights and sirens carries measurably elevated crash risk compared with normal-traffic response, and every unit committed to an unnecessary Code 3 run is a unit unavailable for the next genuinely critical call in the district. Certain determinant codes — particularly "difficulty breathing" and "unknown problem" categories that must cast a wide net — show over-triage rates reported as high as 20–50% in some system audits, because the underlying symptom cluster spans everything from a mild asthma flare to impending respiratory arrest and the protocol cannot always distinguish them from caller answers alone.
Dispatchers also face a well-known incentive toward defensive over-triage: the personal and professional consequence of missing a true emergency is far more severe than the consequence of sending extra resources to a false alarm, which pushes real-world behavior further toward over-triage than the written protocol alone would predict.
Like any diagnostic test, dispatch triage can be described with a sensitivity/specificity curve: moving the decision threshold toward higher sensitivity (catching more true emergencies) necessarily increases false positives, and vice versa. MPDS protocols are intentionally biased toward high sensitivity for the deadliest chief complaints — cardiac arrest, choking, major trauma — because the cost of a missed Echo-level call vastly exceeds the cost of an unnecessary ambulance run.
The promise of AI audio-assist and NLP-based decision support is not to move the threshold along the same curve, but to move the curve itself outward — improving sensitivity and specificity together by extracting more genuine signal (agonal breathing acoustics, subtle language cues) from the same call, rather than simply asking dispatchers to be more suspicious across the board.
In emergency dispatch, the two error types are not symmetric: missing a true cardiac arrest can cost a life, while an unnecessary Code 3 run costs time, money, and marginally elevated crash risk. Protocols are deliberately, and defensibly, engineered to trade one against the other rather than to minimize total errors.
Every stage of this pipeline exists to compress one number: the time between the first ring of a 911 call and effective treatment reaching the patient. Faster, more accurate triage only matters if it translates into measurably better outcomes — while keeping dispatcher workload sustainable and human judgment firmly in control of every automated recommendation.
A determinant code is only useful once it becomes an action: call answer → Case Entry → chief-complaint interrogation → Key Questions → Determinant Code → unit recommendation in the CAD system → crew notification → wheels rolling → on-scene arrival. National benchmarks (NFPA 1221 for call processing, NFPA 1710/1720 for turnout and travel time) set target windows for each link in that chain, typically evaluated at the 90th percentile rather than the average, because tail-end delays are where patients are lost.
Every improvement earlier in the chain — faster ASR-assisted intake, a correctly resolved Determinant Code on the first pass, an AI flag that shaves seconds off recognition — compounds down the line, because dispatch time is the one delay that occurs before any responder has even left the station.
For out-of-hospital cardiac arrest, survival to hospital discharge falls steeply and continuously with time to defibrillation — commonly cited at roughly 7–10% per minute without bystander CPR, with the decline meaningfully slower when high-quality bystander compressions are underway. This is why Dispatcher-Assisted CPR is triggered immediately on Echo-level recognition rather than waiting for crew arrival, and why even a 15–30 second improvement in recognition time — the kind an AI audio-assist tool is designed to deliver — is treated as clinically meaningful rather than a rounding error.
Dispatchers on busy shifts field dozens of calls in a row, each demanding rapid, scripted, high-stakes decision-making under emotional strain — a workload profile associated with measurable stress and burnout in the profession. The entire design philosophy of AI-assist tools discussed in this simulation is augmentation, not replacement: alerts are advisory, the scripted MPDS interrogation continues regardless, and a human dispatcher makes every dispatch decision.
This matters for liability and governance as much as for clinical accuracy. These systems are generally positioned as clinical decision support rather than autonomous medical devices, leaving accountability with the dispatch agency and the human operator, documented through the same Continuous Quality Improvement and case-review processes used to audit MPDS compliance. A poorly tuned or overly aggressive AI layer also risks alarm fatigue — if alerts fire too often on low-probability calls, dispatchers learn to discount them, quietly erasing the benefit the tool was built to provide.
The ultimate measure of success for AI-assisted 911 triage is not model accuracy in isolation — it is system-level outcomes: survival rates, appropriately targeted response tiers, and a workload that trained human dispatchers can sustain shift after shift without alert fatigue eroding trust in the tool.