impacte / laya-medical-escalation

Laya Medical Escalation ( impacte/laya-medical-escalation )

⚠️ NOT A MEDICAL DEVICE — NOT FOR CLINICAL USE

This is a research and educational artifact . It was trained to reproduce the acuity levels recorded in a historical, de-identified dataset. It must never be used to triage, diagnose, or make care decisions for real patients. Acuity assessment is subjective and population dependent; a model that under-triages can cause serious harm.

Laya Medical Escalation is convaiinnovations/laya (421M, ModernBERT-large + typed decision head) fine-tuned to assign the Emergency Severity Index (ESI 1–5) from an emergency-department triage record. Laya is non-autoregressive: it scores each ESI level at its own [MASK] marker and softmaxes over the levels in one forward pass (~23 ms), returning a calibrated distribution rather than free text.

Usage

import laya

router = laya.load("impacte/laya-medical-escalation")

criteria = {
    "esi-1": "ESI 1 Resuscitation: requires immediate life-saving intervention "
             "(e.g. cardiac arrest, intubation, severe respiratory distress).",
    "esi-2": "ESI 2 Emergent: high-risk, drowsy/confused, severe pain; should not wait.",
    "esi-3": "ESI 3 Urgent: needs two or more resources (labs, imaging, IV fluids); vitals stable.",
    "esi-4": "ESI 4 Less urgent: needs one resource (x-ray, prescription, simple procedure).",
    "esi-5": "ESI 5 Non-urgent: needs no resources; can be referred or discharged.",
}

res = router.predict(
    "Emergency department triage record. Patient: 58-year-old male. "
    "Chief complaint: chest pain, shortness of breath. "
    "Triage vitals: HR 118, BP 88/54, RR 28, SpO2 89% on room air, temperature 98.4 F. "
    "Arrival: ambulance. History: 3 prior ED visits.",
    {"route": {
        "type": "choice",
        "instructions": "You are an emergency department triage nurse. Based only on the triage "
                        "record, assign the correct Emergency Severity Index (ESI) level.",
        "criteria": criteria,
    }},
)
print(res["answers"]["route"]["choice"])         # -> esi-2
print(res["answers"]["route"]["probabilities"])

Results

Held-out evaluation — 2500 rows, class-balanced (chance 0.20):

Under-triage = predicted a less urgent ESI level than gold (the dangerous error); the high-acuity (gold ESI-1/2) under-triage rate is 0.189.

Accuracy & response time vs local LLMs (same 300 held-out rows)

LLMs were given a 192-token budget and a parser that strips thinking/tool-call wrappers, so this is not a formatting strawman.

Same-dataset baselines (identical train/test split)

XGBoost uses the full 970-column table; the deterministic model is the partial ESI v5 clinical algorithm (resource count inferred — the dataset has no resource, pain, or mental-status fields, so this is a lower bound on it).

Dataset

Trained on the Yale School of Medicine emergency-department triage dataset:

Hong WS, Haimovich AD, Taylor RA. Predicting hospital admission at emergency department triage using machine learning. PLOS ONE 13(7): e0201016 (2018). DOI 10.1371/journal.pone.0201016

560,486 de-identified ED visits × 972 variables. The model consumes a text state built only from at-triage fields — chief complaint (from 200 chief-complaint one-hots), triage vitals (HR, BP, RR, SpO2, temperature), age, sex, arrival mode, prior ED visits, previous disposition and outpatient medication count. HF mirror used: kondratevakate/hospital-triage-and-patient-history-data (lossless RData→Parquet re-host). Citation is required for any use of this data; do not redistribute it.

Training

• 57,271 training rows (ESI-1 4,771 · ESI-2 15,000 · ESI-3 15,000 · ESI-4 15,000 · ESI-5 7,500).
• 2 epochs, bf16 + gradient checkpointing, one RTX 5060 Ti 16 GB, ~75 min.
• Temperature calibrated to 1.046 on held-out logits.
• No leakage: esi / disposition never enter the state text.

Limitations

• Research artifact only — not for clinical use (see the banner above).
• English only; the base Laya checkpoint is English.
• The state is a template over structured fields, not free-text clinician notes.
• ESI is a nurse-assigned, partly subjective label; class boundaries are fuzzy.
• ESI-2/3 are the hardest classes; under-triage is the metric that matters and it is not zero.
• Performance may not transfer to other hospitals, populations, or triage protocols.

Credits

• convaiinnovations/laya — base model and RLCD recipe (Apache-2.0).
• Hong, Haimovich & Taylor (Yale) / PLOS ONE — the triage dataset.

Citation

@misc{laya_medical_escalation2026,
  title  = {Laya Medical Escalation: a calibrated ESI decision model},
  author = {oamazonasgabriel},
  year   = {2026},
  url    = {https://huggingface.co/impacte/laya-medical-escalation}
}