asep-sepsis-onnx

asep-sepsis-onnx

Early warning for sepsis six hours before onset, from seven bedside vitals plus the patient's age and hours since admission.

385 KB. Runs on a CPU, in-process, with no GPU, no network call, and no Python at inference time. It is the inference half of asep, a triage system for Indonesian puskesmas and other primary care facilities (FKTP), and it loads directly into a Next.js server through onnxruntime-node.

Clinical decision support. It presents an estimate alongside the completeness of the data behind it; the clinical decision stays with the health worker.

---

Competition context

Built by team MyLove for Datathon 2026, a Data & AI implementation competition.

The question we set ourselves was not "how high can AUPRC go on PhysioNet" — that problem is well studied, and a leaderboard number would not have survived contact with a real puskesmas. We asked instead:

What does a sepsis early-warning system look like when the facility running it cannot measure everything?

Most sepsis models assume a full ICU monitor. Ours assumes the opposite and is built to keep working as sensors disappear — because that is the situation in the facilities where early detection would change the most outcomes.

---

What makes this model different

It is designed around missing data, not in spite of it. NaN is a first-class input. A facility with no capnograph sends NaN for EtCO2 every hour, forever, and the model routes it down a split it learned during training. There is no imputation step to configure and no sentinel value to get wrong.

It clears the clinical rules by a wide margin. qSOFA, SIRS and a partial NEWS2 — the scores a clinician can compute at the bedside from these same numbers — reach 1.2–1.6× lift over chance. This model reaches 4.4×.

It degrades gracefully. Scored against synthetic IEEE 11073 device streams with dropouts and dead devices across five facility profiles, it still alarmed 41 of 48 septic patients at a median 16 hours before onset — including facilities down to a blood-pressure cuff and a thermometer.

Physiology drives the score. TreeSHAP attribution is led by temperature (21.8 %), heart rate (15.7 %) and respiratory rate (11.8 %). Length-of-stay influence was deliberately constrained during training so the model reasons from the patient's condition rather than from how long they have been admitted.

The learned risk curves are clinically recognisable — and nobody told the model these thresholds. Temperature is U-shaped with fever dominant, heart rate rises past 104 bpm, respiratory rate jumps past 24, systolic pressure below 98 mmHg raises risk, and MAP below 65 stands out as the shock threshold.

Calibration travels inside the file. The graph ends in a monotone map that places the utility-optimal operating point exactly on the application's high-risk cutoff. The constants live in the .onnx itself, so the calibration can never drift away from the weights it was fitted against.

The contract is machine-verified against the application's own source. The export tooling parses the feature order and the risk thresholds straight out of the TypeScript and fails if either drifts — rather than keeping a second copy that can quietly disagree.

It is fast. 17,824 patient-hours score in 0.029 s — about 620,000 patient-hours per second on one CPU core.

---

The contract

One input tensor, one output tensor.

Units: years, hours, bpm, °C, breaths/min, mmHg, mmHg, %, mmHg.

MAP is computed inside the graph

The booster reads ten columns; the graph accepts nine and derives the tenth itself:

input[:, SBP], input[:, DBP]  ->  DBP + (SBP - DBP)/3  ->  concat  ->  booster

Mean arterial pressure needs an arterial line, so the application never sends it. Deriving it in-graph keeps the [N, 9] contract intact and guarantees the value the model scores is the same one the interface displays. If SBP or DBP is absent, MAP stays NaN rather than becoming a plausible-looking number.

---

Usage

Python

import numpy as np, onnxruntime as ort

session = ort.InferenceSession("sepsis-model.onnx")

# age, hoursSinceAdmission, HR, Temp, Resp, SBP, DBP, O2Sat, EtCO2
# NaN = never measured. This facility has no capnograph.
row = np.array([[65, 10, 126, 38.9, 28, 88, 52, 90, np.nan]], dtype=np.float32)

score = float(session.run(None, {"input": row})[0][0])
print(score)  # 0.9618

Node.js

import ort from "onnxruntime-node";

const session = await ort.InferenceSession.create("sepsis-model.onnx");
const values = Float32Array.from([65, 10, 126, 38.9, 28, 88, 52, 90, NaN]);
const out = await session.run({
  input: new ort.Tensor("float32", values, [1, 9]),
});
console.log(out.riskScore.data[0]); // 0.9618

Reading the score

riskScore is a calibrated ranking, tuned so the cutoffs land on actionable review tiers:

The three tiers separate cleanly: a patient in the high tier is roughly 10× more likely to develop sepsis within six hours than one in the low tier. Read the score as a position in the queue — "this patient is in the top 16 % by risk" — which is what makes it useful for triage.

---

Training data

PhysioNet/CinC Challenge 2019, training sets A and B — 40,276 ICU patients, 1,549,880 patient-hours, 2,929 (7.27 %) septic.

The label is the Challenge's own SepsisLabel: 1 from six hours before clinical onset onward. No relabelling was done, which is what makes this early prediction rather than detection.

Preprocessing, matched to the application

• LOCF. Each vital at hour t is the last observed value at or before t — exactly how the application carries values forward.
• Gating. Only rows the application would actually score are used: age ≥ 18, at least 3 variables present across at least 2 physiological domains.
• Plausibility. Values outside the application's clinical bounds are rejected before the forward fill, so a broken cuff reporting 0 mmHg never becomes a reading.
• float32. Training is quantised to the precision ONNX serves at, so the trainer and the runtime agree on every split.

---

Results

LightGBM, 352 trees. Splits are by patient, never by row — a row split would leak a patient into both sides. Train 910k / tune 299k / test 304k hours. Every threshold was chosen on tune; the test split was scored once.

Against the test split's 1.88 % base rate, AUPRC 0.0825 is a 4.4× lift over chance. Official PhysioNet normalized utility at the operating threshold: 0.0675.

It beats what a clinic already has

Every row below was scored on the same test split from the same nine columns — nothing had more information than anything else.

The bedside rules a clinician already applies — qSOFA, SIRS, a partial NEWS2 — top out at 1.6× lift on this data. This model reaches 2.7× the AUPRC of the best of them, and still adds 16 % over a logistic regression given the exact same columns.

Holding up as equipment disappears

500 patients, five facility profiles, synthetic device streams with dropouts and dead devices:

41 of 48 septic patients alarmed, median lead time 16 hours before onset.

Feature attribution (TreeSHAP, held-out)

Vital signs account for 80.8 % of the attribution — the model reasons from physiology.

---

Verification

Every exported model passes an automated contract check before it ships: graph shape and tensor names, NaN handling on every single-variable dropout, the in-graph MAP derivation, numerical parity against the source LightGBM booster, monotonic separation of the risk tiers against observed sepsis rates, and clinical direction — worsening vitals must never lower the score.

Team

MyLove — Datathon 2026.

Citing the data

Reyna MA, Josef CS, Jeter R, et al. Early Prediction of Sepsis From Clinical Data: The PhysioNet/Computing in Cardiology Challenge 2019. Critical Care Medicine 2020;48(2):210-217.

The weights are derived from that dataset; check its terms before redistributing them.