Skip to content
Yoshio NomuraContact

All work · Gated inference

Pulsemind

Critical-care telemetry — research prototype

  • Team
  • May 2026 — ongoing

What it is

An asynchronous ICU telemetry processor. An XGBoost classifier scores incoming stream events and gates higher-cost LLM rationalisation calls so they fire only on detected anomalies.

Evaluation

Score per reading
~43 ms
including its TreeSHAP explanation — demo service, 7 readings, 31.6–64.7 ms

Pipeline

Pulsemind pipelineTelemetry ingest over an mTLS WebSocket in the serving demo, then an XGBoost classifier (~43 ms including its TreeSHAP explanation — demo service, 7 readings, 31.6–64.7 ms), then a gate: only a detected anomaly continues to LLM rationalisation, which produces a structured rationale. The LLM step takes roughly 15 seconds, on the anomaly path only.Telemetry ingestmTLS WebSocketserving demoXGBoost classifier~43 ms score per readingAnomaly?no → stop hereyesLLM rationalisation~15 sanomaly path onlyStructuredrationaleruns on every eventruns only when asked for
Classifier: ~43 ms score per reading, including its TreeSHAP explanation — demo service, 7 readings, 31.6–64.7 ms. Solid: runs on every event. Dashed: runs only on the anomaly path.

How it works

Asynchronous inference pipeline with conditional explainability gating

Stream events are scored by an XGBoost classifier, selected in a bake-off against LightGBM and CatBoost, trained on the credentialed MIMIC-IV (PhysioNet) de-identified ICU dataset under its data use agreement; demonstrations run on a synthetic derivative built from that dataset. The feature set is 109 columns: eleven frozen ventilator time-series parameters — PEEP, PIP and FiO2 among them — each expanded into observation, staleness and imputation columns, plus static comorbidity, demographic and drug-exposure features. In the serving demo, telemetry reaches the classifier over an mTLS WebSocket. Events scoring below the risk threshold complete on the fast path alone; only those flagged as anomalous trigger the downstream LLM rationalisation call, the expensive step the gate exists to avoid paying for. No identifiable patient data is used, and there is no clinical deployment.