Skip to main content

End-to-end: vitals watch

Advanced ⏱ 25 min healthcare · end-to-end · local llm

You will build: the deterioration-watch pipeline from the Healthcare use case — vitals windowed per patient, escalation thresholds with no model in the alarm path, and a situation summary drafted by a local Ollama so PHI never leaves the machine.

vitals → streaming:trend → rule-based:escalate → llm:summarize (local) → webhook (unit board)

Prerequisites​

  • The quickstart completed.
  • A local Ollama for step 4 — the point of this pipeline is that the model runs inside your perimeter. (Without it, stop after step 3: escalation already works.)

1 · Scaffold​

pulse new vitals --source webhook \
  --stage streaming:trend --stage rule-based:escalate --stage llm:summarize \
  --sink webhook

2 · The trend stage — per-patient windows​

stages:
  - name: trend
    engine: streaming
    operators:
      - { type: keyBy, field: patient_id }
      - type: window
        spec: "sliding(10m,1m)"
        aggregations:
          hr_avg:    "avg(heart_rate)"
          hr_stddev: "stddev(heart_rate)"
          spo2_min:  "min(spo2)"

3 · The escalation stage — deterministic, microseconds, auditable​

  - name: escalate
    engine: rule-based
    rules:
      - { condition: "spo2_min < 92 || hr_stddev > 15", action: flag }

The alarm path contains no model: thresholds your clinical governance signed off, applied identically every time.

4 · The summary stage — a local model drafts the context​

  - name: summarize
    engine: llm
    systemPrompt: |
      Draft a one-paragraph situation summary for the
      responding clinician: what crossed the line, the
      trend, and the window it happened in.

Point Pulse's LLM provider at your local Ollama — the event data is summarized without leaving the machine.

5 · Sink to the unit board and deploy​

pulse secret set UNIT_BOARD_WEBHOOK https://webhook.site/<your-id>
pulse secret set PULSE_WEBHOOK_SECRET dev-secret   # the scaffolded webhook source requires it
pulse deploy .

6 · Simulate a deteriorating patient​

for spo2 in 97 96 95 93 91 90; do
  pulse events publish --topic vitals.in --value '{"patient_id":"p-12","heart_rate":'"$((RANDOM % 40 + 80))"',"spo2":'"$spo2"'}'
done
pulse events tail --topic vitals.escalate.out

The moment spo2_min drops under 92 inside the window, the escalation fires; vitals.summarize.out carries the drafted paragraph; the unit board webhook receives both.

What just happened​

The clinical constraints shaped the architecture: rules decide, the model only describes, the human acts — and every step landed on an auditable topic. Because the LLM is local, the privacy property is architectural, not contractual. The system surfaced attention; it never acted on a patient.

Troubleshooting​

No escalation fires

The window needs readings below 92 within the sliding window — send the burst within a couple of minutes. Check vitals.trend.out to see the computed spo2_min per window.

summarize is unhealthy

Ollama isn't reachable. That stage stays visibly unhealthy — escalation (the safety-critical part) keeps working without it.

Next​