CAPIRA — Smart Capillary Refill Assessment

Research prototype. Not a validated medical device and not a substitute for clinical judgment.

An offline-first PWA that helps nurses at primary care facilities and in remote areas:

1. compute qSOFA (Sepsis-3) deterministically from vital signs;
2. measure Capillary Refill Time (CRT) with a phone camera (~22 s: rest 2 · press 10 · release 10), with a signal curve, quality checks, and an honest rejection when the data is poor;
3. compose a referral draft from a template (offline) — copy, SMS, share sheet, or download as .txt.

CAPIRA does not diagnose sepsis or shock. CRT is shown as a supporting finding, separate from the qSOFA score.

Status (honest)

Engine accuracy (synthetic signals)	Measured — MAE 0.20–0.27 s, Bland-Altman ±0.5–0.8 s on results that pass the gate (VALIDATION.md)
Accuracy on real cameras / patients	Not yet measured (Phase 1: PROTOCOL.md)
Dark skin	Higher rejection rate on synthetic data; must be measured and reported per subject
Medical / statistical claims	Every claim has a status in CLAIMS.md

Running

npm install
npm test            # 105 unit tests (engine, qSOFA, referral, storage, simulation, CSV)
npm run dev         # http://localhost:5173
npm run build && npm run preview   # production build + service worker

UI & quality checks

The interface is a responsive dashboard with desktop/mobile navigation, a dark mode that follows the system setting, a history view with search/filter, and handling for camera and storage failures. The home summary is derived from local data, including simulation records, which are labelled as such.

See the app review and follow-up work for the changes made, test coverage, and remaining limitations.

npx playwright install chromium       # once, for browser tests
npm run test:ui                       # run after npm run dev
npm run build
npm run test:ui -- --production       # automatic preview + offline test

Browser tests use an isolated profile. Reports and screenshots are written to artifacts/. For a dev server on another port, use CAPIRA_TEST_URL=http://localhost:PORT npm run test:ui.

The camera requires a secure context (HTTPS or localhost). To develop on an Android phone (Chrome), pick one:

• adb reverse tcp:5173 tcp:5173, then open http://localhost:5173 on the phone; or
• an HTTPS tunnel (e.g. cloudflared tunnel --url http://localhost:5173).

Deploy & offline demo

1. npm run build → upload the dist/ folder to static HTTPS hosting (Netlify, Vercel, GitHub Pages, Cloudflare Pages). Paths are relative, so a subfolder works too.
2. Open the URL once while online in Chrome for Android and wait for the "Ready offline" label on the home screen (the service worker has finished precaching every asset — no CDN).
3. (Optional) Menu ⋮ → Add to Home screen.
4. Turn on Airplane Mode and open the app again. The home screen, history, simulation, and result storage have been tested in headless Chromium with the network disabled; the camera flow still needs to be tested on the target phone.

5-minute demo script: docs/DEMO.md.

Research tools

Command	Purpose
npm run accuracy -- --seeds=10 [--noiseModel=relative]	Engine accuracy sweep on synthetic signals
npm run analyze -- data/raw/*.csv	Compare protocols A/B/C × 5 candidate signals from Experiment-page CSVs
npm run compare -- annotations.csv data/raw/*.csv	Engine CRT vs. blinded frame-by-frame annotation (≥2 annotators)

In the app: Experiment (record + CSV export + camera capabilities), Annotate (mark release/recovery frames without seeing the engine result).

Architecture in brief

src/engine/     Vascular Refill Engine (pure, no DOM) + synthetic generator + CSV
src/clinical/   qSOFA, age-specific CRT interpretation, non-directive recommendation
src/referral/   template referral draft + MOCK facilities
src/storage/    IndexedDB (Dexie) — derived signal + metadata, no video
src/capture/    camera (torch/AE/AWB where supported), live frame sampler (rVFC), protocol cues
src/ui/         React UI (English)

Decisions and rationale: DECISIONS.md. What was deliberately cut: CUTLIST.md.

Limitations

• Not validated on patients; healthy subjects do not prove detection of critical conditions.
• Quality thresholds are provisional (calibrated on synthetic signals).
• CRT is affected by temperature, age, hand position, and the force and duration of pressure; pressure force is not controlled.
• Exposure / white-balance lock is not available on every phone/browser; compensated by a reference patch plus a drift gate.
• The ROI is chosen manually (no automatic nail segmentation).
• Primary target is Chrome for Android; iOS Safari is best-effort (torch is reportedly supported since iOS 17.5 per WebKit, untested).
• Referral facility data is MOCK.

Privacy

All processing happens on the device; the core flow makes no network requests. Stored locally: derived color signals, one small fingertip thumbnail, and form inputs. Video is not stored. Per-frame ROI crops are stored only if enabled in Experiment (with the subject's consent). Delete individual items or all data from History/About. Consult Indonesian Law No. 27/2022 (Personal Data Protection) before using with real patient data.