# Baseline — Product Definition > **Know what moves your numbers.** > An AI lifestyle and blood-pressure companion that learns which of *your* habits raise *your* blood pressure — and proves it with your own data. **Status:** Proof of concept. Static HTML/CSS/JS. All data synthetic. **Register:** Product (app UI + admin console) with one brand surface (`index.html`). **Prepared by:** Shanti Infosoft · for client evaluation --- ## 1. The problem A person is told their blood pressure is "a bit high." They are given generic advice — eat less salt, sleep more, stress less, move more — and no way to know which of those actually matters *for them*. The evidence is clear that lifestyle drives blood pressure. The [2025 AHA/ACC guideline](https://www.jacc.org/doi/10.1016/j.jacc.2025.07.010) makes lifestyle modification the first-line intervention for every adult with elevated BP, and for the first time names **stress management** alongside diet, activity, sodium and alcohol. But the guideline speaks to populations. It cannot tell an individual that *their* systolic rises 7 mmHg the morning after a short night, while their sodium intake barely moves the needle. That gap — population advice versus personal causation — is the product. ## 2. What Baseline does 1. **Establishes a personal baseline.** Twice-daily readings under a guided protocol until there is enough signal to know what "normal for you" is. 2. **Captures the day in under 60 seconds.** A push-prompted micro-questionnaire covering sleep, stress, work strain, movement, diet and sodium, hydration, caffeine, alcohol, smoking, and medication adherence — most of it pre-filled from wearables. 3. **Finds the drivers.** A per-user model correlates lifestyle features against subsequent readings, accounting for lag, and reports effect sizes in mmHg with honest confidence. 4. **Tells the user what to change, in order.** Ranked, specific, quantified: *"Sleep under 6 hours raises your next-morning systolic by 7.4 mmHg. Based on 23 nights. High confidence."* 5. **Proves it worked.** n-of-1 experiments with pre/post readouts, and a monthly report that can go to a GP. ## 3. What Baseline is not - **Not a medical device.** Phase 1 ships as a wellness/lifestyle product. It does not diagnose, does not recommend or titrate medication, and does not replace clinical care. Non-diagnostic disclaimers appear on the risk score and the monthly report. - **Not a blood pressure measurement device.** It records readings from validated cuffs; it does not derive BP from a camera or a wrist optical sensor. - **Not a chatbot.** An LLM phrases insights from structured model output through a constrained template layer. It never generates clinical content freehand. ## 4. Users ### Primary — the member **Sipho Nkosi, 52 · Operations manager · Johannesburg** Told at a company wellness screening that readings were "a bit high" — 138/88. Not on medication yet. Owns a cuff bought at Clicks that is used inconsistently. Works long days, sleeps badly during quarter-end, drinks two or three beers most evenings, walks less than they think. Uses the app at 06:40 sitting on the edge of the bed with the cuff still on, and again around 21:30. Both are dim-room moments. Wants two things and nothing else: *is today okay*, and *what did I do yesterday that made it worse*. Anxious but not ill. Will abandon anything that feels like a hospital, and will abandon anything that takes more than a minute a day. ### Secondary — the clinician **Dr Naledi Khumalo · GP · Randburg** Sees the member for ten minutes twice a year. Does not want an app login. Wants a one-page summary of out-of-office readings, averages, variability, and adherence — the things the guideline now says matter more than the in-office reading. Receives the monthly report as a shared link or PDF. ### Tertiary — the operator **Platform ops and clinical safety team** Monitors engagement and adherence, triages hypertensive-crisis alerts, manages the questionnaire bank and notification schedules, reviews AI model versions and drift, and answers DSARs. Works in a browser at a desk all day. ## 5. Module register Every screen in the prototype carries a `data-sow-module` attribute mapping it to one of these. The annotation layer (`⌘K` or the flag icon) surfaces the mapping on screen, and `blueprint/08-estimate.html` costs the same register line by line. ### Patient application | # | Module | Covers | |---|---|---| | M1 | Account & onboarding | Sign-up, OTP, granular POPIA/GDPR consent, profile wizard, BP goal | | M2 | Daily check-in | Adaptive questionnaire engine, streak, snooze, wearable pre-fill | | M3 | BP capture | Guided measurement protocol, live classification, BLE cuff pairing, context tags | | M4 | Home & today | Status card, baseline delta, next action, learning state | | M5 | Trends & history | Timeline with category bands, AM/PM split, variability, scatter, heatmap | | M6 | AI insights | Ranked drivers, effect sizes, lag analysis, what-if simulator | | M7 | Risk score | Composite score, trajectory, contributors, PREVENT context | | M8 | Experiments & nudges | n-of-1 experiments, pre/post readout, adherence nudges | | M9 | Monthly report | Generated report, clinician share, PDF export | | M10 | Medications | Schedule, adherence tracking, reminders, refill | | M11 | Devices & data sources | HealthKit, Health Connect, Fitbit, BLE cuffs, sync log | | M12 | Care circle | Clinician share, caregiver read-only, emergency contact | | M13 | Settings & privacy centre | Notification windows, units, consent ledger, export, deletion, DSAR | | M14 | Clinical safety | Crisis detection, interstitial, escalation guidance | ### Admin console | # | Module | Covers | |---|---|---| | A1 | Operations dashboard | Engagement, adherence funnel, sync health, alert volume | | A2 | Member management | Search, profile view, permission-gated access, audited support actions | | A3 | Cohort analytics | BP distribution, category migration, retention, feature engagement | | A4 | Questionnaire builder | Versioned bank, conditional logic, localisation | | A5 | Notification manager | Segments, schedules, quiet hours, templates | | A6 | AI model console | Registry, versions, drift, cohort feature importance, retraining, review queue | | A7 | Clinical safety queue | Crisis alerts, escalation workflow, outcome logging | | A8 | Content library | Evidence-tagged recommendations mapped to drivers | | A9 | Compliance & audit | Consent ledger, access log, DSAR queue, retention, incidents | | A10 | Roles & access | RBAC, admin/clinician roles, API keys, integrations | ## 6. The core loop ``` ┌──────────────────────────────────────────────────┐ │ │ ▼ │ ┌────────┐ ┌─────────┐ ┌──────────┐ ┌─────────┐ │ │ PROMPT │──▶│ CAPTURE │──▶│ ANALYSE │──▶│ REFLECT │──┘ │ push │ │ BP + │ │ per-user │ │ drivers │ │ AM/PM │ │ check-in│ │ model │ │ + score │ └────────┘ └─────────┘ └──────────┘ └─────────┘ │ ▼ ┌────────┐ │ ACT │ │ n-of-1 │ │ expt. │ └────────┘ ``` The loop must close in under 60 seconds a day or the data stops arriving and the model starves. Every design decision in M2 and M3 is subordinate to that constraint: large tap targets, chips and steppers instead of text inputs, wearable pre-fill wherever possible, one question per card. ## 7. Clinical rules the product must obey Sourced from the 2025 AHA/ACC guideline. These are encoded in `assets/js/data.js` as a single source of truth, not scattered across screens. **Categories** — classification uses the higher of the two category assignments: | Category | Systolic | | Diastolic | |---|---|---|---| | Normal | `< 120` | and | `< 80` | | Elevated | `120–129` | and | `< 80` | | Stage 1 hypertension | `130–139` | or | `80–89` | | Stage 2 hypertension | `≥ 140` | or | `≥ 90` | | Hypertensive crisis | `> 180` | and/or | `> 120` | **Measurement protocol** — 5 minutes seated rest, back supported, feet flat, arm at heart level, no caffeine or exercise for 30 minutes prior. Two readings one minute apart, averaged. The app enforces this with a rest timer and takes the average, because the guideline now treats home monitoring as the diagnostic gold standard. **Lifestyle targets** used in recommendation copy: - Sodium `< 2300 mg/day`, ideally `< 1500 mg/day`; potassium-based salt substitutes encouraged - Alcohol ≤ 2 units/day (men), ≤ 1 unit/day (women); reduction or abstinence preferred - Physical activity 75–150 min/week - DASH-style eating pattern - Stress management — meditation, breathing, yoga — now a named first-line strategy - Weight management where BMI is elevated **Crisis handling** — any reading `> 180` systolic or `> 120` diastolic triggers an immediate full-screen interstitial directing the user to seek care. No AI insight, no recommendation, no dismissal into a normal flow. This path never involves the model. ## 8. Non-negotiables 1. **The numbers agree everywhere.** Every figure on every screen derives from `assets/js/data.js`. A prototype whose dashboard disagrees with its report is a prototype nobody believes. 2. **No insight without confidence.** The model gates on credible interval and sample size. Where confidence is low the app says so rather than inventing certainty. 3. **Colour is never the only signal.** Every BP state carries a label and a distinct glyph alongside its colour. 4. **The daily loop stays under 60 seconds.** 5. **Consent is granular and revocable.** Separate toggles per processing purpose, with a visible ledger and a working export/delete path. ## 9. Out of scope for the POC No backend, no real authentication, no persistence beyond `localStorage` for theme and tour state. No real device pairing — BLE and wearable flows are simulated at the UI layer. All member data is synthetic and labelled as such. Excel estimate sheet and formal Word SOW are deliberately deferred; both can be generated from `blueprint/08-estimate.html` and this document.