Balance Meter — How It Scales (No-Wearable → +Mark I → +Mark II)¶
Date: 2026-06-07 · Author: Gene Lang, PharmD — Director of Pharmacy Operations · Rx360
For: Jamie (product) · Matilda (clinical) · Peyman · counsel
Status: ✅ Current product direction — the landed Balance Meter model, re-projected onto three sensing envelopes (one model, graceful degradation).
Spine: Fall_Risk/Balance_Meter_Canonical_Model_2026-05-20.md (the locked two-layer model). Any weight/threshold here traces to that document or is labeled a proposal.
The shift to flag up front: this spec makes the Balance Meter member-facing without a wearable — questionnaire + medications alone produce a member reading on day one. That extends member-facing output below Mark II, which revisits the 2026-04-28 positioning decision ("consumer output begins Mark II"). It is surfaced as a named sign-off (Decision 1), not assumed.
⟳ Reconciled 2026-06-07 to
Balance_Meter_Aligned_Direction_2026-06-07(the re-grounding against the 6/02 decision kit + 6/03 Master Plan). Carry these corrections through this spec: - Band words (LOCKED 2026-06-29): Steady / Some Change / Worth a Look — the wellness-lane register of Oura's "Pay attention" / Whoop's recovery states: calm, awareness-prompting, never clinical-flat or alarming. - Med-change card = NEUTRAL, no attribution (Matilda-approved, locked 6/07): "You entered a new medication this week. It can be helpful to speak to your pharmacist when your medicines change." — no "likely tied to the new [drug]," no "when convenient," never name a drug as the cause. The causal gait↔med link is a backend/clinical view only. - Per the 6/02 kit the consumer band is a Mark II surface; Mark I = "Phase Zero" (not surfaced). Surfacing a member read at No-Wearable / +Mark I is exactly the proposed extension (Decision 1) — keep it flagged as the room's call, not assumed. - Insulin / high-stakes meds are served via the user-initiated optional second reminder (Patent Brief §3.6) — the system never labels a drug "time-critical." (Replaces the stale "insulin off the band" note.) - The recurring "30-second self-check" idea (added to an earlier mockup) is retracted — it was not the room's design; the only performance test is the one-time Mark-II onboarding stand check.
1 · Framing — one model, the inputs scale¶
Per Jamie, we need three versions of the Balance Meter by what's available:
- No-Wearable — questionnaire + medications, no device.
- +Mark I — adds the Mark I band's motion sensor (IMU gait).
- +Mark II — adds the full Mark II sensor suite (barometer, wear-detect, optional PPG, Pill Bottle).
The important design insight: these are not three algorithms. The locked canonical model is already two layers —
- Layer 1 — Baseline risk: a capped 0–100 composite from fall history, medications, conditions, mobility & function, sensory/physiologic, and mood, computed at onboarding. Needs no wearable.
- Layer 2 — Live monitoring: gait change from the wearer's personal baseline → the band (Steady / Some Change / Worth a Look). Sensor-dependent.
So the three versions are one model where Layer 1 is constant (the backbone, present with zero hardware) and Layer 2 is progressively enriched. This honors the "one weighting decision" rule — we lock one model and re-project it onto three input sets — and it matches the graceful-degradation principle already in the 5-20 explainer deck: "the band still works with no pharmacist and no Medicare link… verification upgrades confidence; it is never a gate." The medication layer is the through-line in all three — Pharmacy Operations' lane, and the reason No-Wearable is viable on day one.
2 · The architecture — three input regimes of one model¶
How the gait signals are actually computed: see
Product_Specs/Fall_Barometer_Gait_Biomechanics_Signal_Pipeline_2026-06-30.md— per-signal sensor + computation method + fall-risk threshold + Apple HealthKit precedent, cited. This table names the inputs; that doc specifies how they're derived from raw IMU samples (the gap this table's numbers assumed was already solved).
| Dimension | No-Wearable | +Mark I | +Mark II |
|---|---|---|---|
| Inputs | STEADI-Rx questionnaire + meds + conditions + mobility + mood | + IMU gait (speed, asymmetry, step-timing variability) | + barometer (fall-confirm) + full gait + optional PPG + Pill Bottle adherence |
| Layer 1 (baseline) | Full (0–100) | Full | Full |
| Layer 2 (live) | None (no live stream) | Gait-only (no altitude, no physiologic) | Full |
| Member output | A standing balance read (profile-based) | A live band at reduced precision | A live band, full precision |
| Signal weight vs full | Baseline-only | ~60–70% of the live framework | 100% |
| Confidence surfaced to member | Screening-grade ("based on your profile + meds") | Moderate ("early movement signal") | Full (unqualified) |
| Calibration track | Layer 1 (med portion done) | + IMU gait-Δ thresholds | + full-sensor calibration |
| Surfaces | Member card (standing) + clinical view | Member card (live) + clinical | Member + caregiver + Base Station + clinical |
The rule that ties them together: Layer 1 is always fully computed (its inputs exist in every version); Layer 2 is what scales. A version with less sensing does not get a lower score — it gets a wider confidence envelope and a capped claim (Decision 6a). Less data never means "safer-looking."
3 · No-Wearable — Layer 1 only, member-facing¶
Inputs (all present without a device): the STEADI-Rx 12-item screen (fall history, mobility, sensory, mood) + the pharmacist-verified medication layer (FRID by class with diminishing returns, polypharmacy, the cognitive + polypharmacy interaction bonus) + the 8 named conditions. This is the full Layer 1 baseline (caps: fall history 30, meds 22, conditions 22, mobility 18, sensory 14, mood 10 → ceiling 100).
What the member sees — and the design problem we must solve: the canonical wearer bands (Steady / Some Change / Worth a Look) describe change, and No-Wearable has no live stream to change against. So No-Wearable cannot show a "Some Change" band. It shows a standing read — a wellness "balance profile," not a live trend. The recommendation:
- A standing, plain-language read derived from the Layer 1 band cutoffs (elevated ≈ 45, high ≈ 70 in the clinical view) — but rendered in wellness-safe, non-diagnostic language, never "your fall risk is High."
- Every read carries a confidence qualifier ("based on your health profile and medications") and the approved soft prompt ("Evaluation by your healthcare provider may be helpful").
- It explicitly sets up the upgrade: "Add the band to see this update as your movement changes."
Worked member string (proposed, pending lexicon sign-off): "Your balance profile looks steady. This is based on your health history and medications — adding the Rx360 band will let it track day-to-day changes. Worth mentioning your profile to your pharmacist at your next visit."
Claim ceiling: No-Wearable is a composite risk index informed by validated screens — not a validated fall predictor (the canonical model's own language). Member framing is a wellness profile, never a prediction. This version drifts closest to "health-risk-assessment," so its strings need the heaviest lexicon/legal discipline (Decision 2). No "fall risk," "frailty," or outcome language.
Calibration: the medication portion is already calibrated — the cap of 22 lands at the 95th percentile of 754 anchor-site seniors. The remaining Layer 1 weights are literature/odds-ratio-grounded but not yet validated against fall outcomes (that needs the pilot). Label accordingly.
4 · +Mark I — Layer 1 + IMU gait¶
Adds: the Mark I band's IMU motion-tier → gait speed Δ (the −15% delta gate before scoring), step-timing variability, walking asymmetry, all compared to the wearer's personal baseline. No barometer (no altitude-verified fall confirmation) and no PPG (no physiologic tier).
What changes for the member: a live band becomes possible — Steady / Some Change / Worth a Look now reflects actual movement, not just profile. But at ~60–70% of the full framework's signal weight (the documented Mark I constraint), so:
- The band ships with a wider confidence interval and a slightly higher delta gate (single-sensor gait is noisier — require the drift to clear a margin / persist before tripping a band) so we don't over-call on IMU-only noise.
- Acute-event (hard fall) confirmation is IMU-only / manual — no barometric altitude check — so fall-confirmation language is softer than Mark II.
Positioning note: a member-facing band on Mark I is exactly what Decision 1 reopens — today's locked rule says Mark I has no consumer card. This version assumes that rule is lifted.
5 · +Mark II — Layer 1 + full Layer 2¶
Adds: the full Mark II minimum set — barometric pressure (height-loss-to-floor fall confirmation, Tier A), dedicated wear-detection (gates all alerts), the full gait suite, optional PPG (physiologic corroboration), and the Pill Bottle adherence signal feeding the medication weighting. This is the productized RxBALANCE and the current canonical target — full confidence, all four surfaces (member, caregiver, Base Station, clinical), auto-emergency calling.
Nothing in the model changes here; this is the version the 5-20 canonical was written for. The other two versions are this one with Layer 2 progressively removed.
6 · The four cross-cutting design decisions (the engineering heart)¶
These four are what make "three versions" safe rather than three diverging products.
a. Missing-input handling — widen and cap, never zero-out¶
Layer 1 is always full, so no re-normalization is needed there. Where inputs are missing is Layer 2: - No-Wearable: no Layer 2 → output is the standing Layer 1 read with a widened confidence envelope and a capped claim (a profile, not change-detection). - +Mark I: partial Layer 2 → re-normalize the live signal across the available gait metrics, widen the band's confidence interval, and raise the delta gate — do not zero the missing tiers. Zeroing them would make a Mark I wearer look steadier than reality. The invariant: a missing input may only widen uncertainty or cap the claim — it may never lower the apparent reading. - +Mark II: full.
b. Per-version confidence + calibration tracks¶
Each version declares its own confidence, surfaced to the member as a qualifier (No-Wearable: "based on your profile + meds"; +Mark I: "+ early movement signal"; +Mark II: unqualified). Three calibration tracks, one model: No-Wearable = Layer 1 (med portion at p95 on 754 anchor-site seniors; the rest pilot-validated); +Mark I = IMU gait-Δ thresholds; +Mark II = full-sensor. Calibration is downstream (Decision 5) — this spec sets the method, not the numbers.
c. Version-transition continuity — the sneaky one¶
Stamp every reading with the version that produced it (provenance on the score). When a member upgrades (No-Wearable → +Mark I → +Mark II), the basis changes — they gain or enrich the live layer. Without care, that basis change looks like a risk spike. So: - Re-baseline the personal-normal window at every upgrade (treat an upgrade like a "change event" — the same 7–14-day re-baseline the model already uses after a med change), so the new sensor stream establishes its own baseline before it can trip a band. - Frame the upgrade as a precision gain — "your reading just got sharper" — never a "Worth a Look." This ties directly to the personal-baseline framing already in the canonical model.
d. Evidence + claim discipline per version¶
- No-Wearable validates on STEADI-Rx, Seppälä FRID, Beers 2023, the Drug Burden Index, and the Layer-1 odds-ratio anchors (Deandrea, Mol, Ellmers). Claim ceiling = wellness profile.
- +Mark I / +Mark II add the gait-variability (Hausdorff) and Timed-Up-and-Go evidence for the live layer.
- All three stay wellness-lane; the lower the sensing, the more the strings are confidence-qualified and HCP-prompted.
7 · Member experience across versions¶
| No-Wearable | +Mark I | +Mark II | |
|---|---|---|---|
| Reading type | Standing profile | Live band (reduced precision) | Live band (full) |
| Example member line | "Your balance profile looks steady — based on your history + meds." | "Your balance has been steady this week." (with a wider confidence note) | "Your balance has been steady this week." (on a med change → the neutral pharmacist nudge, no drug named) |
| Confidence qualifier | Always | When precision is reduced | None |
| HCP soft prompt | Always present | Present | As warranted |
| Caregiver view | Decision 4 | Decision 4 | Daily digest + band-change push |
The upgrade narrative is a feature: the same member can start at No-Wearable (day one, no hardware), add the band (live tracking begins), and reach Mark II (full precision) — each step framed as "your reading just got sharper." All member-facing strings stay inside the approved lexicon (Steady / Some Change / Worth a Look; "Evaluation by your healthcare provider may be helpful"; no "fall risk," "frailty," "detects gait abnormalities," "prevents falls," "medical-grade," or "clinically validated").
8 · The medication layer is the backbone (Pharmacy Operations' lane)¶
The medication layer is the one input present in all three versions — it's what makes No-Wearable viable with zero hardware, and it is our strategic moat. It is provenance-weighted (self-reported ×0.5 → pharmacist-verified ×1.0; verification upgrades confidence, never gates entry), uses diminishing returns within each FRID class (100% → 70% → 50% → 30%), the one interaction bonus (cognitive + polypharmacy), and the calibrated cap of 22. It draws on the same curated drug intelligence we built for the medication-input engine — the FRID / Beers / NTI tagging schema and the Smart-Scheduling forgiveness data — so the Balance Meter and the medication engine share one source of drug truth. This is why the three versions reinforce each other rather than fragment: same medication brain, three sensing envelopes.
Companion deck: Balance_Meter_Three_Versions_Deck_2026-06-07.pptx (pre-read for Jamie). Spine: Balance_Meter_Canonical_Model_2026-05-20.md. Reuse: the medication-engine guardrails + tagging schema + the Smart-Scheduling forgiveness KB.