teleo-codex/agents/vida/musings/research-2026-03-16.md

status	type	stage	created	last_updated	tags
seed	musing	developing	2026-03-16	2026-03-16	glp-1 adherence value-based-care capitation ai-healthcare clinical-ai epic abridge openevidence research-session

Research Session: GLP-1 Adherence Interventions and AI-Healthcare Adoption

Research Question

Can GLP-1 adherence interventions (care coordination, lifestyle integration, CGM monitoring, digital therapeutics) close the adherence gap that makes capitated economics work — or does solving the math require price compression to ~$50/month before VBC GLP-1 coverage becomes structurally viable?

Secondary question: What does the actual adoption curve of ambient AI scribes tell us about whether the "scribe as beachhead" theory for clinical AI is materializing — and does Epic's entry change that story?

Why This Question

Priority justification: The March 12 session ended with the most important unresolved tension in the entire GLP-1 analysis: MA plans are restricting access despite theoretical incentives to cover GLP-1s. The BALANCE model (May 2026 Medicaid launch) is the first formal policy test of whether medication + lifestyle can solve the adherence paradox. Three months out from launch is exactly when preparatory data should be available.

The secondary question comes from the research directive: AI-healthcare startups are a priority. The KB has a claim that "AI scribes reached 92% provider adoption in under 3 years" — but this was written without interrogating what adoption actually means. Is adoption = accounts created, or active daily use? Does the burnout reduction materialize? Is Abridge pulling ahead?

Connections to existing KB:

• Active thread: GLP-1 cost-effectiveness under capitation requires solving the adherence paradox (March 12 claim candidate)
• Active thread: MA plans' near-universal prior auth demonstrates capitation alone ≠ prevention incentive (March 12 claim candidate)
• Existing KB claim: "ambient AI documentation reduces physician documentation burden by 73 percent but the relationship between automation and burnout is more complex than time savings alone" — needs updating with 2025-2026 evidence

What would change my mind:

• If BALANCE model design includes an adherence monitoring component using CGM/wearables, that strengthens the atoms-to-bits thesis (physical monitoring solves the behavioral gap)
• If purpose-built MA plans (Devoted, Oak Street) are covering GLP-1s while generic MA plans restrict, that strongly validates the "VBC form vs. substance" distinction
• If AI scribe adoption is plateauing at 30-40% ACTIVE daily use despite 90%+ account creation, the "beachhead" theory needs qualification
• If AI scribe companies are monetizing through workflow data → clinical intelligence (not just documentation), the atoms-to-bits thesis gets extended

Direction Selection Rationale

Following active inference principles: these questions have the highest learning value because they CHALLENGE the attractor state thesis (GLP-1 question) and TEST a KB claim empirically (AI scribe question). Both are areas where I could be wrong in ways that matter.

GLP-1 adherence is the March 12 active thread with highest priority. AI scribe adoption is in the research directive and has a KB claim that may be stale.

---

What I Found

Track 1: GLP-1 Adherence — The Digital Combination Works (Observationally)

The headline finding: Multiple convergent 2025 studies show digital behavioral support substantially improves GLP-1 outcomes AND may reduce drug requirements:

1. JMIR retrospective cohort (Voy platform, UK): Engaged patients lost 11.53% vs. 8% body weight at 5 months. Digital components: live video coaching, in-app support, real-time weight monitoring, adherence tracking.

2. Danish digital + treat-to-target study: 16.7% weight loss at 64 weeks — matching clinical trial outcomes — while using HALF the typical semaglutide dose. This is the most economically significant finding: same outcomes, 50% drug cost.

3. WHO December 2025 guidelines: Formal conditional recommendation for "GLP-1 therapies combined with intensive behavioral therapy" — not medication alone. First-ever WHO guideline on GLP-1 explicitly requires behavioral combination.

4. Critical RCT finding on weight regain after discontinuation (the 64.8% scenario):

   • GLP-1 alone: +8.7 kg regain — NO BETTER than placebo (+7.6 kg)
   • Exercise-containing arm: +5.4 kg
   • Combination (GLP-1 + exercise): only +3.5 kg

The core insight this changes: The existing March 12 framing assumed the adherence paradox is about drug continuity — keep patients on the drug and they capture savings. The new evidence suggests the real issue is behavioral change that OUTLASTS pharmacotherapy. GLP-1 alone doesn't produce durable change; the combination does. The drug is a catalyst, not the treatment itself.

CLAIM CANDIDATE: "GLP-1 medications function as behavioral change catalysts rather than standalone treatments — combination with structured behavioral support achieves equivalent outcomes at half the drug cost AND reduces post-discontinuation weight regain by 60%, making medication-plus-behavioral the economically rational standard of care"

Track 2: BALANCE Model Design — Smarter Than Expected

The design is more sophisticated than the original March 12 analysis captured:

1. Two-track payment mechanism: CMS offering BOTH (a) higher capitated rates for obesity AND (b) reinsurance stop-loss. This directly addresses the two structural barriers identified in March 12: short-term cost pressure and tail risk from high-cost adherents.

2. Manufacturer-funded lifestyle support: The behavioral intervention component is MANUFACTURER FUNDED at no cost to payers. CMS is requiring drug companies to fund the behavioral support that makes their drugs cost-effective — shifting implementation costs while requiring evidence-based design.

3. Targeted eligibility: Not universal coverage — requires BMI threshold + evidence of metabolic dysfunction (heart failure, uncontrolled hypertension, pre-diabetes). Consistent with the sarcopenia risk argument: the populations most at cardiac risk from obesity get the drug; the populations where GLP-1 muscle loss is most dangerous (healthy elderly) are filtered.

4. Timeline: BALANCE Medicaid May 2026, Medicare Bridge July 2026, full Medicare Part D January 2027.

The March 12 question was: "does capitation create prevention incentives?" The BALANCE answer: capitation alone doesn't, but capitation + payment adjustment + reinsurance + manufacturer-funded lifestyle + targeted access might.

CLAIM CANDIDATE: "CMS BALANCE model's dual payment mechanism — capitation rate adjustment plus reinsurance stop-loss — directly addresses the structural barriers (short-term cost, tail risk) that cause MA plans to restrict GLP-1s despite theoretical prevention incentives"

Track 3: AI Scribe Market — Epic's Entry Changes the Thesis

Epic AI Charting launched February 4, 2026 — a native ambient documentation tool that queues orders AND creates notes, accessing full patient history from the EHR. Key facts:

• 42% of acute hospital EHR market, 55% of US hospital beds
• "Good enough" for most documentation use cases at fraction of standalone scribe cost
• Native integration is structurally superior for most use cases

Abridge's position (pre- and post-Epic entry):

• $100M ARR, $5.3B valuation by mid-2025
• $117M contracted ARR (growth secured even pre-Epic)
• Won top KLAS ambient AI slot in 2025
• Pivot announced: "more than an AI scribe" — pursuing real-time prior auth, coding, clinical decision support inside Epic workflows
• WVU Medicine expanded across 25 hospitals in March 2026 — one month after Epic entry (implicit market validation of continued demand)

The "beachhead" thesis needs revision: Original framing: "ambient scribes are the beachhead for broader clinical AI trust — documentation adoption leads to care delivery AI adoption." Epic's entry creates a different dynamic: the incumbent is commoditizing the beachhead before standalone AI companies can leverage the trust into higher-value workflows.

CLAIM CANDIDATE: "Epic's native AI Charting commoditizes ambient documentation before standalone AI scribes can convert beachhead trust into clinical decision support revenue, forcing Abridge and competitors to complete a platform pivot under competitive pressure"

Burnout reduction confirmed (new evidence): Yale/JAMA study (263 physicians, 6 health systems): burnout dropped from 51.9% → 38.8% (74% lower odds). Mechanism: not just time savings — 61% cognitive load reduction + 78% more undivided patient attention. The KB claim about burnout complexity is now supported.

Track 4: OpenEvidence — Beachhead Thesis Holds for Clinical Reasoning

OpenEvidence operates in a different workflow (clinical reasoning vs. documentation) and is NOT threatened by Epic AI Charting:

• 40%+ of US physicians daily (same % as existing KB claim, much larger absolute scale)
• 20M clinical consultations/month by January 2026 (2,000%+ YoY growth)
• $12B valuation (3x growth in months)
• First AI to score 100% on USMLE (all parts)
• March 10, 2026: first 1M-consultation single day

The benchmark-vs-outcomes tension is now empirically testable at this scale. Concerning: 44% of physicians still worried about accuracy/misinformation despite being heavy users. Trust barriers persist even in the most-adopted clinical AI product.

Key Surprises

1. Digital behavioral support halves GLP-1 drug requirements. At half the dose and equivalent outcomes, GLP-1s may be cost-effective under capitation without waiting for generic compression. This is the most important economic finding of this session.

2. GLP-1 alone is NO BETTER than placebo for preventing weight regain. The drug doesn't create durable behavioral change — only the combination does. Plans that cover GLP-1s without behavioral support are paying for drug costs without downstream savings.

3. BALANCE model's capitation adjustment + reinsurance directly solves the March 12 barriers. CMS has explicitly designed around the two structural barriers I identified. The question is whether plans will participate and whether lifestyle support will be substantive.

4. Epic's AI Charting is the innovator's dilemma in reverse. The incumbent is using platform position to commoditize the beachhead. Abridge must complete a platform pivot under competitive pressure.

5. OpenEvidence at $12B valuation with 20M monthly consultations. Clinical AI at scale — but the outcomes data doesn't exist yet.

Belief Updates

Belief 3 (structural misalignment): PARTIALLY RESOLVED. The BALANCE model's dual payment mechanism directly addresses the misalignment identified in March 12. The attractor state may be closer to policy design than I thought.

Belief 4 (atoms-to-bits boundary): REINFORCED for physical data, COMPLICATED for software. Digital behavioral support is the "bits" that makes GLP-1 "atoms" work — supporting the thesis. But Epic's platform move shows pure software documentation AI is NOT defensible against platform incumbents. The physical data generation (wearables, CGMs) IS the defensible layer; documentation software is not.

Existing GLP-1 claim: Needs further scope qualification beyond March 12's payer-level vs. system-level distinction. The half-dose finding changes the economics under capitation if behavioral combination becomes the implementation standard.

---

Follow-up Directions

Active Threads (continue next session)

• BALANCE model Medicaid launch (May 2026): The launch is in 6 weeks. Look for: state Medicaid participation announcements, manufacturer opt-in/opt-out decisions (Novo Nordisk, Eli Lilly), early coverage criteria details. Key question: does the lifestyle support translate to structured exercise programs, or just nutrition apps?

• GLP-1 half-dose + behavioral support replication: The Danish study is observational. Look for: any RCT directly testing dose reduction + behavioral combination, any managed care organization implementing this protocol. If replicated in RCT, it changes GLP-1 economics more than any policy intervention.

• Abridge platform pivot outcomes (Q2 2026): Look for revenue data post-Epic entry, any contract cancellations citing Epic, KLAS Q2 scores, whether coding/prior auth capabilities are gaining traction. The test: can Abridge maintain growth while moving up the value chain?

• OpenEvidence outcomes data: 20M consults/month creates the empirical test for benchmark-vs-outcomes translation. Look for any population health outcomes study using OpenEvidence vs. non-use. This is the missing piece in the clinical AI story.

Dead Ends (don't re-run these)

• Tweet feeds: Four sessions, all empty. The pipeline (@EricTopol, @KFF, @CDCgov, @WHO, @ABORAMADAN_MD, @StatNews) produces no content. Do not open sessions expecting tweet-based source material.

• Devoted Health GLP-1 specifics: No public data distinguishing Devoted's GLP-1 approach from generic MA plans. Plan documents confirm PA required; no differentiated protocols available publicly.

• Compounded semaglutide: Flagged as dead end in March 12; confirmed. Legal/regulatory mess, not analytically relevant.

Branching Points (one finding opened multiple directions)

• GLP-1 + behavioral combination at half-dose:

  • Direction A: Write the standard-of-care claim now (supported by convergent observational + WHO guidelines), flag experimental until RCT replication
  • Direction B: Economic modeling of capitation economics under half-dose + behavioral assumptions
  • Recommendation: A first. Write the claim now; flag for RCT replication. Direction B is a Vida + Rio collaboration.

• Epic AI Charting threat:

  • Direction A: Write a claim about Epic platform commoditization of documentation AI (extractable now as a structural mechanism)
  • Direction B: Track Abridge pivot metrics through Q2 2026 and write outcome claims when market structure is clearer
  • Recommendation: A for mechanism, B for outcome. The commoditization dynamic is extractable now. Abridge's fate needs 6-12 months more data.

SOURCE: 9 archives created (7 new + 2 complementing existing context)