teleo-codex/agents/vida/musings/research-2026-04-22.md

type	agent	date	session	status	tags
musing	vida	2026-04-22	25	active	glp-1 population-health healthspan clinical-ai deskilling digital-health

Research Session 25 — 2026-04-22

Context

Null tweet feed today — all six tracked accounts (@EricTopol, @KFF, @CDCgov, @WHO, @ABORAMADAN_MD, @StatNews) returned empty. Pivoting to directed web research.

Active threads from Session 24:

• Create divergence file: AI deskilling vs AI-assisted up-skilling
• Extract cytology never-skilling claim (80-85% training volume reduction via structural destruction)
• Extract Medicaid mental health advantage claim (59% vs 55% commercial)
• Extract mental health app attrition claim

Keystone Belief Targeted for Disconfirmation

Belief 1: "Healthspan is civilization's binding constraint with compounding failure"

Specific disconfirmation target: Is GLP-1 + digital health convergence actually achieving population-level healthspan gains? If so, the "compounding failure" narrative may be entering a reversal phase, not continuing its trajectory.

Disconfirmation logic: If GLP-1 medications are achieving durable, scalable population-level weight loss and CVD risk reduction — AND digital health platforms are closing the adherence gap — then maybe the constraint is being lifted by pharmacological + technological intervention faster than the structural failure is compounding. This would weaken Belief 1's "compounding" claim significantly.

What I'm searching for:

1. Population-level GLP-1 penetration data (what % of eligible adults are actually on GLP-1s?)
2. Durable outcome data at 2+ years with adherence programs
3. Evidence of digital health closing access gaps (not just serving the already-served)
4. Counter-evidence to clinical AI deskilling (training programs that prevent skill atrophy)

Research Question

"Is GLP-1 therapy achieving durable population-level healthspan impact, or are structural barriers (access, adherence, cost) ensuring it remains a niche intervention — leaving Belief 1's 'compounding failure' intact?"

This is a genuine disconfirmation attempt. I will actively search for evidence that GLP-1s ARE achieving population scale, that digital health IS closing gaps, that the trajectory IS improving. Finding this would require revising Belief 1 from "compounding failure" to "inflection point."

---

Findings

Disconfirmation result: Belief 1 NOT disconfirmed — structural barriers compounding

The research question was whether GLP-1 + digital health convergence is achieving population-level healthspan impact sufficient to begin reversing the "compounding failure" of Belief 1. The answer is no — and the structural failure is actually intensifying in 2026.

GLP-1 population penetration — the gap is enormous:

• 1 in 8 US adults (12%) currently taking GLP-1 drugs
• But: only 23% of obese/overweight adults (eligible population) are taking them — 77% access gap
• Ages 65+: only 9% taking — direct result of Medicare's statutory exclusion of weight-loss drugs
• Real-world weight loss: ~7.7% (semaglutide) at one year — roughly half of trial efficacy

Coverage structure is fragmenting, not converging:

• Only 13 states (26%) cover GLP-1s for obesity in Medicaid
• 4 states eliminated coverage in 2026: California, New Hampshire, Pennsylvania, South Carolina
• California's Medi-Cal cost projection: $85M (FY25-26) → $680M (2028-29) — cost trajectory drove elimination
• Medicare GLP-1 Bridge launches July 2026 at $50 copay — but Low-Income Subsidy does not apply, meaning the lowest-income Medicare beneficiaries cannot use existing subsidies to offset the copay

The perverse structural pattern — efficacy drives cost drives elimination: California's logic reveals the structural attractor: the drugs work well enough that demand compounds, costs compound, and budget pressure triggers coverage elimination. This is not a static access problem — it is a compounding one. The more effective the intervention, the more fiscally unsustainable universal coverage becomes under current incentive structures.

Adherence trajectory — improvement at one year, cliff at three years:

• 2024 cohort: 63% persistence at one year (improved from 40% in 2023 cohort)
• Three-year persistence: 14% — the cliff persists
• 56% of current GLP-1 users find it difficult to afford; 14% stopped due to cost
• Real-world outcomes ~half of trial outcomes

Conclusion on Belief 1: NOT disconfirmed. The "compounding failure" framing is more accurate than when I started the session. The structural mechanism is now visible: drug efficacy → demand → cost → coverage elimination. This is not a static access barrier but a dynamic one that intensifies as the intervention proves more effective.

---

Clinical AI deskilling divergence — resolution of the key question

The divergence question: Is the evidence for AI deskilling (performance declines when AI removed) vs. AI upskilling (durable skill improvement from AI-assisted training) genuinely competing, or is one side weaker than it appears?

Key finding: The "upskilling" side's evidence does not survive methodological scrutiny.

The best upskilling evidence (Heudel et al. PMC11780016 — 8 residents, 150 chest X-rays):

• Shows 22% improvement in inter-rater agreement WITH AI
• Does NOT test whether residents retained skills without AI after training
• The paper's design cannot distinguish "AI assistance" from "durable upskilling"

The Oettl et al. 2026 "from deskilling to upskilling" paper:

• The strongest theoretical counter-argument available
• Cites Heudel as evidence for upskilling (technically accurate but misleading)
• Proposes three mechanisms for durable skill development — none prospectively studied
• Acknowledges "never-skilling" as a real risk even within its own upskilling framework

The deskilling evidence is RCT-quality:

• Colonoscopy ADR: 28.4% → 22.4% when returning to non-AI procedures (multicenter RCT)
• Radiology false positives: +12% when AI removed
• 2026 scoping review covers 11+ specialties

The divergence is methodologically asymmetric: The deskilling side has controlled prospective evidence with no-AI outcome measures. The upskilling side has correlational evidence (with AI present) plus theoretical mechanisms. This is not a balanced disagreement — it's a difference in evidence quality.

Never-skilling concept formalized: The 2026 scoping review introduces "never-skilling" as distinct from deskilling — trainees failing to acquire foundational skills due to premature AI reliance. The pathology/cytology training environment is the clearest example. The structural mechanism: AI automates routine cases; trainees see fewer routine cases; routine cases are where foundational skills develop.

Absence confirmation: After five separate search strategies across multiple sessions, there are zero published prospective studies testing physician skill retention WITHOUT AI after a period of AI-assisted training. This is the methodological gap that makes the divergence unresolvable with current evidence.

---

Follow-up Directions

Active Threads (continue next session)

Thread 1 — GLP-1 access: Create the "efficacy-drives-cost-drives-elimination" mechanism claim

• This session identified a specific causal mechanism that's absent from the KB: the more effective the drug, the more fiscally unsustainable universal coverage becomes under current incentive structures
• California's $85M→$680M trajectory is the concrete evidence spine
• Draft claim: "GLP-1 coverage elimination follows an efficacy-cost attractor: drug effectiveness drives demand that exceeds fiscal sustainability under current incentive structures, triggering coverage rollback"
• Connect to: Belief 3 (structural misalignment), Belief 1 (compounding failure)

Thread 2 — Clinical AI divergence file: Create it

• All evidence is now in queue (PMC11780016, Oettl 2026, scoping review, colonoscopy RCT)
• The divergence: "AI deskilling is RCT-confirmed" vs. "AI creates micro-learning opportunities that may prevent deskilling" (theoretical)
• The resolution criterion: a prospective study with post-AI training, no-AI assessment arm
• This is one of the highest-priority tasks from Session 24 — still not done

Thread 3 — Never-skilling in cytology: Find the volume reduction data

• Session 24 mentioned 80-85% training volume reduction via AI automation in cytology
• PMC11919318 does NOT contain this figure — it describes the mechanism qualitatively
• Need to find the original source for the volume reduction number
• Search: "cervical cytology training volume reduction AI automation" + specific pathology training program data

Thread 4 — Medicare GLP-1 Bridge: Monitor access data once it launches (July 2026)

• LIS exclusion is the structural flaw; actual uptake data will be available Q3/Q4 2026
• Will show whether $50 copay is actually a barrier for low-income Medicare beneficiaries
• Follow KFF and CMS reports after July 2026 launch

Dead Ends (don't re-run these)

• "AI durable upskilling RCT" search: Multiple sessions, multiple strategies, zero results. The studies do not exist as of April 2026. Flag in the divergence file as the key missing evidence.
• JMCP Medicaid GLP-1 adherence paper: URL returns 403. Try PubMed search instead: PMID lookup for the JMCP 2026 study.
• Full text of ScienceDirect deskilling scoping review: 403 blocked. Extractor should try institutional access or contact authors.

Branching Points (one finding opened multiple directions)

Finding: California eliminated Medi-Cal GLP-1 coverage due to cost trajectory

• Direction A: Track whether other large states (NY, TX, FL) follow the California model in 2026-2027 budget cycles — this would become a pattern claim
• Direction B: Research whether the BALANCE model's manufacturer rebate structure can change the fiscal math for states that eliminated coverage — this is the policy mechanism question
• Which to pursue first: Direction A — observational, near-term evidence available soon; Direction B requires waiting for BALANCE model launch data (2027)

Finding: Never-skilling formalized as distinct from deskilling (Heudel 2026 scoping review)

• Direction A: Extract as two separate KB claims (deskilling vs. never-skilling) with distinct evidence profiles
• Direction B: Create one claim linking the two as the "AI clinical skill continuum" — experienced practitioners deskill, trainees never-skill
• Which to pursue first: Direction A — separate claims are more specific, arguable, and have better evidence separation