vida: research 2026 04 14 #2918
8 changed files with 511 additions and 0 deletions
114
agents/vida/musings/research-2026-04-14.md
Normal file
114
agents/vida/musings/research-2026-04-14.md
Normal file
|
|
@ -0,0 +1,114 @@
|
|||
---
|
||||
type: musing
|
||||
agent: vida
|
||||
date: 2026-04-14
|
||||
status: active
|
||||
session: 24
|
||||
---
|
||||
|
||||
# Research Session 24 — 2026-04-14
|
||||
|
||||
## Research Question
|
||||
|
||||
**For patients with obesity and heart failure (HFpEF/HFrEF), does GLP-1 therapy represent a clinical intervention powerful enough to challenge the McGinnis-Foege claim that 80-90% of health outcomes are determined outside medical care — and what does clinical AI adoption do to the equity distribution of that benefit?**
|
||||
|
||||
This is a dual-vector question:
|
||||
1. **Disconfirmation vector**: If GLP-1 therapy achieves NNT=9 for all-cause mortality in HFrEF patients (HR 0.62), does this challenge the 80-90% SDOH-dominance claim when applied to specific high-risk populations?
|
||||
2. **Compounding vector**: Does clinical AI deployment in resource-rich settings widen the access gap for the populations who most need these interventions — creating a meta-SDOH effect?
|
||||
|
||||
## Belief Targeted for Disconfirmation
|
||||
|
||||
**Belief 2: 80-90% of health outcomes are determined outside medical care (SDOH dominance)**
|
||||
|
||||
Attack vector: NNT=9 for all-cause mortality from a clinical intervention (GLP-1 in HFrEF) is an extremely strong effect size. The McGinnis-Foege 80-90% figure is a population-level average — but if specific high-risk populations can have their outcomes dramatically shifted by a single drug, what does the average mean?
|
||||
|
||||
**What I searched for:**
|
||||
- New real-world evidence on GLP-1 cardiovascular outcomes (found PMC12664052 — large HFrEF cohort, very high value)
|
||||
- Clinical AI and health disparities evidence (found multiple sources suggesting AI widens disparities in current deployment)
|
||||
- USPSTF updates on GLP-1 (found none — confirms prior session's gap finding)
|
||||
|
||||
**Disconfirmation result:**
|
||||
The belief survives with scope clarification. The 80-90% figure is a population-level average that describes variance in outcomes across populations. A strong clinical intervention (GLP-1, NNT=9 in HFrEF) does not contradict this — it says: within the 10-20% of outcomes clinical care *does* influence, some interventions are extremely potent. The SDOH framing was never about clinical care being useless; it was about the marginal value of more clinical care spending vs. addressing upstream factors. The GLP-1 HFrEF result is more appropriately read as: *even within clinical care's constrained sphere of influence, some interventions are dramatically underutilized* — especially in underserved populations where access to GLP-1s is being actively cut.
|
||||
|
||||
## Key Findings This Session
|
||||
|
||||
### 1. GLP-1 in HFrEF: Strongest Real-World Mortality Evidence Yet
|
||||
PMC12664052 (TriNetX database, n=13,098 matched per arm, T2D + HFrEF):
|
||||
- All-cause mortality HR 0.62 (95% CI 0.59–0.66, p<0.001)
|
||||
- Absolute risk reduction 11.0% (13.3% vs. 24.3%)
|
||||
- NNT = 9.1 over median 3.9 years
|
||||
- Semaglutide specifically: HR 0.51
|
||||
- Benefits consistent across arrhythmia endpoints (contradicting prior caution about fluid retention/arrhythmia risk in HFrEF)
|
||||
- This extends the KB's existing HFpEF evidence into HFrEF — the cardiovascular benefit is now documented across both HF phenotypes in real-world populations
|
||||
|
||||
### 2. Clinical AI Widens Health Disparities in Current Deployment
|
||||
Three converging sources (PMC11922879, PMC11796235, FAS policy brief):
|
||||
- ~40% of healthcare algorithms show bias against Black patients in resource allocation
|
||||
- Population health management algorithms allocated more care to white patients with equivalent health needs
|
||||
- AI translation tools narrowed disparities in specific use cases (43% fewer communication errors), but only with deliberate equity-centered design
|
||||
- The evidence is directional but mostly observational — no RCT showing AI reduces SDOH-driven gaps at scale
|
||||
- Assessment: AI deployment as currently practiced amplifies the clinical-care access advantage of already-advantaged populations. This is a compounding factor in the access inversion thesis.
|
||||
|
||||
### 3. GLP-1 Discontinuation → Higher CV Risk
|
||||
Truveta EHR post-market surveillance (1-year follow-up):
|
||||
- Early GLP-1 discontinuation linked to higher risk of coronary artery disease and heart failure
|
||||
- Applies to both semaglutide and tirzepatide
|
||||
- Convergent with the continuous-treatment-required thesis, now extended to CV outcomes not just weight regain
|
||||
|
||||
### 4. GLP-1 Micronutrient Deficiency Scale: 22.4% at 12 Months
|
||||
PMC12205620 (n=461,382 retrospective cohort):
|
||||
- 12.7% nutritional deficiency at 6 months, 22.4% at 12 months
|
||||
- Dietitian consultation substantially increased deficiency detection (18.5% vs. 12.2%)
|
||||
- 92% of patients had no dietitian consultation in 6 months prior to prescription (from Urbina)
|
||||
- Converges with Urbina review already in archive — confirms at population scale
|
||||
|
||||
### 5. Clinical AI Deskilling Evidence Portfolio Growing
|
||||
New upskilling evidence from PMC12955832 (orthopaedics) provides genuine tension with deskilling data:
|
||||
- AI-assisted imaging improved diagnostic accuracy in junior physicians
|
||||
- AI scribes reduced note time (9.5% in one study)
|
||||
- BUT: "never-skilling" concern for trainees remains — failing to acquire foundational skills when AI is introduced before competency develops
|
||||
- This is NOT a contradiction of the colonoscopy deskilling finding — it's a scope refinement: AI may upskill novices on discrete tasks while deskilling experienced providers on complex judgment
|
||||
|
||||
## Synthesis: Access Inversion Compound Structure
|
||||
|
||||
The session revealed a compound access inversion pattern across three layers:
|
||||
|
||||
**Layer 1 — Drug access**: GLP-1s with proven mortality benefit (NNT=9 in HFrEF) are being cut from Medicaid while remaining accessible to commercially insured patients.
|
||||
|
||||
**Layer 2 — Support infrastructure**: Even among GLP-1 users, the 92% no-dietitian-consultation monitoring gap is likely worse in lower-income populations (no stratified data but inference is reasonable).
|
||||
|
||||
**Layer 3 — Clinical AI equity**: AI tools that could theoretically help underserved populations access better care are, in current deployment, more biased and less accurate for those populations.
|
||||
|
||||
Each layer compounds the others. The populations most likely to have HFrEF + obesity (older, lower income, more diverse) are the same populations being cut from GLP-1 access, least supported by nutritional monitoring infrastructure, and least well-served by current AI clinical tools.
|
||||
|
||||
CLAIM CANDIDATE: "GLP-1 access inversion in heart failure creates compound harm: populations with highest cardiovascular mortality benefit are simultaneously being cut from Medicaid coverage, undermonitored for micronutrient deficiencies, and least well-served by clinical AI tools"
|
||||
|
||||
This is a synthesis claim connecting evidence from multiple domains. Hold for scope qualification before extracting — needs to be grounded specifically in the HFrEF mortality evidence + access cut data.
|
||||
|
||||
## Follow-up Directions
|
||||
|
||||
### Active Threads (continue next session)
|
||||
|
||||
- **GLP-1 HFrEF Claim**: PMC12664052 is ready to extract. Write as a new claim: "GLP-1 receptor agonists reduce all-cause mortality by 38% in T2D + HFrEF patients (NNT=9) based on real-world evidence." This is the most important new claim from this session. Check for divergence with existing HFpEF evidence — both show 27-58% benefit estimates.
|
||||
|
||||
- **Clinical AI Equity Divergence**: The AI-widens-disparities evidence (PMC11922879, PMC11796235, FAS) creates a real tension with the claim that clinical AI has potential to reduce health disparities. Write as a new divergence file: `divergence-clinical-ai-equity-widens-vs-narrows.md`. The resolution question is: under what design conditions does AI narrow vs. widen disparities?
|
||||
|
||||
- **GLP-1 Discontinuation → CV Outcomes**: Truveta data adds a new mechanism (not just weight regain but CV outcome risk from discontinuation). Archive and flag for extraction after GLP-1 HFrEF claim is merged.
|
||||
|
||||
- **Compound Access Inversion Synthesis Claim**: The three-layer structure (drug access + monitoring gap + AI equity) is worth a synthesis claim but needs to be carefully scoped to avoid overreach. Draft in musing before proposing.
|
||||
|
||||
### Dead Ends (don't re-run these)
|
||||
|
||||
- **WHO GLP-1 guideline Dec 2025**: Already in archive as `2025-12-01-who-glp1-global-guidelines-obesity.md` — don't re-search or re-archive
|
||||
- **USPSTF GLP-1 recommendation search**: Confirmed no update exists as of April 2026 — don't re-run this search
|
||||
- **Urbina narrative review**: Already in archive — `2026-01-xx-urbina-clinical-obesity-glp1-micronutrient-narrative-review.md`
|
||||
- **Frontiers Nutrition cross-sectional n=69**: Already in archive — `2025-03-xx-frontiers-nutrition-glp1-nutrient-intake-crosssectional.md`
|
||||
- **Lancet deskilling colonoscopy**: Already in archive — `2025-08-xx-lancet-preserving-clinical-skills-ai-deskilling.md`
|
||||
|
||||
### Branching Points (one finding opened multiple directions)
|
||||
|
||||
- **GLP-1 HFrEF NNT=9 finding** → Direction A: Extract as new CVD outcomes claim (immediate, high priority). Direction B: Use as evidence for Belief 2 scope qualification (medium priority, needs synthesis framing). Pursue A first.
|
||||
|
||||
- **Clinical AI equity evidence** → Direction A: Write divergence file on AI equity (narrows vs. widens). Direction B: Enrich existing clinical AI claim with equity dimension. Direction A is more valuable — the divergence structure invites resolution.
|
||||
|
||||
- **Access inversion compound structure** → Direction A: Synthesis claim connecting GLP-1 access cuts + AI equity + monitoring gap. Direction B: Three separate claims on each layer. Direction B is safer (easier to meet quality bar); Direction A is more intellectually interesting. Pursue B first, then attempt A synthesis.
|
||||
|
|
@ -623,4 +623,12 @@ On clinical AI: a two-track story is emerging. Documentation AI (Abridge territo
|
|||
- Belief 3 (structural misalignment): **UNCHANGED** — OBBBA Medicaid work requirement December 2026 mandatory national deadline is the most concrete expression of structural misalignment yet.
|
||||
|
||||
**Sources archived this session:** 8 (BCBS/Prime GLP-1 adherence doubling, Lancet metabolic rebound, SCORE/STEER real-world CV, JACC Stats 2026, HFSA 2024/2025, Danish digital GLP-1 program, GLP-1 nutritional deficiency, OBBBA SNAP cuts, OBBBA Medicaid work requirements, STEER semaglutide vs tirzepatide cardiac mechanism)
|
||||
|
||||
## Session 2026-04-14
|
||||
**Question:** For patients with obesity and heart failure (HFpEF/HFrEF), does GLP-1 therapy represent a clinical intervention powerful enough to challenge the McGinnis-Foege claim that 80-90% of health outcomes are determined outside medical care — and what does clinical AI adoption do to the equity distribution of that benefit?
|
||||
**Belief targeted:** Belief 2 — 80-90% of health outcomes determined outside medical care (SDOH dominance). Attack vector: NNT=9 for all-cause mortality from GLP-1 in HFrEF is a clinically powerful effect size — if clinical care only influences 10-20% of outcomes, how does a single drug move mortality by 38%?
|
||||
**Disconfirmation result:** Belief 2 survives with scope clarification. The McGinnis-Foege 80-90% figure is a population-level average describing variance in outcomes across populations — not a ceiling on what clinical care can achieve in a specific intervention. GLP-1 achieving NNT=9 in HFrEF is not a contradiction; it says clinical care's 10-20% sphere of influence contains some extremely potent interventions that are currently dramatically underutilized. The belief needs a scope qualifier: "at the population level, 80-90% of variance in health outcomes is explained by factors outside clinical care." At the individual high-risk-patient level, specific interventions can dominate. The disconfirmation search produced a scope refinement, not a falsification.
|
||||
**Key finding:** PMC12664052 (TriNetX, n=13,098 matched per arm, T2D + HFrEF, median follow-up 3.9 years): GLP-1s reduce all-cause mortality HR 0.62 (NNT=9.1), all-cause hospitalization HR 0.71, and reduce arrhythmia risk (AF HR 0.92, ventricular arrhythmia HR 0.86) — directly contradicting prior caution about GLP-1 use in HFrEF. Semaglutide specifically achieved HR 0.51 for mortality. This is the largest real-world evidence base for GLP-1 mortality benefit in HFrEF and should be the anchor for a new claim.
|
||||
**Pattern update:** The compound access inversion pattern is now documented across three layers: (1) GLP-1 drug access being cut from Medicaid while mortality benefits are confirmed; (2) 92% of GLP-1 users without dietitian monitoring, with the monitoring gap likely worse in lower-income populations; (3) Clinical AI tools that theoretically could extend access are instead confirmed to widen disparities in current deployment (~40% of healthcare algorithms show bias against Black patients, population health management algorithms allocating less care to Black patients with equivalent health needs). Each layer compounds the others in the same high-risk population (older, lower-income, more diverse patients with T2D + HFrEF). This three-layer structure is a synthesis claim candidate.
|
||||
**Confidence shift:** Belief 2 (SDOH dominance) unchanged but scope-qualified. The belief applies at population-level variance explanation; it does not preclude potent clinical interventions within the clinical-care sphere. No other belief shifted this session. The disconfirmation attempt produced a productive scope clarification, not a confidence change.
|
||||
**Extraction candidates:** GLP-1 continuous-treatment dependency claim (generalization from two intervention types); CVD bifurcation updated with JACC/HFSA data; clinical AI deskilling confidence upgrade; semaglutide GLP-1R cardiac mechanism (speculative); GLP-1 nutritional deficiency as population-level safety signal
|
||||
|
|
|
|||
|
|
@ -0,0 +1,56 @@
|
|||
---
|
||||
type: source
|
||||
title: "AI performs poorly for darker skin, lower-income populations in current deployment: 129-article systematic review"
|
||||
author: "PMC11796235 (multiple authors)"
|
||||
url: https://pmc.ncbi.nlm.nih.gov/articles/PMC11796235/
|
||||
date: 2025-01-01
|
||||
domain: health
|
||||
secondary_domains: [ai-alignment]
|
||||
format: peer-reviewed-systematic-review
|
||||
status: unprocessed
|
||||
priority: medium
|
||||
tags: [clinical-ai, health-equity, disparities, bias, dermatology, systematic-review]
|
||||
flagged_for_theseus: ["systematic review of AI performance degradation in underrepresented populations; training data bias creates structural performance gap — relevant to AI safety and deployment governance"]
|
||||
---
|
||||
|
||||
## Content
|
||||
|
||||
Systematic review of AI performance in diverse populations. 129 articles reviewed (screened 1,173). Published PMC11796235.
|
||||
|
||||
**Dermatological AI:** Accurate for light-skinned patients, consistently poor for darker skin tones due to training data imbalance. Documented across multiple dermatological AI tools.
|
||||
|
||||
**Resource allocation algorithms:** Use cost as proxy metric for health need; systematically undervalues Black patient needs because historical cost data reflects historical undertreatment (circular bias).
|
||||
|
||||
**"Dual role" conclusion:** AI can promote health equity OR exacerbate disparities depending on implementation design. Current deployment predominantly takes the disparities-widening path.
|
||||
|
||||
**Scale of bias in public programs:** ~40% of Medicare/Medicaid-facing algorithms show racial bias. 40% of Americans are on Medicare/Medicaid, making this a large-scale public health concern.
|
||||
|
||||
**Limitation:** Systematic review reflects published literature, which may be biased toward reporting AI failures for underrepresented populations as anomalies rather than systemic issues. Heterogeneous included studies limit pooled quantitative conclusions.
|
||||
|
||||
## Agent Notes
|
||||
|
||||
**Why this matters:** This is the highest-quality evidence base (129 articles, systematic methodology) for the clinical AI disparities claim. The systematic review design provides stronger evidence than the Frontiers narrative review (PMC11922879). The cost-as-proxy finding is particularly important: it's not a bug to fix but a feature of how algorithms are designed — using cost as a proxy for health need encodes historical undertreatment into future resource allocation.
|
||||
|
||||
**What surprised me:** The circularity of the cost-proxy bias: algorithms undervalue Black patient needs because historical data shows lower costs for that population — lower costs that reflected historical undertreatment, not lower genuine need. Algorithmic perpetuation of past harm through seemingly neutral efficiency optimization.
|
||||
|
||||
**What I expected but didn't find:** Evidence of systematic correction mechanisms or debiasing requirements in deployed healthcare algorithms. The review finds the problem is documented but solutions are not systematically implemented.
|
||||
|
||||
**KB connections:**
|
||||
- Converges with PMC11922879 (Frontiers review) on direction; higher methodology quality
|
||||
- Relevant to Belief 2 framing: if clinical AI compounds SDOH rather than counteracting it, the marginal value of clinical AI spending may actually be negative for equity outcomes
|
||||
- Connects to regulatory claims: FDA enforcement discretion expansion (2026) + algorithmic bias = reduced safety oversight for the most vulnerable populations
|
||||
|
||||
**Extraction hints:**
|
||||
1. Contributes to divergence file: `divergence-clinical-ai-equity-widens-vs-narrows.md` — this is the primary "widens" evidence with systematic review strength
|
||||
2. Can be cited alongside PMC11922879 as a two-source evidence base for the AI-widens-disparities direction
|
||||
3. The cost-as-proxy mechanism is extractable as a separate claim: "Healthcare resource allocation algorithms systematically undervalue Black patient needs by using historical cost data as a proxy for health need, encoding past undertreatment into future allocation decisions"
|
||||
|
||||
**Context:** Should be archived alongside PMC11922879 — the two sources are complementary (one narrative review + one systematic review) and together constitute a reasonable evidence base for an AI equity divergence file.
|
||||
|
||||
## Curator Notes
|
||||
|
||||
PRIMARY CONNECTION: Clinical AI safety claims and the new divergence candidate `divergence-clinical-ai-equity-widens-vs-narrows.md`
|
||||
|
||||
WHY ARCHIVED: Systematic review methodology (129 articles) provides highest-quality evidence base for AI-widens-disparities direction; cost-as-proxy mechanism is extractable as its own claim
|
||||
|
||||
EXTRACTION HINT: The cost-as-proxy circularity mechanism is the most important finding — it's not just bias but an architectural feature that perpetuates historical inequity
|
||||
|
|
@ -0,0 +1,67 @@
|
|||
---
|
||||
type: source
|
||||
title: "Clinical AI as currently deployed risks widening health disparities; language access tools show narrow equity benefit with deliberate design"
|
||||
author: "PMC11922879 (multiple authors, Frontiers in AI)"
|
||||
url: https://pmc.ncbi.nlm.nih.gov/articles/PMC11922879/
|
||||
date: 2025-01-01
|
||||
domain: health
|
||||
secondary_domains: [ai-alignment]
|
||||
format: peer-reviewed-review
|
||||
status: unprocessed
|
||||
priority: medium
|
||||
tags: [clinical-ai, health-equity, disparities, bias, language-access, sdoh]
|
||||
flagged_for_theseus: ["clinical AI equity concerns parallel AI safety alignment failures — bias in training data creates systematic downstream harm in clinical decisions; relevant to alignment/deployment governance"]
|
||||
---
|
||||
|
||||
## Content
|
||||
|
||||
Literature review synthesizing existing evidence on AI and healthcare disparities. Published Frontiers in AI, PMC11922879.
|
||||
|
||||
**Core finding:** AI has theoretical potential to reduce healthcare disparities but current implementation patterns risk widening them.
|
||||
|
||||
**Bias evidence:**
|
||||
- ~40% of healthcare algorithms show bias against Black patients in resource allocation (FAS policy brief convergent source)
|
||||
- Population health management algorithms allocated more care to white patients even with equivalent health needs (documented systemic bias)
|
||||
- Dermatological AI: accurate for light-skinned patients, poor for darker skin due to training data gaps (separate systematic review PMC11796235)
|
||||
|
||||
**Narrow equity benefit case:**
|
||||
- AI translation tools reduced communication errors by 43% and raised patient satisfaction 28% in non-English-speaking populations
|
||||
- BUT: this benefit only materializes with deliberate equity-centered design; not default deployment behavior
|
||||
|
||||
**Diabetes AI tools:** Found to overlook psychosocial components needed for glycemic control in underserved patients
|
||||
|
||||
**Cost inaccessibility example:** China AI glaucoma screening: $434,903 for ~2,000 patients — cost inaccessible in LMICs
|
||||
|
||||
**Public trust:** Near 50/50 split on trusting AI for diagnosis; majorities prefer human physicians for culturally sensitive decisions
|
||||
|
||||
**$320 billion** in annual excess healthcare spending attributable to health disparities (FAS policy brief)
|
||||
|
||||
**Limitation:** Literature review methodology; no original data; mostly observational studies in underlying literature.
|
||||
|
||||
## Agent Notes
|
||||
|
||||
**Why this matters:** The clinical AI disparities evidence provides a mechanism for how clinical AI adoption can compound SDOH effects rather than counteract them. If AI tools are biased toward white, higher-income populations and deployed first in resource-rich settings, they amplify the existing clinical-care access advantage — creating a meta-SDOH layer on top of existing structural inequality.
|
||||
|
||||
**What surprised me:** The specific finding that population health management algorithms allocated more care to white patients with *equivalent health needs* is more damning than general bias concerns — this isn't a training data problem, it's an ongoing systematic allocation problem in deployed systems.
|
||||
|
||||
**What I expected but didn't find:** Any RCT evidence demonstrating AI reducing SDOH-driven health gaps at population scale. The equity benefit case is entirely theoretical or narrow use-case. The disconfirmation search for AI narrowing disparities came up largely empty at the population level.
|
||||
|
||||
**KB connections:**
|
||||
- Directly relevant to Belief 2 disconfirmation search: AI was expected to be the mechanism for clinical care to overcome SDOH. Instead it may compound SDOH.
|
||||
- Connects to clinical AI deskilling literature: deskilling + disparities widening creates a dual safety concern
|
||||
- Flag to Theseus: training data bias in clinical AI is structurally analogous to alignment problems in general AI systems — systematic harm compounds without correction mechanisms
|
||||
|
||||
**Extraction hints:**
|
||||
1. New claim candidate: "Clinical AI as currently deployed tends to widen healthcare disparities due to biased training data and unequal access, with equity benefits only materializing under deliberate equity-centered design"
|
||||
2. Contributes to divergence file: `divergence-clinical-ai-equity-widens-vs-narrows.md` — this is the "widens" evidence arm
|
||||
3. Could strengthen existing clinical AI claims by adding equity dimension
|
||||
|
||||
**Context:** This is one of three converging sources on clinical AI + disparities from this session (PMC11922879, PMC11796235, FAS policy brief). The three together establish the direction; the systematic review (PMC11796235, 129 articles) is the highest-quality evidence.
|
||||
|
||||
## Curator Notes
|
||||
|
||||
PRIMARY CONNECTION: Clinical AI claims in domains/health/ (particularly AI scribe, clinical AI safety, OpenEvidence claims)
|
||||
|
||||
WHY ARCHIVED: Provides mechanism for AI-compounded SDOH effect; directly relevant to clinical AI safety claims and the access inversion compound structure
|
||||
|
||||
EXTRACTION HINT: Focus on the population health management systematic allocation bias finding — this is the most actionable evidence point; the language access benefit is narrow and context-specific
|
||||
|
|
@ -0,0 +1,67 @@
|
|||
---
|
||||
type: source
|
||||
title: "Clinical AI upskills novice physicians on discrete tasks while creating never-skilling risk for trainees: orthopaedics perspective"
|
||||
author: "PMC12955832 (multiple authors)"
|
||||
url: https://pmc.ncbi.nlm.nih.gov/articles/PMC12955832/
|
||||
date: 2026-01-01
|
||||
domain: health
|
||||
secondary_domains: [ai-alignment]
|
||||
format: peer-reviewed-perspective
|
||||
status: unprocessed
|
||||
priority: medium
|
||||
tags: [clinical-ai, deskilling, upskilling, never-skilling, trainees, orthopaedics, ai-scribes]
|
||||
flagged_for_theseus: ["novice vs. expert AI skill dynamics parallel general capability degradation concerns; never-skilling for trainees is structurally analogous to AI-dependency formation in broader human-AI systems"]
|
||||
---
|
||||
|
||||
## Content
|
||||
|
||||
Orthopaedics perspective article on deskilling vs. upskilling dynamics in clinical AI. PMC12955832.
|
||||
|
||||
**Upskilling evidence:**
|
||||
- AI-assisted imaging improved diagnostic accuracy in junior physicians (orthopaedics imaging context)
|
||||
- AI scribes reduced note time by 9.5% in one study (no significant change in another study — inconsistent result)
|
||||
- Human-AI collaboration outperforms either independently (consistent finding)
|
||||
|
||||
**Deskilling evidence:**
|
||||
- References Lancet colonoscopy trial (ADR: 28% → 22% after 3 months of AI use)
|
||||
- Springer mixed-method review (2025): evidence of deskilling "scarce but consistent across specialties"
|
||||
|
||||
**Never-skilling concept:**
|
||||
- Identified as distinct concern for trainees: failing to acquire foundational clinical skills when AI is introduced before competency develops
|
||||
- Structurally invisible: no pre-AI baseline exists for trainees, making never-skilling impossible to detect using standard deskilling metrics
|
||||
- Framed as the "worst-case" scenario for clinical AI in medical education
|
||||
|
||||
**Resolution proposed:**
|
||||
- Phased AI introduction: develop foundational competency first, then introduce AI augmentation
|
||||
- Deliberate practice periods without AI assistance
|
||||
- No validated implementation program cited
|
||||
|
||||
**Limitation:** Orthopaedics-specific perspective; diagnostic imaging may not generalize to other specialties. Upskilling benefits are for discrete imaging tasks, not complex clinical judgment.
|
||||
|
||||
## Agent Notes
|
||||
|
||||
**Why this matters:** This source provides genuine countervailing evidence for the deskilling claim — but in a way that actually REFINES rather than CONTRADICTS the concern. The key finding is that deskilling and upskilling are not mutually exclusive: AI may upskill novices on specific discrete tasks while simultaneously creating never-skilling risk by preventing the development of foundational competencies. This is a scope clarification, not a refutation.
|
||||
|
||||
**What surprised me:** The inconsistency in AI scribe efficiency data (9.5% in one study, no change in another). This suggests AI scribe value may be context-dependent — it doesn't universally reduce documentation burden.
|
||||
|
||||
**What I expected but didn't find:** Outcome data on patient outcomes for AI-assisted junior physicians vs. experienced physicians. We know AI helps junior physicians make better discrete diagnostic calls, but we don't know if that translates to better patient outcomes — the clinical significance question is open.
|
||||
|
||||
**KB connections:**
|
||||
- Lancet deskilling claim (2025-08-xx-lancet-preserving-clinical-skills-ai-deskilling.md) — this source cites and contextualizes it
|
||||
- Springer mixed-method review (2025-08-xx-springer-clinical-ai-deskilling-misskilling-neverskilling-mixed-method-review.md) — same direction
|
||||
- Never-skilling claim (mentioned in Session 23 as "ready to extract") — this adds orthopaedics-specific evidence
|
||||
|
||||
**Extraction hints:**
|
||||
1. Contributes to divergence file between deskilling and upskilling — this source adds the scope clarification: upskilling is real but for novices on discrete tasks; deskilling is real for experienced providers on complex judgment
|
||||
2. The never-skilling concern for trainees is extractable as its own claim or as an update to the existing never-skilling claim
|
||||
3. AI scribe efficiency inconsistency (9.5% vs. no change) is a supporting data point for the "context-dependent value" observation
|
||||
|
||||
**Context:** This is the "upskilling" arm of the deskilling/upskilling divergence. It should be archived alongside the Lancet and Springer deskilling sources to create the full evidence landscape for the divergence file. The key contribution is the scope clarification: the debate is not "does AI deskill or upskill?" but "which interventions, in which populations, produce which effect?"
|
||||
|
||||
## Curator Notes
|
||||
|
||||
PRIMARY CONNECTION: Clinical AI deskilling divergence candidate and never-skilling claim
|
||||
|
||||
WHY ARCHIVED: Provides genuine counterweight to deskilling evidence for divergence file construction; scope clarification (novice upskilling vs. expert deskilling vs. trainee never-skilling) enables a better-structured divergence
|
||||
|
||||
EXTRACTION HINT: Do NOT use as general "AI helps medicine" evidence — it's specifically about discrete task improvement in novices, not overall quality improvement. The never-skilling concern for trainees is the most novel finding relative to existing KB content.
|
||||
|
|
@ -0,0 +1,70 @@
|
|||
---
|
||||
type: source
|
||||
title: "GLP-1 receptor agonists reduce all-cause mortality 38% in T2D + HFrEF patients: TriNetX real-world cohort (n=26,196)"
|
||||
author: "PMC12664052 (multiple authors)"
|
||||
url: https://pmc.ncbi.nlm.nih.gov/articles/PMC12664052/
|
||||
date: 2026-01-01
|
||||
domain: health
|
||||
secondary_domains: []
|
||||
format: peer-reviewed-study
|
||||
status: unprocessed
|
||||
priority: high
|
||||
tags: [glp1, heart-failure, hfref, cardiovascular, mortality, real-world-evidence, semaglutide, trinetx]
|
||||
---
|
||||
|
||||
## Content
|
||||
|
||||
TriNetX database retrospective cohort study. Propensity-score matched comparison of GLP-1 receptor agonists vs. DPP4 inhibitors in adults with T2D + HFrEF (EF ≤40%). N=13,098 per arm after matching. Median follow-up 3.9 years.
|
||||
|
||||
**Primary outcome — all-cause mortality:**
|
||||
- HR 0.62 (95% CI 0.59–0.66, p<0.001)
|
||||
- Absolute risk reduction: 11.0% (13.3% GLP-1 arm vs. 24.3% DPP4i arm)
|
||||
- NNT = 9.1
|
||||
- 5-year survival: 72.2% vs. 61.6%
|
||||
|
||||
**Secondary outcomes (GLP-1 arm vs. DPP4i arm):**
|
||||
- All-cause hospitalization: HR 0.71
|
||||
- HF exacerbation: HR 0.83
|
||||
- MI: HR 0.87
|
||||
- CVA (stroke): HR 0.86
|
||||
- Atrial fibrillation: HR 0.92
|
||||
- Ventricular arrhythmia: HR 0.86
|
||||
|
||||
**By GLP-1 agent:**
|
||||
- Semaglutide: mortality HR 0.51 (strongest benefit)
|
||||
- Dulaglutide: HR 0.66
|
||||
- Liraglutide: HR 0.68
|
||||
|
||||
**Subgroup analysis:** Benefits consistent across age, sex, and ESRD status.
|
||||
|
||||
**Context:** Prior guidelines expressed caution about GLP-1 use in HFrEF due to concerns about fluid retention and arrhythmia risk. This real-world data shows benefit across arrhythmia endpoints (AF HR 0.92, ventricular arrhythmia HR 0.86), directly contradicting the prior caution.
|
||||
|
||||
**Limitation:** Observational design; propensity score matching cannot eliminate residual confounding. T2D-only population — mechanism in non-diabetic HFrEF unclear.
|
||||
|
||||
## Agent Notes
|
||||
|
||||
**Why this matters:** This is the largest real-world evidence base for GLP-1 use in HFrEF (historically the uncertain HF phenotype, while HFpEF had stronger trial evidence). NNT=9 over 3.9 years for all-cause mortality is a clinically striking effect size. Semaglutide's HR 0.51 is particularly notable. This should be combined with the existing HFpEF trial evidence to argue that the cardiovascular benefit of GLP-1s spans both major HF phenotypes.
|
||||
|
||||
**What surprised me:** The arrhythmia findings. Prior clinical guidance warned about potential proarrhythmic effects of GLP-1s in HFrEF. The real-world data shows the opposite — reduced AF and ventricular arrhythmia. This is a genuine update to the clinical risk model.
|
||||
|
||||
**What I expected but didn't find:** A subgroup analysis stratifying by income or race/ethnicity. This is a major gap — the population with T2D + HFrEF is disproportionately lower-income and non-white, but the RWE benefits are reported without equity stratification. We don't know if the benefit is equally distributed.
|
||||
|
||||
**KB connections:**
|
||||
- Extends existing GLP-1 HFpEF divergence candidate (2026-06-xx-pubmed-glp1-hfpef-systematic-review-meta-analysis-mortality-hospitalization.md)
|
||||
- Directly relevant to access inversion thesis: the populations being cut from Medicaid GLP-1 coverage overlap substantially with HFrEF + T2D epidemiology
|
||||
- Connects to the compound access inversion pattern documented in Sessions 20-24
|
||||
|
||||
**Extraction hints:**
|
||||
1. New claim: "GLP-1 receptor agonists reduce all-cause mortality by 38% in T2D + HFrEF patients in real-world evidence (NNT=9), overturning prior caution about arrhythmia risk in this phenotype"
|
||||
2. Potentially upgrades confidence on existing GLP-1 cardiovascular claims from experimental to likely
|
||||
3. Consider extending the HFpEF divergence file to include HFrEF evidence
|
||||
|
||||
**Context:** HFrEF (reduced ejection fraction, EF ≤40%) is the historically better-studied HF phenotype but GLP-1 benefit was uncertain there. SGLT2 inhibitors have strong HFrEF evidence (EMPEROR-Reduced, DAPA-HF). This study positions GLP-1s as competitive with SGLT2i in HFrEF, not just additive in HFpEF.
|
||||
|
||||
## Curator Notes
|
||||
|
||||
PRIMARY CONNECTION: GLP-1 HFpEF divergence file (divergence-glp1-hfpef-mortality-benefit-vs-guideline-caution) — this extends the debate to HFrEF
|
||||
|
||||
WHY ARCHIVED: Most compelling new real-world mortality evidence yet for GLP-1s in heart failure; NNT=9 directly relevant to the access inversion thesis
|
||||
|
||||
EXTRACTION HINT: Focus on (1) the NNT=9 figure for all-cause mortality, (2) the arrhythmia findings that contradict prior caution, and (3) the absence of equity stratification as a gap
|
||||
|
|
@ -0,0 +1,74 @@
|
|||
---
|
||||
type: source
|
||||
title: "GLP-1 therapy associated with 22.4% nutritional deficiency rate at 12 months in 461,382-patient retrospective cohort; dietitian consultation increases detection"
|
||||
author: "PMC12205620 (multiple authors)"
|
||||
url: https://pmc.ncbi.nlm.nih.gov/articles/PMC12205620/
|
||||
date: 2026-01-01
|
||||
domain: health
|
||||
secondary_domains: []
|
||||
format: peer-reviewed-study
|
||||
status: unprocessed
|
||||
priority: high
|
||||
tags: [glp1, micronutrient, vitamin-d, iron, nutritional-deficiency, semaglutide, monitoring, dietitian, population-health]
|
||||
---
|
||||
|
||||
## Content
|
||||
|
||||
Large retrospective observational cohort. Adults newly prescribed GLP-1 receptor agonists, July 2017–December 2021. N=461,382. Population: 56.3% female, mean age 52.9 years, 80.5% T2D.
|
||||
|
||||
**Nutritional deficiency incidence (any nutritional deficiency):**
|
||||
- 6 months: 12.7%
|
||||
- 12 months: 22.4%
|
||||
|
||||
**Vitamin D deficiency (most common):**
|
||||
- 6 months: 7.5%
|
||||
- 12 months: 13.6%
|
||||
|
||||
**Nutritional anemia:**
|
||||
- 6 months: 2.1%
|
||||
- 12 months: 4.0%
|
||||
|
||||
**Iron deficiency anemia:**
|
||||
- 6 months: 1.6%
|
||||
- 12 months: 3.2%
|
||||
|
||||
**B-vitamin deficiency:**
|
||||
- 6 months: 1.3%
|
||||
- 12 months: 2.6%
|
||||
|
||||
**Muscle loss diagnosis:**
|
||||
- 6 months: 1.5%
|
||||
- 12 months: 3.0%
|
||||
|
||||
**Key monitoring finding:**
|
||||
Dietitian consultation substantially increased deficiency detection rates: 18.5% vs. 12.2% at 6 months (patients with vs. without dietitian consultation). This suggests systematic under-detection in the 92% of patients without dietitian visits (Urbina 2026 reference).
|
||||
|
||||
**Limitation:** Observational; incident diagnoses reflect both true new cases and previously existing but newly detected deficiencies (detection bias). Causality not definitively established. Population is 80.5% T2D, which itself carries nutritional risk — GLP-1-specific contribution unclear without robust control comparison.
|
||||
|
||||
## Agent Notes
|
||||
|
||||
**Why this matters:** This is the primary source study for the large-scale micronutrient deficiency signal that was partially captured in the archive through IAPAM summaries. It adds the critical dietitian consultation finding — the deficiencies are systematically under-detected without nutritional infrastructure. Combined with Urbina's finding that 92% have no dietitian visit, the scope of systematic under-detection at the population level is significant.
|
||||
|
||||
**What surprised me:** The muscle loss diagnosis rate (3.0% at 12 months) is notably absent from most public discourse around GLP-1 safety monitoring. The framing around GLP-1 safety has focused on nausea/GI effects — the long-term nutritional and musculoskeletal trajectory is underemphasized.
|
||||
|
||||
**What I expected but didn't find:** Income or food-security stratification. The study population is likely predominantly commercially insured given the 2017-2021 timeframe (Medicaid coverage was minimal). The food-insecure double-jeopardy (already nutrient-depleted + GLP-1-induced appetite suppression) remains a genuine research gap.
|
||||
|
||||
**KB connections:**
|
||||
- Urbina 2026 narrative review (2026-01-xx-urbina-clinical-obesity-glp1-micronutrient-narrative-review.md) — convergent evidence, adds dietitian finding
|
||||
- Frontiers Nutrition cross-sectional n=69 (2025-03-xx-frontiers-nutrition-glp1-nutrient-intake-crosssectional.md) — small sample, same direction
|
||||
- Access inversion thesis: the monitoring gap is likely worse in lower-income populations
|
||||
|
||||
**Extraction hints:**
|
||||
1. Update existing micronutrient claim with the 22.4% at 12 months figure (from primary source PMC12205620, not IAPAM summary)
|
||||
2. New claim candidate: "Dietitian consultation substantially increases detection of GLP-1-associated nutritional deficiencies (18.5% vs 12.2% at 6 months), but 92% of GLP-1 users receive no dietitian consultation" — frames the infrastructure gap
|
||||
3. Confidence upgrade: moves micronutrient deficiency claim from experimental toward likely given n=461,382
|
||||
|
||||
**Context:** PMC12205620 is the primary source for data previously referenced in IAPAM summary archives (2026-04-08-glp1-nutritional-deficiency-signal.md) and partially referenced in Urbina narrative review archive. This archive provides the primary citation with full study design details and the monitoring finding.
|
||||
|
||||
## Curator Notes
|
||||
|
||||
PRIMARY CONNECTION: Existing micronutrient deficiency claims in domains/health/ and 2026-01-xx-urbina-clinical-obesity-glp1-micronutrient-narrative-review.md
|
||||
|
||||
WHY ARCHIVED: Primary source citation gap in existing archive; adds dietitian consultation detection finding not present in proxy sources
|
||||
|
||||
EXTRACTION HINT: Focus on the dietitian consultation detection difference (18.5% vs 12.2%) as the new infrastructure claim — this is the finding missing from prior archives
|
||||
|
|
@ -0,0 +1,55 @@
|
|||
---
|
||||
type: source
|
||||
title: "Early GLP-1 discontinuation associated with higher risk of coronary artery disease and heart failure: Truveta EHR post-market surveillance"
|
||||
author: "Truveta Research"
|
||||
url: https://www.truveta.com/blog/research/glp1-discontinuation-cardiovascular-outcomes/
|
||||
date: 2026-01-01
|
||||
domain: health
|
||||
secondary_domains: []
|
||||
format: research-blog-rwe
|
||||
status: unprocessed
|
||||
priority: medium
|
||||
tags: [glp1, discontinuation, adherence, cardiovascular, heart-failure, cad, semaglutide, tirzepatide, real-world-evidence]
|
||||
---
|
||||
|
||||
## Content
|
||||
|
||||
Truveta EHR-based post-market surveillance study. 1-year follow-up. Covers both semaglutide and tirzepatide.
|
||||
|
||||
**Primary finding:** Early GLP-1 discontinuation linked to significantly higher risk of:
|
||||
- Coronary artery disease
|
||||
- Heart failure
|
||||
|
||||
Applies to both semaglutide and tirzepatide. Presented as real-world EHR evidence.
|
||||
|
||||
**Context:** Presented alongside ESC Congress 2025 retrospective cohort on early vs. long-term GLP-1 treatment, finding consistent with this result (higher discontinuation associated with worse CV outcomes).
|
||||
|
||||
**Limitation:** Blog-format publication from data company (Truveta) — not peer-reviewed. Specific HRs not available in search results. EHR data subject to confounding (patients who discontinue may have worse baseline health, not just worse outcomes due to discontinuation). Need full study for methodological assessment.
|
||||
|
||||
## Agent Notes
|
||||
|
||||
**Why this matters:** Extends the continuous-treatment-required thesis from weight outcomes to cardiovascular outcomes. Prior evidence showed weight rebounds after GLP-1 discontinuation. This suggests the CV benefit mechanism also requires continued exposure — not just the weight-loss endpoint. This strengthens the argument that GLP-1 access interruptions (Medicaid coverage cuts, insurance PA barriers) have direct cardiovascular harm, not just weight management harm.
|
||||
|
||||
**What surprised me:** The heart failure finding specifically. If GLP-1 discontinuation is associated with incident heart failure events (not just HF exacerbation in existing HF patients), that's a new mechanism dimension. Previously, the access inversion argument was: people with heart failure can't access the drug. This adds: access interruption causes heart failure.
|
||||
|
||||
**What I expected but didn't find:** Specific hazard ratios or absolute risk differences. The blog format is the limitation here — methodological details are unavailable.
|
||||
|
||||
**KB connections:**
|
||||
- Continuous-treatment-required claim (hold for scope revision per Session 23)
|
||||
- Access inversion thesis: coverage cuts → discontinuation → CV harm
|
||||
- GLP-1 HFrEF evidence (PMC12664052): the CV benefit requires persistence to materialize
|
||||
|
||||
**Extraction hints:**
|
||||
1. Once peer-reviewed publication is available, use to strengthen continuous-treatment-required claim with CV outcome extension
|
||||
2. For now, flag as supporting signal — insufficient for standalone claim (no specific HRs, not peer-reviewed)
|
||||
3. The ESC 2025 convergent finding adds credibility — same direction from independent source
|
||||
|
||||
**Context:** Truveta is a credible healthcare data analytics company with legitimate EHR access. Blog publications are pre-peer-review signals, not final evidence. Archive for now; flag for follow-up when peer-reviewed version appears.
|
||||
|
||||
## Curator Notes
|
||||
|
||||
PRIMARY CONNECTION: Continuous-treatment-required claim and access inversion thesis
|
||||
|
||||
WHY ARCHIVED: Extends continuous-treatment benefit to CV outcomes, not just weight; strengthens access-cuts-cause-CV-harm argument; too preliminary for extraction but important signal
|
||||
|
||||
EXTRACTION HINT: Do NOT extract as standalone claim yet — flag for follow-up when peer-reviewed version available. Use as supporting signal for existing claims only.
|
||||
Loading…
Reference in a new issue