# Research Log — alzhub.org data harvest

**Date of research:** 2026-09-21
**Method:** Five parallel research tracks (dispatched as subagents), each assigned a topic
slice, instructed to use WebSearch to find leads and WebFetch (with `r.jina.ai` as a fallback
reader proxy for blocked/paywalled pages) to actually open and read every source before
extracting any fact, number, author, DOI, or PMID. Search-engine snippets were treated as
leads only, never as evidence, per the project's honesty rules.

## What was searched, track by track

### Track 1 — Biomarkers & Diagnosis
Searched PubMed/NIH for blood biomarker papers (p-tau217, p-tau181, GFAP, NfL) 2023-2026;
NIA newsroom; JAMA Neurology, Lancet Neurology, Alzheimer's & Dementia, Brain, Annals of
Neurology, NEJM; the 2024 revised Alzheimer's Association diagnostic criteria and the
clinical-community pushback against them; C2N PrecivityAD2, Fujirebio Lumipulse, and
Quest/ALZpath assay validation papers.
**Yield:** 12 verified studies (9 biomarkers, 3 diagnosis).
**Not found / excluded:** The NIA newsroom page on blood-test accuracy
(`nia.nih.gov/news/accurate-blood-test-alzheimers-disease`) was CAPTCHA-blocked on both direct
fetch and the reader proxy; the underlying study it covered was independently verified via
PubMed instead, so the NIA page itself is not cited. A 2024 AAIC conference abstract by the
Pitt group (Sehrawat et al., Alzheimer's Dement. 2024;20:e092598) was excluded in favor of
that group's later full peer-reviewed 2025 journal paper, since conference abstracts are not
full peer review. Several news/marketing write-ups appearing only as search snippets (CNN,
NeurologyLive, Xtalks, PRNewswire) were not opened and are not used as sources for any claim.

### Track 2 — Treatment & Clinical Trials
Searched FDA press announcements for aducanumab, lecanemab, and donanemab approvals and the
Aduhelm discontinuation; the primary pivotal trial papers (Clarity AD/NEJM, TRAILBLAZER-ALZ
2/JAMA, EMERGE-ENGAGE/JPAD); EMA CHMP decisions and reversals for lecanemab and donanemab; the
CMS national coverage determination; ClinicalTrials.gov for GLP-1 and tau-antibody trials; and
ARIA safety literature.
**Yield:** 15 verified studies/regulatory items (9 treatment, 5 clinical-trial, plus the ARIA
safety review folded into treatment).
**Not found / excluded:** The original FDA.gov press-release URLs for the 2021 aducanumab and
January 2023 lecanemab accelerated approvals both returned 404 (FDA appears to have
archived/removed them); the corresponding Biogen investor-relations releases, which quote the
FDA's approval language directly, were substituted and independently verified instead. A CMS
PDF (MM13598, registry data-submission specifics) returned unreadable garbled binary and was
not used. The raw ClinicalTrials.gov HTML page for NCT04777396 rendered only glossary
boilerplate (a site rendering issue); the official ClinicalTrials.gov API v2 JSON endpoint was
used instead and returned the full record.

### Track 3 — Prevention, Genetics & Basic Science
Searched for the US POINTER and FINGER multidomain lifestyle trials; the ACHIEVE hearing-aid
trial; the 2024 Lancet Commission report; SPRINT MIND blood-pressure/cognition follow-up;
cataract-surgery/vision-loss meta-analysis; sleep-duration cohort studies; GLP-1/semaglutide
observational and RCT data; APOE4/APOE2 and TREM2 genetics papers; and basic-science
mechanism papers on glymphatic clearance and microglia.
**Yield:** 16 verified studies (8 prevention, 3 genetics, 4 basic-science, minus 1 duplicate
merged — see below).
**Not found / fully verified:** Nedergaard & Goldman's 2020 *Science* review, "Glymphatic
failure as a final common pathway to dementia," was paywalled on Science.org and PubMed even
through the reader proxy; it is included (as a review/perspective, evidence_strength "early")
on the strength of PubMed metadata and independent corroborating coverage, not a direct read
of the full text — this is flagged explicitly in `studies.json`. Similarly, the exact
percentage breakdown in the Lancet Commission 2024 results table (e.g., the ~7% LDL and ~2%
vision-loss attributable fractions) was corroborated by multiple secondary summaries
(EurekAlert, Alzheimer's Disease International, Alzheimer Europe) rather than a direct read of
the paywalled primary results table; the risk-factor list and overall 45% headline figure
were confirmed via a directly-fetched hosted PDF mirror of the report.
**Duplicate found:** Track 3 and Track 2 independently surfaced the same Lancet 2026 paper
(oral semaglutide EVOKE/EVOKE+ trials). The two entries were merged into one
(`evoke-evoke-plus-semaglutide-lancet-2026`, categorized clinical-trial) during compilation to
avoid double-counting; the companion 2024 observational semaglutide study
(`semaglutide-real-world-alzheimers-2024`) was kept as a separate, distinct study since it is
a different paper.

### Track 4 — Caregiving Studies & Statistics
Searched for landmark caregiver-intervention trials (REACH II, Mittelman/NYU spousal
counseling), caregiver mortality/burden research (Schulz & Beach, Pinquart & Sörensen,
Adelman), a Cochrane review of respite care, and recent CMS GUIDE model implementation papers.
For statistics: the Alzheimer's Association 2025 Facts and Figures report (both the
peer-reviewed Alzheimer's & Dementia article and the executive-summary PDF), CDC/NCHS
mortality data briefs, the WHO dementia fact sheet, NIH RePORT categorical spending, the NIA
FY2025 budget justification, and the Lancet Commission's headline prevention statistic.
**Yield:** 9 verified caregiving studies; 43 verified statistics (exceeding the 30-stat
target).
**Not found / excluded:** The `alzheimers.gov` NIH research-progress-report page was blocked
by a CAPTCHA/human-verification wall on both direct fetch and proxy and was not used; the NIH
RePORT categorical spending table (fetched successfully via proxy) was used instead for
funding figures. The live NIA FY2025 budget page was also bot-gated directly; a Wayback
Machine snapshot was used as a substitute and is cited as such. Two live CDC data-portal pages
(`cdc.gov/aging/agingdata/`, `cdc.gov/healthy-aging-data/`) opened fine but contained no
extractable named statistics — only a data-portal landing page — and were excluded.
**Version-drift note:** the live `alz.org/alzheimers-dementia/facts-figures` page has already
rolled over to 2026-cycle figures (7.4 million prevalence, $409B cost) as of this September
2026 research date. Figures in `stats.json` are deliberately pulled from the dated 2025-cycle
report text (the peer-reviewed article and executive-summary PDF) rather than the live page,
so every stat is traceable to a specific report year; the one 2026-cycle figure that is
included is explicitly labeled as such and dated 2026.

### Track 5 — News (last ~12-18 months)
Searched AP, Reuters, STAT News, NPR, BBC Health, and the newsrooms of FDA, EMA, Alzheimer's
Association, and NIA/NIH for drug approvals, blood-biomarker test launches, trial
results/failures, funding news, and caregiving/policy news roughly September 2024 – September
2026.
**Yield:** 25 verified news items across research (8), funding (5), drug-approval (4),
diagnosis (3), caregiving (3), and policy (2).
**Not found / excluded:** A CNN piece on NIH funding cuts was blocked (HTTP 451) with the
proxy hitting a Cloudflare challenge page; Scientific American's independently-verified
equivalent story was used instead. An NPR piece on state caregiver-leave policy timed out
twice and hit a Cloudflare challenge via proxy; excluded, leaving the dataset without a
state-level caregiving-policy item (a genuine gap — see below). An NIA FY2027
professional-judgment budget page returned HTTP 405. An alzimpact.org Senate companion-bill
page returned an empty body, and the congress.gov bill page returned HTTP 403 — as a result,
the Senate companion to the House ASAP Act is not separately verified/cited (only the House
bill is). A NeurologyLive article on remternetug self-injection returned HTTP 403 and was
excluded — no remternetug item appears in the dataset. A CMS fact sheet PDF on GUIDE respite
dollar caps (MLN7172818) returned corrupted binary and was not used, so exact PY2026 respite
reimbursement caps are not stated anywhere in this dataset.
**Outlet mix achieved:** FDA (2), Alzheimer's Association/AAIC (9), STAT News (4), Scientific
American (1), Clinical Trials Arena (1), Roche (1), Quest Diagnostics (1), PharmExec (1),
BioSpace/C2N (1), UCLA (1), Alzheimer's Disease International (1), Association of Health Care
Journalists (1), peer-reviewed report (1).

## Deliverable counts (final, after de-duplication)

| File | Count | Breakdown |
|---|---|---|
| `studies.json` | 49 | biomarkers 9, treatment 9, prevention 8, caregiving 8, clinical-trial 5, basic-science 4, diagnosis 3, genetics 3 |
| `news.json` | 25 | research 8, funding 5, drug-approval 4, diagnosis 3, caregiving 3, policy 2 |
| `stats.json` | 45 | prevalence/demographics, mortality, cost/economic burden, caregiver burden, workforce, global (WHO), prevention (Lancet Commission), federal funding, program enrollment |

**Date range covered:** studies span 1999-12-15 (foundational caregiving research) through
2026-04-21 (most recent ARIA safety workgroup update); news spans 2024-11-25 through
2026-08-20.

## Honesty-driven exclusions (items considered but NOT included)

1. **Nedergaard & Goldman 2020 Science glymphatic review** — included but flagged, since full
   text could not be directly opened even via proxy (see Track 3 notes above).
2. **Sehrawat et al. 2024 AAIC conference abstract** — excluded in favor of the full 2025
   peer-reviewed paper by the same authors (conference abstracts are not full peer review).
3. **Senate companion to the ASAP Act** — not included as a separate item because the
   congress.gov bill page 403'd and the alzimpact.org summary page returned empty; only the
   House bill (verified via the Alzheimer's Association newsroom) appears in `news.json`.
4. **Remternetug self-injection news** — excluded; source article was blocked (403) and no
   alternate verified source was found in the time available.
5. **State-level caregiver-leave policy news** — a planned item (NPR) could not be verified due
   to a Cloudflare block; this is a genuine gap in the `policy`/`caregiving` news coverage that
   a future pass should fill from an alternate outlet.
6. **GUIDE model PY2026 respite dollar caps** — a CMS fact-sheet PDF meant to source this could
   not be parsed (corrupted binary); no specific respite reimbursement cap figure is stated
   anywhere in this dataset as a result.

No study, statistic, author, DOI, PMID, or URL in the three data files was invented; every
entry traces to a source actually opened and read, with the two explicitly-flagged partial
exceptions above (glymphatic review, Lancet Commission sub-percentages) documented rather than
silently included.

## Genuinely thin areas (reported rather than papered over)

- **State/local caregiving policy** is underrepresented in `news.json` (only federal
  ASAP Act coverage) due to the blocked NPR source noted above.
- **TREM2/microglia basic science** is represented by 2 papers (2020, 2024) rather than a
  larger set; the field moves slower in fully peer-reviewed human-tissue studies than the
  biomarker space, so this reflects real publication volume rather than a search gap.
- **Genetics** category has only 3 studies (APOE4, APOE2, TREM2 splicing) — this is a
  narrower slice than biomarkers/treatment by design (the research brief asked for broad
  coverage across many risk-factor domains, and genetics-specific breakthrough papers in the
  12-18 month news window were sparser than biomarker or drug news).

## 2026-09-28 — Individual-level MCI→Alzheimer's progression dataset survey

Separate task from the above (which targets aggregate/ecological sources); this one hunted
individual-level, longitudinal cohort data for modelling MCI-to-AD conversion. Full writeup:
`data/PROGRESSION_DATASETS_20260928.md`.

**Verified/opened this session:** OASIS-1/2 tabular derivative on Kaggle (genuinely OPEN, no
login — downloaded and measured: 436 OASIS-1 rows, 373 OASIS-2 rows/150 subjects/14 converters,
but baseline is CDR 0 normal, not MCI, so it does not answer the actual research question);
LASI-DAD (`lasi-dad.org`, FORM/MANAGED, free application); EPAD LCS (`ep-ad.org`, MANAGED via AD
Workbench despite "Open Source Data" branding — its Data Information Pack PDF is openly
downloaded, 423KB); a restricted Zenodo deposit (normal→MCI conversion, wrong direction for this
question, `access_right: restricted`, no files attached). ADAMS (HRS supplement) design was
verified via its NCBI-hosted abstract only — the entire `hrsdata.isr.umich.edu` domain
CAPTCHA-blocked every fetch attempt, including through the `r.jina.ai` proxy that works
elsewhere on this project. BIOCARD (Johns Hopkins) 403'd on every attempt. PhysioNet, Hugging
Face, and ICPSR were checked and yielded nothing usable (no PhysioNet AD/MCI dataset found;
Hugging Face only has cross-sectional OASIS-derived image sets; ICPSR HRS page 403'd).

**Headline honesty note:** no dataset found this session is both openly downloadable and has a
genuine diagnosed-MCI baseline with a tracked conversion outcome. The only open, no-login
individual-level longitudinal download tracks normal→mild-decline (CDR 0→0.5), not MCI→AD.
Everything with the right design (LASI-DAD, EPAD, BIOCARD, plus the already-catalogued
ADNI/NACC/AIBL/BioFINDER/OASIS-3-4 from the 2026-09-28 gated-access-register pass) is behind a
managed application. Zenodo's direct API also started 403ing this session ("unusual traffic from
your network"), likely from this project's earlier heavy Zenodo use — broader Zenodo searching
for other MCI-conversion datasets was cut short as a result and is a genuine, stated gap.

## 2026-09-28 — Data research: unblocking the correlation gaps

Consolidated and verified ~35 source pulls already sitting uncommitted in `data/sources/`
from same-day earlier sessions (timestamps 2026-09-27 23:14 through 2026-09-28 03:20) that had
never been written up in the deliverable format this mission requires. Cross-referenced every
claimed row count against the actual file (`wc -l` / `csv.DictReader`, not the README's own
claim) — all matched within 1 (header row). Built the gap list from `correlation_studies.json`
(1 hard-blocked card: VAX-3, COVID-vaccine/dementia; 3 thin-coverage country cards: I4/J1/J4)
and from `analysis/run_analysis.py`'s `CANNOT_ANSWER` list, then checked each claimed "NOT
SOURCED" column's real fill rate in `panel_county.csv`/`panel_state.csv` directly.

**New work this session:** built `data/sources/cms_nursing_home_provider_info/county_certified_beds_2026.csv`
(2,826/3,143 counties matched, 89.9%, 1,542,523 total certified beds) from the already-downloaded
CMS Care Compare file — the first real data for the `nursing_home_beds_per_1000` gap. Confirmed
live today: EPA's SDWIS Envirofacts API (keyless, 2,098,792 violation records nationwide,
`data.epa.gov/efservice/VIOLATION` and `/WATER_SYSTEM`) is real and working but was not turned
into `pct_water_system_violations` — the violation-type/time-window/system-type choice needs
review before publishing a number. Re-confirmed live: CDC WONDER still structurally blocks
county-level mortality grouping (real HTTP 500s, not fabricated); no individual-level study
exists yet measuring dementia risk by COVID-19 vaccination status (checked PubMed today).
CDC's county lead-exposure page exists but resisted 3 automated extraction attempts — flagged,
not solved.

Full detail: `data/DATA_RESEARCH_20260928.md`.