All findings (11)
- About 3 in 10 people with mild cognitive impairment develop dementia within 3 years
- Memory and thinking tests lift dementia-risk prediction far more than scans or spinal-fluid markers
- Cognitive tests alone predicted dementia better than all biomarkers combined
- Two spinal-fluid tau measures move almost in lockstep (correlation 0.98)
- ADAS-Cog 13 and ADAS-Cog 11 correlate at 0.95
- A blood p-tau217 to amyloid-42 ratio tracks amyloid on PET scans (correlation 0.75)
- Executive-function tests track amyloid, but mostly through memory
- Genetic liability to ADHD did not predict memory or thinking scores or their decline in older adults
- No ADHD-gene DNA-methylation difference by cognitive status beyond a cohort-wide shift
- In a small genetic scan the only genome-wide hit for Alzheimer's was the APOE gene region
- Mendelian randomisation finds no robust causal link from ADHD liability to Alzheimer's disease
Publishable
About 3 in 10 people with mild cognitive impairment develop dementia within 3 years
We followed 988 people who started with mild cognitive impairment and no dementia, and counted how many were later diagnosed with dementia. About 8 in 100 had been diagnosed by one year, 21 in 100 by two years, 30 in 100 by three years and 38 in 100 by four years. Half had been diagnosed by roughly six and a half years. These are averages across a large group and say little about any one person.
| Measure | Value | n |
|---|---|---|
| Diagnosed with dementia by 12 months | 8.2% (95% CI 6.6-10.1) | 988 |
| by 24 months | 20.9% (18.4-23.7) | 988 |
| by 36 months | 30.2% (27.2-33.4) | 988 |
| by 48 months | 37.7% (34.3-41.3) | 988 |
Caveat. This is a research cohort of volunteers seen at memory clinics, so rates are higher than in the general population, and allowing for the timing of clinic visits moves the one-year figure up to about 13%.
Source report: /home/arodrig5/adni_pipeline/out/summary.md
Publishable
Memory and thinking tests lift dementia-risk prediction far more than scans or spinal-fluid markers
We built a risk model for who with mild impairment would develop dementia and tested it on 297 people it had never seen. Using age, sex, education and APOE4 alone, the model ranked people correctly about 64 times in 100 (0.635, where 0.5 is a coin toss). Adding memory and thinking tests raised this to about 85 in 100 (0.847). Adding spinal-fluid markers, brain MRI and amyloid/glucose PET on top moved it by less than one point each (0.855, 0.860, 0.863), and every one of those gains is within chance.
| Measure | Value | n |
|---|---|---|
| Age/sex/education/APOE4 only (concordance) | 0.635 (0.577-0.686) | 297 |
| + cognitive tests | 0.847 (0.812-0.877) | 297 |
| + spinal fluid | 0.855 (0.822-0.882) | 297 |
| + MRI | 0.860 (0.829-0.888) | 297 |
| + amyloid PET and glucose PET | 0.863 (0.832-0.890) | 297 |
Caveat. Several of the cognitive tests are used by clinicians to decide the dementia diagnosis itself, which flatters their predictive power.
Source report: /home/arodrig5/adni_pipeline/out/cognitive.md
Not ready to publish
Cognitive tests alone predicted dementia better than all biomarkers combined
A model using only memory and thinking tests and one using only biomarkers (spinal fluid, MRI, PET, plus age and APOE4) were compared on the same 297 held-out people. The test-based model ranked people better (0.848 versus 0.806), and tests added 0.057 (0.036 to 0.080) on top of the biomarkers.
| Measure | Value | n |
|---|---|---|
| Biomarkers without cognition (concordance) | 0.806 (0.767-0.843) | 297 |
| Cognitive tests, extra gain beyond biomarkers | +0.057 (+0.036 to +0.080) | 297 |
Caveat. Not for stand-alone use: the tests help define the diagnosis being predicted, and most biomarker values were estimated by imputation for people who lacked them, which handicaps the biomarker side.
Source report: /home/arodrig5/adni_pipeline/out/cognitive.md
Publishable
Two spinal-fluid tau measures move almost in lockstep (correlation 0.98)
In 693 people with spinal-fluid results, total tau and phosphorylated tau (p-tau181) were almost perfectly correlated (rank correlation 0.98). A model or report that includes both is adding very little new information over using one.
| Measure | Value | n |
|---|---|---|
| Rank correlation, total tau vs p-tau181 (log scale) | 0.98 (0.98-0.99) | 693 |
Caveat. Phosphorylated tau is a fraction of total tau, so a high correlation is partly built in; it does not mean the two carry identical clinical meaning.
Source report: /home/arodrig5/adni_pipeline/out/correlations.md
Not ready to publish
ADAS-Cog 13 and ADAS-Cog 11 correlate at 0.95
The 13-item and 11-item versions of the same cognitive scale agree very closely (0.95 in 983 people).
| Measure | Value | n |
|---|---|---|
| Rank correlation, ADAS-Cog 13 vs 11 | 0.95 (0.94-0.95) | 983 |
Caveat. Not worth publishing: the 13-item test contains the 11 items of the other, so the correlation is true by construction.
Source report: /home/arodrig5/adni_pipeline/out/correlations.md
Publishable
A blood p-tau217 to amyloid-42 ratio tracks amyloid on PET scans (correlation 0.75)
In 113 people who had both a blood test and an amyloid PET scan, the blood ratio of p-tau217 to amyloid-42 rose and fell with the amount of amyloid on the scan (rank correlation 0.75, 95% confidence range 0.66 to 0.81). That is a strong agreement but far from a one-to-one match.
| Measure | Value | n |
|---|---|---|
| Rank correlation, blood p-tau217/amyloid-42 ratio vs amyloid PET | 0.75 (0.66-0.81) | 113 |
| Rank correlation, blood amyloid-42/40 ratio vs amyloid PET (inverse) | -0.68 (-0.75 to -0.59) | 114 |
| Benchmark: spinal-fluid amyloid-42/p-tau181 ratio vs amyloid PET (inverse) | -0.81 (-0.83 to -0.78) | 505 |
Caveat. Only 113 people, one blood-test platform, and people with mild impairment, so it should not be read as how well the test performs for diagnosis in general.
Source report: /home/arodrig5/adni_pipeline/out/correlations.md
Not ready to publish
Executive-function tests track amyloid, but mostly through memory
Tests of planning and processing speed were related to amyloid on PET (rank correlation -0.27, n=497). After accounting for memory the link shrank to -0.11 (-0.19 to -0.02) and did not clear the multiple-testing threshold; the link to the spinal-fluid amyloid marker was stronger (0.33, n=693).
| Measure | Value | n |
|---|---|---|
| Executive-function composite vs amyloid PET, raw | -0.27 (-0.36 to -0.19) | 497 |
| Same, after adjusting for memory | -0.11 (-0.19 to -0.02), corrected p=0.113 | 497 |
| Executive-function composite vs spinal-fluid amyloid-42 | +0.33 (+0.26 to +0.40) | 693 |
Caveat. Not for stand-alone use: the result flips between significant and not depending on the marker and the correction, the composites share test components with the memory score, and it says nothing about ADHD itself.
Source report: /home/arodrig5/adni_pipeline/out/correlations.md
Publishable
Genetic liability to ADHD did not predict memory or thinking scores or their decline in older adults
We scored 751 older adults of European ancestry on a polygenic score for ADHD, leaving out the APOE region, and tested it against ten measures of cognition at the start and their change over time, including whether the link differed by amyloid status. Nothing was convincing: the smallest corrected p-value was 0.79. In 391 further people genotyped on a different platform it was again null (smallest corrected p 0.96), and pooling all 1,142 people gave 0.93.
| Measure | Value | n |
|---|---|---|
| Smallest corrected p-value, sequenced sample | 0.793 | 751 |
| Smallest corrected p-value, independent array sample | 0.958 | 391 |
| Smallest corrected p-value, pooled | 0.925 | 1142 |
| Score agreement between two genotyping methods | r = 0.969 | 542 |
Caveat. ADHD is not recorded in this cohort, the score captures only a few percent of ADHD liability, European ancestry only, and the second sample is mainly early-enrolled people with limited amyloid data, so the test of an amyloid interaction is weak.
Source report: /mnt/adni/prs_adhd/report.md; /home/arodrig5/adni_pipeline/out/imputed_prs_replication.md
Not ready to publish
No ADHD-gene DNA-methylation difference by cognitive status beyond a cohort-wide shift
At 1,014 sites in ADHD-related genes, none differed between cognitively normal people and people with mild impairment or Alzheimer's after correction; 4,998 random sites gave the same answer (none).
| Measure | Value | n |
|---|---|---|
| Panel sites passing correction, normal vs mild impairment | 0 of 1014 | 603 |
| Random control sites passing correction | 0 of 4998 | 603 |
| Normal vs Alzheimer's: panel sites passing correction | 0 of 1014 | 252 |
| Test inflation, panel vs random (normal vs impaired) | 1.53 vs 1.25 | 603 |
Caveat. Not for stand-alone use: only 36 Alzheimer's participants, and a cohort-wide same-direction methylation shift (inflation up to 2.35) means small effects could not be detected, so this is a weak null rather than a clean one.
Source report: /mnt/adni/methylation_analysis/report.md
Publishable
In a small genetic scan the only genome-wide hit for Alzheimer's was the APOE gene region
We compared the genomes of roughly 320 cognitively normal people and people with Alzheimer's dementia. One variant in the APOE region, the best-known Alzheimer's risk gene, reached genome-wide significance (odds ratio 5.4, p = 6.4e-9); no other region did. Comparing normal with mild impairment in about 760 people found nothing significant, as expected for a sample this small. That the known gene shows up and the statistics are well behaved (inflation factor 0.97) is a useful check on the data.
| Measure | Value | n |
|---|---|---|
| Odds ratio, APOE-region variant, normal vs dementia | 5.4 | 321 |
| p-value | < 0.001 | 321 |
| Genome-wide-significant regions, normal vs dementia | 1 (APOE) | 321 |
| Genome-wide-significant regions, normal vs mild impairment | 0 | 761 |
Caveat. Only a few hundred people, so the odds ratio is inflated by comparing two extreme groups and this scan could not find new genes.
Source report: /mnt/adni/gwas_analysis/top20_CNvAD_m1_agesex_pc10.csv; /mnt/adni/gwas_analysis/lambda_gc.csv
Not ready to publish
Mendelian randomisation finds no robust causal link from ADHD liability to Alzheimer's disease
Using genetic variants as natural experiments, we tested whether liability to ADHD changes Alzheimer's risk in three large biobank outcomes. The main estimate for late-onset Alzheimer's was an odds ratio of 0.98 (0.90 to 1.08) with every robustness method agreeing. One larger study gave a nominal 5% lower risk (0.945, p = 0.025) that was not supported by other methods and weakened when variants were selected more strictly (p = 0.074).
| Measure | Value | n |
|---|---|---|
| IVW odds ratio, late-onset Alzheimer's (FinnGen) | 0.984 (0.898-1.078) | 34 |
| Weighted median odds ratio | 1.025 (0.899-1.167) | 34 |
| Pleiotropy-outlier test (global) | p = 0.856, no outliers | 34 |
| Larger Alzheimer's study, IVW | 0.945 (0.900-0.993), p = 0.0250 | 34 |
Caveat. Not for stand-alone use yet: the earlier 27-variant instrument set could not be reproduced (34 used), the MR-RAPS package was unavailable so a home-built robust estimator was used, and one nominal signal is unresolved.
Source report: /mnt/adni/gwas_analysis/mr_robust.md