How to read an Alzheimer's research paper without a science degree

A research paper is an argument with numbers attached. The abstract is the argument's best version of itself. The six questions below are a way of reading past the abstract to what the study can actually support.

1. Who was in it?

Every result applies first to the people who were studied, and only by inference to anyone else. So the first thing to find is the participants: how many, how old, how ill, how they were recruited, and who was excluded.

Take the lecanemab trial in The New England Journal of Medicine. It enrolled 1,795 people aged 50 to 90 with early Alzheimer's, meaning mild cognitive impairment or mild dementia, and every one of them had amyloid confirmed by a PET scan or spinal-fluid test. That is a precise population. The result says nothing about people with moderate dementia, nothing about people without confirmed amyloid, and nothing about the people the trial screened out. A later paper in Annals of Neurology looked at exactly that: across four early-Alzheimer's trials, 10,804 people were screened in the US, 65% of them non-Hispanic white, and Black, Hispanic and Asian participants were two to four times more likely to be ruled ineligible on the amyloid criteria. A trial population is a selection, and the selection shapes what can be concluded.

Size matters, but less than people think. A small study can be decisive if the effect is large and the design is clean. A huge study can be worthless if everyone in it was chosen in a way that produces the answer. Our sample-size analysis shows how the library's studies are distributed.

2. What was compared with what?

This is the study design, and it decides what kind of claim is possible. The rough hierarchy, from the claim "this causes that" down to "these go together":

• Randomised controlled trial. People are assigned to treatment or control by chance, so the two groups differ only in the treatment. The only design that can show a treatment caused an outcome.
• Meta-analysis or systematic review. Pools the published studies on a question. Only as good as the studies it pools, and sensitive to which ones were found and included.
• Cohort study. Follows a group over time and sees who develops the outcome. Can show that an exposure comes before an outcome, but not that it caused it, because people who differ in the exposure usually differ in other things too.
• Case-control study. Starts with people who have the outcome and looks back for differences. Faster and cheaper, and more vulnerable to bias in how cases and controls were chosen.
• Cross-sectional study. A snapshot. Can show two things go together now; cannot say which came first.
• Modelling. A computer simulation built on assumptions drawn from other studies. Its result is a consequence of the assumptions.

The meta-analysis of risk factors for MCI progression in the Journal of Neurology, Neurosurgery and Psychiatry is a good example of the second rung. It pooled 60 cohort studies with 14,821 participants and found relative risks for progression of 3.77 for an abnormal spinal-fluid tau-to-amyloid ratio and 2.59 for hippocampal atrophy. Those are associations across many observational cohorts. They are strong evidence that the markers predict progression and no evidence at all that changing them would prevent it.

3. What exactly was measured?

Find the primary outcome: the one measure the study was designed and sized to test. Then find out what it is in human terms.

The lecanemab trial's primary outcome was the change at 18 months in the Clinical Dementia Rating Sum of Boxes, a scale from 0 to 18 where higher means worse, built from a structured interview with the patient and a family member about memory, orientation, judgement, community affairs, home life and personal care. The donanemab trial in JAMA used a different composite, the integrated Alzheimer's Disease Rating Scale. Two trials of similar drugs with different primary outcomes cannot be compared number for number, and neither scale is something a patient or family would notice directly.

Watch for surrogate outcomes. A reduction in amyloid on a PET scan is a biological event; whether it translates into a person remembering more is a separate question that the scan cannot answer. Watch also for outcomes that were not the primary one. A paper that leads with a secondary result usually has a primary result it would rather not lead with.

4. How big was the effect, in absolute terms?

This is where most headlines go wrong. A relative figure, "27% slower decline", sounds like a quarter of the disease removed. The absolute figure is what actually happened to the scale.

In the lecanemab trial, the adjusted mean change on the 18-point CDR-SB was 1.21 points with lecanemab and 1.66 with placebo: a difference of 0.45 points, with a 95% confidence interval from 0.67 to 0.23. The 27% in the press release is 0.45 divided by 1.66. Both numbers are true. Only the absolute one tells you the size of the thing. The trial's own conclusion used the word "moderately". Whether 0.45 points on that scale is noticeable to a family is a question the trial did not test, and clinicians still argue about it.

The donanemab trial reported a 35.1% slowing in participants with low or medium tau and 22.3% in the whole population. The gap between those two numbers is itself a finding: the drug did more for people earlier in the process. The same trial reported brain swelling, ARIA-E, in 24.0% of treated participants against 2.1% on placebo, and small brain bleeds in 31.4% against 13.6%. Benefits and harms should be read on the same absolute footing.

Look for the confidence interval. It is the range of effects the data are compatible with. An interval that includes zero means the study could not rule out no effect. A narrow interval around a small number is a precise estimate of a small effect, which is a different thing from no effect and a different thing from a big one.

Finally, when a meta-analysis pools studies, look at how much they disagree. A 2025 meta-analysis in Neurology of nine phase 3 anti-amyloid trial cohorts with 6,315 patients reported a pooled ARIA-E incidence of 9.5%, with a 95% confidence interval from 2.8% to 27.3%, and an adjusted figure of 25.5% after removing outliers. An interval that wide is the study telling you the trials it pooled were not measuring the same thing in the same way.

5. Who paid, and who wrote it?

Funding does not make a result wrong. It does tell you which direction the pressure runs. The lecanemab trial was funded by Eisai and Biogen, the companies that sell the drug, and says so in its final sentence. Industry trials are often the best-run trials in the field, because the stakes justify the expense; they are also the ones where the choice of outcome, comparator, population and follow-up length was made by people with an interest in the answer.

Modelling studies deserve the same question. A 2024 cost-effectiveness analysis in Neurology concluded that lecanemab was not cost-effective against standard care at its current price, and would become so only below about $5,100 a year. That conclusion is a consequence of the model's assumptions about how long the benefit lasts and what a year of better function is worth. The paper's limitations section says exactly that. A different set of assumptions gives a different number, which is why models from different groups disagree.

6. What do the authors admit?

The limitations paragraph, usually near the end of the discussion, is the most honest part of most papers and the least read. Good authors list the things that could make their result wrong: who was excluded, what was not measured, how long people were followed, whether the finding has been repeated elsewhere. The lecanemab authors ended their abstract with "longer trials are warranted". The donanemab trial's population was pre-selected for confirmed amyloid and intermediate or high tau, which limits how far its result travels. If a paper has no limitations section, that is itself a limitation.

One more thing to admit on the reader's side. A single study, however good, is one observation. The question is never "what did this study find?" but "what does this study add to what was already known?" That is what the library is for.

What this library looks like through those questions

This site holds 27,380 study summaries, and the six questions above are roughly what our usefulness score tries to automate. Applied to the library as a whole, on the day this page loaded:

• Design. A study design could be read from the record, either from the summary or from PubMed's own classification, for 7,460 studies, 27.2% of the library. For the other 19,920 no design is recorded, which is a gap in our data, not a verdict on the study. 1,324 are recorded as randomised or phase 2/3 trials and 1,514 as meta-analyses or systematic reviews.
• Size. A participant count is recorded for 9,360 studies, 34.2% of the library.
• Strength. Our summaries rate the evidence as high for 2,887 studies, moderate for 5,244 and low for 18,905, which is 69.0% of the library. "Low" usually means a single observational study, a laboratory finding or a review without pooled data. It is where most research sits, in any field.
• Review. 26,348 studies, 96.2%, are classed as peer-reviewed. The rest are preprints, reports, abstracts, press releases and regulatory decisions, each labelled as such on its page.
• Score. The median usefulness score is 46 out of 100. A low score means a study cannot bear much weight on its own, not that it is wrong.

Those figures are computed from the library file each time this page loads, so they will change as the library grows. The point of them is the shape, not the decimals: most research is small, observational and preliminary, and the handful of large randomised trials carry a disproportionate share of what is actually known. Reading well means knowing which kind of paper you are holding.