An autopsy report is rarely a single word. In one of the most cited studies of the problem, Patricia Boyle and colleagues at Rush examined the brains of 1,079 people who had been followed with repeated cognitive testing until death, and found that 78 per cent carried two or more distinct neuropathologies and 58 per cent carried three or more (Annals of Neurology, 2018). A peer-reviewed autopsy analysis published in April 2026 in the Journal of Neuropathology and Experimental Neurology, looking at broader diagnostic categories rather than individual lesions, found only about a third of its sample had Alzheimer's change on its own. So the honest question for anyone sitting in a memory clinic is not "is it Alzheimer's" but "how much of the list can anyone see while I am alive?"
A preprint posted to bioRxiv on 23 May 2026 by Christopher Patterson and colleagues, with senior author Paul Thompson at the University of Southern California, put a number on that. Using harmonised ante-mortem data from the Alzheimer's Disease Sequencing Project Phenotype Harmonization Consortium, covering 7,894 people who had both research assessments in life and a neuropathological examination after death (cohorts include the Adult Changes in Thought study and the National Alzheimer's Coordinating Center), the team ran automated machine learning across 12 different predictor sets: clinical and cognitive composites, structural MRI, diffusion MRI, image-derived white matter hyperintensity measures, and cerebrospinal fluid biomarkers. The target was 26 autopsy-confirmed neuropathological outcomes. Models were ensembles trained with stratified cross-validation, scored with Spearman's rank correlation for graded outcomes and the Matthews correlation coefficient for yes or no ones, a statistic that sits at zero when a model is doing no better than chance. This is a preprint: it has not been through peer review.
What held up was Alzheimer's staging, and it held up in a way that flatters the routine clinic. From clinical and cognitive data alone, the models tracked Braak tau stage at a rank correlation of 0.577, with the amyloid B score at 0.515, Thal phase at 0.497 and CERAD neuritic plaque score at 0.463. Adding spinal fluid pushed the top figures to 0.683, 0.678 and 0.666. That is the biology working as advertised: the two proteins we have assays for are the two the models can stage. But the added value was narrow. In only four of the 26 outcomes did adding CSF significantly beat clinical features alone in the same people.
The collapse is the story. In the CSF analyses, Lewy body disease came in at a Matthews coefficient of 0.098, not statistically significant, which is to say the model was close to guessing. Pathologic TDP-43 reached 0.231. Most vascular brain injury outcomes fell away too. Ordinal Parkinson's disease Braak staging, the authors write, was poorly predicted across every feature set they tried. Their own phrasing for the pattern is a quantified "diagnostic gap for non-AD proteinopathies". A caution on interpretation: the paper does not report a person-by-person confidence score that dips when a brain is mixed. What it reports is more useful and more damning, that the specific pathologies which most often accompany Alzheimer's change are the ones the models predict worst, so a mixed brain is by construction the case a model gets least right.
The sample arithmetic is its own warning. The base clinical models drew on 3,345 to 7,865 people per outcome. The CSF models had 39 to 281. The diffusion MRI models had 49 to 220, because diffusion scans are not routinely collected across cohorts. The authors flag that small CSF subset as underpowered and potentially selected. And every one of these numbers comes from cross-validation inside research autopsy cohorts, where people are highly educated, willing to donate their brain, and referred to specialist centres. Nothing here tells you how such a model behaves in general practice, on a first visit, in a patient nobody has scanned twice.
There is a constructive half. Imaging carried signal exactly where fluid did not: diffusion metrics best captured vascular brain injury and white matter injury and outperformed both clinical measures and anatomical MRI for Lewy body disease; structural MRI beat standard clinical assessment for neurodegeneration and hippocampal sclerosis; white matter hyperintensities added to cognitive measures for TDP-43. For families, that maps onto something plain. The blood test the FDA cleared on 16 May 2025 answers one question, whether amyloid plaques are present, and answers it well. It says nothing about the second, third or fourth entry on an eventual autopsy report, and for now neither does a well-built machine learning model.