Improving AI Accuracy for Alzheimer’s Tau-Positivity Prediction

Why is predicting tau-positivity difficult across cohorts?

Apparent differences in machine-learning models for predicting tau-positivity often stem from the way research cohorts are structured and which variables are available, rather than genuine biological divergence. Researchers from the University of Chicago released a preprint on October 21, 2026, investigating this issue across major Alzheimer’s disease research initiatives.

The ADNI study has been running since 2004 and has contributed data to over 1,000 scientific publications. Despite this, cross-cohort generalizability remains a challenge because cohorts like ADNI, OASIS3, and NACC differ in their enrollment criteria, clinical populations, and endpoint definitions.

The researchers used a feature-block ablation framework to determine if these prediction gaps persist after harmonizing features and separating the signal from clinical severity. The ADNI data used in this work were downloaded from the repository on March 6, 2026.

How do researchers harmonize data across ADNI, OASIS3, and NACC?

To allow for a fair comparison between research cohorts, scientists select a set of shared variables that are consistently present across the different datasets. The researchers harmonized data from ADNI, OASIS3, and NACC by mapping available information to a common set of predictors.

Initial harmonization included variables such as age, sex, education, APOE ε4 count, APOE ε4 carrier status, CDR-SB, FAQ total score, and MoCA total score. For the final matched-feature analysis, the team restricted the selection to age, APOE ε4 count, APOE ε4 carrier status, education, sex, CDR-SB, and FAQ total score. This process ensures that the model performance is not skewed by different available data points between the ADNI and OASIS3 cohorts.

How do Alzheimer’s disease research cohorts impact AI?

Clinical severity is a strong predictor of tau burden, as evidenced by models that rely solely on severity measures retaining 86-99% of the performance seen in reference models. In Alzheimer’s research, tau positron emission tomography is the standard for quantifying neurofibrillary tangle burden, which links directly to disease stage and clinical outcomes. While many individuals show a canonical concordance between their tau burden and clinical severity, others are classified as resilient or vulnerable.

Researchers define amyloid status using specific cutoffs, such as 1.42 for Pittsburgh compound B and 1.19 for florbetapir in the OASIS-3 cohort. To ensure consistency, the Gaussian mixture model midpoint was selected as the primary OASIS-3 definition to align with ADNI's meta-temporal tau positivity endpoint. This alignment is important because previous work by Lin et al. used structural MRI and amyloid-PET alongside demographics to classify tau status in amyloid-positive individuals.

The OASIS-3 cohort data shows that the 1.22 cutoff for tau-positivity classified 46.5% of tau180 participants and 45.1% of tau90 participants as positive. At this 1.22 cutoff, 51.1% of tau-positive participants were actually amyloid-negative, a figure that dropped to 5.6% when using the Gaussian mixture model midpoint. Because the OASIS-3 tau90 cohort had only 12 positive cases, it was treated as supplementary alongside the NACC strict A/T cohort. Demographic and severity blocks, including measures like animal and vegetable category fluency, were used to standardize these comparisons. The ADNI initiative has was a model for the design and implementation of other initiatives related to Alzheimer's disease and other conditions.

How was the machine learning model built and validated?

The machine learning approach relied on balanced logistic regression, which was chosen for its interpretability, stability with modest sample sizes, and clinical auditability. The team trained these models using five-fold stratified cross-validation with a fixed random seed of 42. To ensure findings were not dependent on the primary method, they also evaluated random forests and extra-trees classifiers as robustness checks in supplementary analyses.

Data preprocessing was essential to handle variations across cohorts before training began. The team used median imputation and standard scaling for numeric features. For categorical features, they applied most-frequent imputation and one-hot encoding to prepare the data for analysis.

What were the results of the matched-feature comparison?

In the primary matched-feature comparison, the model achieved an area under the receiver operating characteristic curve of 0.851 for ADNI tau90 and 0.860 for ADNI tau180. The performance for the OASIS-3 tau180 cohort reached an area under the curve of 0.933. Researchers summarized these results using mean cross-validated area under the curve, average precision, balanced accuracy, and F1 score. They estimated 95% bootstrap confidence intervals for these metrics using 1000 subject-level resamples.

The ADNI tau90 cohort included 438 subjects with 78 testing positive, while the ADNI tau180 cohort included 547 subjects with 84 testing positive. The participants in these ADNI cohorts had a mean age of 72.90 years and 72.78 years respectively, within a 40 to 100 year range. In comparison, the OASIS-3 tau90 cohort included 142 subjects with 12 testing positive, and the tau180 cohort included 368 subjects with 45 testing positive. A supplementary NACC strict A/T cohort added 413 subjects, including 109 positive cases, using flortaucipir and MK-6240 tracers.

How did the team handle missing data and cohort differences?

An audit of OASIS3 files showed that while age and APOE data were recoverable, clinical scores like MoCA and CDR were absent from the master extract. To address this, the team built subject-level tables by selecting the best candidate row for each person, prioritizing data that was closest in time to the tau-PET scan and had the most complete information.

For the initial analysis, the team excluded features with more than 35% missing data. A later matched-feature rerun used a shared set of variables and applied foldwise imputation to handle remaining gaps. The ADNI study, which provides much of this data, was funded through 2021.

What sensitivity analyses were performed to ensure robustness?

The researchers performed sensitivity analyses to verify the robustness of their findings. To define OASIS3 tau-positivity, researchers fitted a two-component Gaussian mixture model to the tauopathy SUVR summary variable using 10 random initializations with a fixed random seed, setting the threshold at the midpoint of the two component means.

A threshold-sensitivity analysis confirmed these results using a 1.22 cutoff. They also used ridge-regression to model continuous Tauopathy SUVR in OASIS3 to verify that conclusions did not depend solely on the binary GMM threshold. Finally, the researchers analyzed the CDR-SB = 0 subgroup to see if cross-cohort stage asymmetry persisted in subjects who had no detectable functional impairment.

What is the background of the ADNI study?

In 2003, the Alzheimer’s Disease Neuroimaging Initiative (ADNI) began as a public-private partnership under the leadership of Principal Investigator Michael W. Weiner, MD. The study was set up to test whether biological markers and clinical assessments could measure the progression of mild cognitive impairment and early Alzheimer’s disease.

ADNI, operating as a multisite project, seeks to enhance clinical trials focused on preventing and treating the condition. Researchers at 63 sites across the United States and Canada monitor disease progression in subjects with Alzheimer’s, those who may develop it, and controls who show no signs of cognitive impairment. The study tracks these changes using genetic, biochemical, and neuroimaging markers.

To ensure results are comparable across different research centers, ADNI developed a set of standardized protocols. The initiative follows a data-sharing policy that provides all information to qualified researchers worldwide without an embargo.