Status
Published
Role
First author
Posted
Revised

Early prediction of Alzheimer's disease and related dementias from electronic health records

Machine learning models trained on OneFlorida+ EHR data to predict Alzheimer's disease and related dementias up to five years before diagnosis, comparing knowledge-driven and data-driven feature sets across four computable phenotypes.

Prediction design: observation window, prediction window, index dateA patient's record runs left to right. Predictors are taken from the observation window. A prediction window of 0, 1, 3 or 5 years separates the end of the observation window from the index date, which for cases is the date of ADRD diagnosis.timeObservation windowEHR history used to build predictorsPrediction window0, 1, 3, 5 yearsIndex datediagnosis, for casesPrediction design: observation window, prediction window, index dateObservation windowEHR history used tobuild predictorsPrediction window0, 1, 3, 5 yearsIndex datediagnosis, for cases

Conceptual illustration

How each prediction was set up

Predictors come only from the observation window. The prediction window, from 0 to 5 years, separates them from the index date. Redrawn for this page from the design described in the paper.

The question

Alzheimer’s disease and related dementias (ADRD) are usually diagnosed late. If routine electronic health records carry signals years before diagnosis, models built on them could help identify people at higher risk earlier, when there is more room for prevention, planning and trial enrollment.

Data and design

The study used EHR data from the OneFlorida+ Research Consortium: 23,835 patients with ADRD and 1,038,643 controls. Cases were identified with four different computable phenotyping algorithms, so the results could be compared across case definitions rather than resting on one. Each model looked at a patient’s record up to a cutoff and predicted diagnosis 0, 1, 3 or 5 years later, as in the figure above.

Two strategies defined the predictors:

  • Knowledge-driven: established ADRD risk factors chosen from the literature, such as stroke, depression and diabetes.
  • Data-driven: the broad set of diagnoses, procedures, medications and other features available in the record, with the model left to find what matters.

Logistic regression and gradient boosting trees were trained under each strategy.

Results

Gradient boosting trees with data-driven features performed best, with areas under the ROC curve of 0.939, 0.906, 0.884 and 0.854 for prediction 0, 1, 3 and 5 years before diagnosis. Performance declined as the prediction window lengthened, as expected. SHAP analysis identified the clinical and sociodemographic factors contributing most to the models, with age and cerebrovascular disease among them.

SHAP summary plot for the five-year data-driven gradient boosting model. Each row is a feature, ordered by its contribution; points show individual patients, colored by feature value.

Published figure

Features contributing most to the five-year model

SHAP values for the data-driven gradient boosting model predicting ADRD five years before diagnosis. Shown unmodified.

Source: Li et al., Alzheimer's & Dementia 2023, CC BY-NC-ND 4.0

What it does not show

These are retrospective models in one research network. External validation in other populations, and evidence that earlier identification changes care, are separate questions that this study does not answer.

Later work

A 2026 paper in JAMIA Open, on which I am the second author, examined whether factors related to ADRD can be identified in real-world data.

First published on my earlier site, andeeli.github.io/posts/AD-Prediction/, and revised for this one.