UCSF model predicts Alzheimer's from health records up to seven years ahead

Researchers at the University of California, San Francisco (UCSF) trained machine learning models on routine electronic health records to predict the onset of Alzheimer's disease up to 7 years before it was diagnosed. The study, by Alice Tang, Marina Sirota and colleagues at UCSF's Bakar Computational Health Sciences Institute, was published in Nature Aging on 21 February 2024 ¹.

What the study found

The team trained random forest models on a cohort of 749 individuals with Alzheimer's disease and 250,545 controls drawn from UCSF's records. The models predicted onset with a mean area under the receiver operating characteristic curve (AUROC) of 0.72 seven years before onset, rising to 0.81 one day before ¹. The UCSF clinical database from which the cohort was drawn holds records on more than 5 million patients, and the university's news release describes the seven-year figure as 72% predictive power ².

Matched cohort models identified conditions with predictive power before onset. In both men and women these included high cholesterol, hypertension and vitamin D deficiency. In men they also included erectile dysfunction and an enlarged prostate, and in women osteoporosis ². Knowledge networks, including UCSF's SPOKE, highlighted genes shared between several of the top predictors and Alzheimer's disease, among them APOE, ACTB, IL6 and INS ¹. Genetic colocalisation analysis supported an association between Alzheimer's and hyperlipidaemia at the APOE locus, and a stronger association in women between Alzheimer's and osteoporosis at a locus near MS4A6A ¹.

Alice Tang, an MD/PhD student in the Sirota laboratory and the paper's first author, said: "The power of this AI approach comes from identifying risk based on combinations of diseases" ².

What it means for the market

For health systems, the study shows a prediction signal in data they already hold: diagnoses recorded in routine care, with no imaging, blood test or genetic sequencing added. The authors write that clinical data "can be utilized for early AD prediction and identification of personalized biological hypotheses" ¹. For suppliers of risk-stratification and population-health tools, the method is a random forest on coded diagnoses rather than a proprietary biomarker, and the sex-specific findings point to different risk profiles for men and women ¹.

The limits are in the numbers. The abstract reports results from UCSF's own records only, and describes the biological findings as hypotheses to be prioritised rather than confirmed mechanisms ¹. The cohort was 749 cases against 250,545 controls, and the seven-year AUROC of 0.72 sits well below the 0.81 reached one day before onset ¹.