Development and Internal Validation of a Prognostic Model for 15-year Risk of Alzheimer Dementia in Primary Care Patients

Background: The exploitation of routinely collected clinical health information is warranted to optimize the case detection and diagnostic workout of Alzheimer's disease (AD). We aimed to derive an AD prediction score based on routinely collected primary care data.

Methods: We built a cohort selecting 199,978 primary care patients 60 + part of the Health Search Database between January 2002 and 2009, followed up until 2019 to detect incident AD cases. The cohort was randomly divided into a derivation and validation sub-cohort. To identify AD and non-AD cases, we applied a clinical algorithm that involved two clinicians. According to a nested case-control design, AD cases were matched with up to 10 controls based on age, sex, calendar period, and follow-up duration. Using the derivation sub-cohort, 32 potential AD predictors (sociodemographic, clinical, drug-related, etc.) were tested in a logistic regression and selected to build a prediction model. The predictive performance of this model was tested on the validation sub-cohort by mean of explained variation, calibration, and discrimination measurements.

Results: We identified 3223 AD cases. The presence of memory disorders, hallucinations, anxiety, and depression and the use of NSAIDs were associated with future AD. The combination of the predictors allowed the production of a predictive score that showed an explained variation (pseudo-R) for AD occurrence of 13.4%, good calibration parameters, and an area under the curve of 0.73 (95% CI: 0.71-0.75). In accordance with this model, 7% of patients presented with a high-risk score for developing AD over 15 years.

Conclusion: An automated risk score for AD based on routinely collected clinical data is a promising tool for the early case detection and timely management of patients by the general practitioners.