Forecasting Subjective Cognitive Decline: AI Approach Using Dynamic Bayesian Networks

Abstract

BackgroundSeveral potentially modifiable risk factors are associated with subjective cognitive decline (SCD). However, developmental patterns of these risk factors have not been used before to forecast later SCD. Practical tools for the prevention of cognitive decline are needed. ObjectiveWe examined multifactorial trajectories of risk factors and their associations with SCD using an artificial intelligence (AI) approach to build a score calculator that forecasts later SCD. In addition, we aimed to develop a new risk score tool to facilitate personalized risk assessment and intervention planning and to validate SCD against register-based dementia diagnoses and dementia-related medications. MethodsFive repeated surveys (2000-2022) of the Helsinki Health Study (N=8960; n=7168, 80% women, aged 40-60 years in phase 1) were used to build dynamic Bayesian networks for estimating the odds of SCD. The model structure was developed using expert knowledge and automated techniques, implementing a score-based approach for training dynamic Bayesian networks with the quotient normalized maximum likelihood criterion. The developed model was used to predict SCD (memory, learning, and concentration) based on the history of consumption of fruit and vegetables, smoking, alcohol consumption, leisure time physical activity, BMI, and insomnia symptoms, adjusting for sociodemographic covariates. Model performance was assessed using 5-fold cross-validation to calculate the area under the receiver operating characteristic curve. Bayesian credible intervals were used to quantify uncertainty in model estimates. ResultsOf the participants, 1842 of 5865 (31%) reported a decline in memory, 2818 of 5879 (47.4%) in learning abilities, and 1828 of 5888 (30.7%) in concentration in 2022. Physical activity was the strongest predictor of SCD in a 5-year interval, with an odds ratio of 0.76 (95% Bayesian credible interval 0.59-0.99) for physically active compared to inactive participants. Alcohol consumption showed a U-shaped relationship with SCD. Other risk factors had minor effects. Moreover, our validation confirmed that SCD has prognostic value for diagnosed dementia, with individuals reporting memory decline being over 3 times more likely to have dementia in 2017 (age 57-77 years), and this risk increased to more than 5 times by 2022 (age 62-82 years). The receiver operating characteristic curve analysis further supported the predictive validity of our outcome, with an area under the curve of 0.78 in 2017 and 0.75 in 2022. ConclusionsA new risk score tool was developed that enables individuals to inspect their risk profiles and explore potential targets for interventions and their estimated contributions to later SCD. Using AI-driven predictive modeling, the tool can aid health care professionals in providing personalized prevention strategies. A dynamic decision heatmap was presented as a communication tool to be used at health care consultations. Our findings suggest that early identification of individuals with SCD could improve targeted intervention strategies for reducing dementia risk. Future research should explore the integration of AI-based risk prediction models into clinical workflows and assess their effectiveness in guiding lifestyle interventions to mitigate SCD and dementia.