Multidimensional analysis and predictive modeling of cognitive decline risk in the United States using propensity...

Cognitive decline, an early indicator of neurodegenerative disorders, presents a growing public health challenge. This study aimed to integrate causal inference and machine learning to quantify the causal impact of high-risk status on cognitive function, explore geographic and temporal heterogeneity, and develop predictive models for early identification of at-risk individuals in the United States. We analyzed 22,182 records from the behavioral risk factor surveillance dystem. Causal effects were estimated using propensity score matching and doubly robust estimation to mitigate confounding. Geographic and temporal heterogeneity were assessed through stratified analyses. Predictive models were developed using random forest and XGBoost, trained on 80% of the dataset and evaluated on 20%, with performance assessed by accuracy, precision, recall, and F1 score. Propensity score matching estimated an average treatment effect (ATE) of 19.92 points (95% confidence interval: 19.70-20.15), and doubly robust estimation yielded an ATE of 16.95 points (95% confidence interval: 16.77-17.14), confirming a significant causal link between high-risk status and cognitive decline. Regional heterogeneity was pronounced, with US territories showing the highest ATE (59.40). Temporal analysis from 2015 to 2022 revealed no significant overall trend (P = .54), although annual fluctuations were observed. The random forest model achieved the best predictive performance (accuracy = 0.77, F1 = 0.67), outperforming XGBoost in balancing precision and recall. The findings demonstrate a significant causal relationship between high-risk status and cognitive decline, with marked geographic disparities. The integration of causal inference with machine learning provides a robust framework for both understanding risk factors and building practical screening tools. These results offer actionable insights for targeting public health interventions and underscore the potential of data-driven approaches in cognitive health management.