Longitudinal muti-modal data prediction model for mild cognitive impairment by deep survival analysis

Background: Timely prediction of cognitive deterioration patterns in Mild Cognitive Impairment (MCI) patients proves essential for implementing optimal therapeutic interventions. We aimed to develop a deep survival analysis model using longitudinal multi-modal data to predict dementia conversion probabilities, enabling personalized treatment planning in clinical practice. Methods: We used a deep neural network model designed for survival analysis to predict the progression from MCI to Alzheimer’s Disease (AD) using longitudinal biomarkers encompassing neuropsychological assessments and neuroimaging results, along with baseline demographic characteristics and genetic risk indicators from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. Results: This study enrolled 922 baseline MCI patients for analysis. The prognostic model exhibited outstanding predictive capability, attaining cdAUC values of 0.9089±0.01 alongside Brier scores of 0.1651±0.01 at Δt=1 on the test set, when all variable sets were incorporated into the time-dependent Cox survival neural network (tdCoxSNN) model. Through feature significance evaluation, the Functional Activities Questionnaire (FAQ) emerged as the most influential predictive element. Conclusions: Through the systematic integration of various longitudinal biological markers, we developed a dynamic prediction model for MCI using deep survival analysis, enabling effective individual risk stratification and facilitating early identification of high-risk individuals as well as supporting clinical decision-making.