Fig. 1

Method framework visualization. Data from the first six exam visits were used for prediction of time-to-CVD event in the subsequent 17 years. Six strategies were employed to incorporate longitudinal data in 35 variables. All models were trained and tested under the same 5-fold x 2 times cross-validation scheme. Model output was survival probabilities (1-predicted CVD risk) over time. Model performance was quantified by C-index, AUC, Brier Score, and other metrics. CVD: cardiovascular disease; Cox: Cox proportional hazards; LASSO-Cox: Cox Proportional Hazards penalized by Least Shrinkage and Selection Operator; Dynamic-Deephit: recurrent neural network-based survival method for longitudinal data; JMBayes: joint modeling under Bayesian approach; RSF: random survival forest