Table 4.

The performance comparison of the proposed models and the baseline models.

Models	Accuracy	PPIA	SPIA
LSTM with TI	0.9906 ± 0.0043	0.9894 ± 0.0074	0.9912 ± 0.0039
LSTM w/o TI	0.9843 ± 0.0057	0.9792 ± 0.0117	0.9849 ± 0.0053
LR with average aggregation	0.7955 ± 0.0216	0.7126 ± 0.0345	0.7986 ± 0.0211
LR with two most recent visits	0.6652 ± 0.0162	0.5057 ± 0.0409	0.6674 ± 0.0163
LR with the most recent visit	0.6803 ± 0.0243	0.5209 ± 0.0370	0.6825 ± 0.0243
SVM with average aggregation	0.7445 ± 0.0237	0.6465 ± 0.0516	0.7468 ± 0.0226
SVM with two most recent visits	0.6533 ± 0.0165	0.4825 ± 0.0445	0.6552 ± 0.0163
SVM with the most recent visit	0.6746 ± 0.0209	0.4931 ± 0.0245	0.6757 ± 0.0208
DT with average aggregation	0.7035 ± 0.0206	0.6223 ± 0.0267	0.7058 ± 0.0200
DT with two most recent visits	0.5810 ± 0.0199	0.4463 ± 0.0470	0.5829 ± 0.0196
DT with the most recent visit	0.5916 ± 0.0204	0.4705 ± 0.0458	0.5934 ± 0.0196
RF with average aggregation	0.6916 ± 0.0223	0.5786 ± 0.0487	0.6944 ± 0.0227
RF with two most recent visits	0.6373 ± 0.0181	0.4517 ± 0.0430	0.6399 ± 0.0179
RF with the most recent visit	0.6416 ± 0.0183	0.4570 ± 0.0422	0.6441 ± 0.0186

LSTM with TI and LSTM w/o TI are implemented based on the dataset of patients with more than 3 visits. Baseline models with average aggregation are trained with aggregated features derived from the longitudinal data. Baseline models with two most recent visits are trained directly with the information of the $\left(\mathit{N}-1\right){}^{\mathrm{th}}$ visit and the N^th visit among historical visits. Baseline models with the most visit are trained directly with the N^th visit. LSTM with TI model and all the baseline models are trained with time intervals, while LSTM w/o TI is trained without time intervals. Note that the results presented here are mean values and the standard deviation values of the 10-fold cross validation, and the performances of each fold are provided in the Supplementary Tables S2–S7.