Table 4 Long length of stay prediction on MIMIC-III. U, T, S represents unstructured data, temporal signals, and static information respectively
From: Combining structured and unstructured data for predictive models: a deep learning approach
| Model | Model inputs | F1 | AUROC | AUPRC | P value | |
|---|---|---|---|---|---|---|
| Baseline models | LR | \(T + S\) | 0.668 (0.658, 0.678) | 0.735 (0.732, 0.738) | 0.615 (0.611, 0.619) | 1 |
| LR | U | 0.686 (0.683, 0.689) | 0.736 (0.732, 0.740) | 0.614 (0.610, 0.618) | 0.5643 | |
| LR | \(U + T + S\) | 0.703 (0.699, 0.707) | 0.773 (0.770, 0.776) | 0.642 (0.637, 0.647) | < 0.001 | |
| RF | \(T + S\) | 0.523 (0.462, 0.584) | 0.695 (0.689, 0.701) | 0.586 (0.577, 0.595) | < 0.001 | |
| RF | U | 0.568 (0.479, 0.657) | 0.651 (0.642, 0.660) | 0.559 (0.553, 0.565) | < 0.001 | |
| RF | \(U + T + S\) | 0.537 (0.533, 0.541) | 0.718 (0.714, 0.722) | 0.597 (0.591, 0.603) | < 0.001 | |
| Deep models | Fusion-CNN | \(T + S\) | 0.674 (0.667, 0.681) | 0.748 (0.745, 0.751) | 0.640 (0.635, 0.645) | < 0.001 |
| Fusion-CNN | U | 0.695 (0.683, 0.707) | 0.742 (0.741, 0.743) | 0.635 (0.632, 0.638) | < 0.001 | |
| Fusion-CNN | \(U + T + S\) | 0.725 (0.718, 0.732) | 0.784 (0.781, 0.787) | 0.662 (0.658, 0.666) | < 0.001 | |
| Fusion-LSTM | \(T + S\) | 0.690 (0.684, 0.696) | 0.757 (0.756, 0.758) | 0.644 (0.643, 0.645) | < 0.001 | |
| Fusion-LSTM | U | 0.702 (0.697, 0.707) | 0.746 (0.745, 0.747) | 0.637 (0.634, 0.640) | < 0.001 | |
| Fusion-LSTM | \(U + T + S\) | 0.716 (0.711, 0.721) | 0.778 (0.776, 0.780) | 0.660 (0.657, 0.663) | < 0.001 |