Table 4.
Comparison of currently available Deep learning models for mortality in the inpatient setting.
| Current | Li et al. (2020)* | Zhu et al. (2020)** | Vaid et al. (2021)Ø | Booth et al. (2021)^ | Chowdhury MEH et al. (2021)*** | |
|---|---|---|---|---|---|---|
| Sample Size (n) | 1214 | 102 | 181 | 4029 | 398 | 375 |
| Endpoints | End of admission | ICU admission and mortality | End of admission | 7-day mortality | End of admission | End of admission |
| Clinical setting | In-patient | Non-ICU and ICU | In-patient | In-patient | In-patient | In-patient |
| Training and Test (set) split | 70/30 | 90/10 | 85/15 | 70/30 | 80/20 | |
| Datapoint/Variables | 47 | 15 | Top 5 lab features | 38 | Top 5 lab features | Top 5 lab features |
| Testing accuracy | 87.7 | 85.3 | nr | 76.2# (median) | nr | nr |
| Training accuracy | 98.7 | 89.2 | nr | nr | nr | nr |
| AUROC | 0.885 | 0.844 | 0.968 | 0.809# (median) | 0.930§ | 0.970 |
| Sensitivity | 33.3 | 75 | nr | 77.8# (median) | 91.0§ | 89.4–94.6 |
| Specificity | 95.3 | 87.2 | nr | 69.9# (median) | 91.0§ | 89.0–95.0 |
| PPV | 50.0 | 52.2 | nr | nr | 62.5§ | 9.1# |
| NPV | 91.1 | 94.9 | nr | nr | 98.4§ | 92.9# |
nr = not reported.
Calculated from data given in paper.
Top 5 serum markers for mortality form 26 only lab parameters, with an unbalanced training set.
Non-normalized AUROC, sensitivity, specificity, PPV, NPV.
MLP federated model (best model, median values were calculated from Supplemental Table 6), some variables have large amount of missing data Supplemental Table 1.
Mortality data used.
Top 5 variables and cut off (very high cutoffs) > 6.7 mg/L for D‐dimer, <94 for O2 index, >10 for NE:LY, >93 mg/L for CRP, and >450 U/L for LDH.
Top 5 feature for mortality prediction: lactate dehydrogenase, neutrophils (%), lymphocytes (%), high-sensitivity C-reactive protein, and age score (TP = 160, FP = 16, FN = 14 and TP = 184).