Table 3.
Main results from the prediction models. Predictions were performed with data available from four different time frames in the patient disease trajectories (left column): On diagnosis (Diagnoses model), On hospital admission and 12-h into admission (Admission model), 12 h leading up to Intensive Care Unit (ICU) admission (Pre-ICU model) and 12 h after ICU admission (post-ICU model).
| Hospital admission | ICU admission | Ventilator treatment | Death | |||||
|---|---|---|---|---|---|---|---|---|
| TPR/FPR | Pre/Rec | TPR/FPR | Pre/Rec | TPR/FPR | Pre/Rec | TPR/FPR | Pre/Rec | |
| Diagnosis | ||||||||
| Age + Gender + BMI | 0.820 | 0.705 | 0.802 | 0.173 | 0.815 | 0.184 | 0.902 | 0.412 |
| +Comorbidities | 0.822 | 0.705 | 0.844* | 0.206 | 0.851* | 0.192 | 0.906 | 0.412 |
| +Temporal Features | – | – | – | – | – | – | – | – |
| +In-hospital Tests | – | – | – | – | – | – | – | – |
| Admission | ||||||||
| Age + Gender + BMI | – | – | 0.685 | 0.226 | 0.675 | 0.200 | 0.785 | 0.435 |
| +Comorbidities | – | – | 0.752* | 0.282 | 0.743* | 0.238 | 0.794 | 0.445 |
| +Temporal Features | – | – | 0.763* | 0.308 | 0.762* | 0.289 | 0.796 | 0.444 |
| +In-hospital Tests | – | – | 0.805*# | 0.418 | 0.786* | 0.345 | 0.818* | 0.540 |
| Pre-ICU | ||||||||
| Age + Gender + BMI | – | – | – | – | 0.598 | 0.892 | 0.733 | 0.575 |
| +Comorbidities | – | – | – | – | 0.567 | 0.869 | 0.735 | 0.548 |
| +Temporal Features | – | – | – | – | 0.563 | 0.871 | 0.738 | 0.567 |
| +In-hospital Tests | – | – | – | – | 0.502 | 0.867 | 0.721 | 0.567 |
| Post-ICU | ||||||||
| Age + Gender + BMI | – | – | – | – | 0.598 | 0.892 | 0.733 | 0.575 |
| +Comorbidities | – | – | – | – | 0.530 | 0.843 | 0.724 | 0.552 |
| +Temporal Features | – | – | – | – | 0.584 | 0.861 | 0.739 | 0.569 |
| +In-hospital Tests | – | – | – | – | 0.671 | 0.928 | 0.741 | 0.568 |
Models were trained to predict risk of hospital admission, ICU admission, ventilator treatment and death (top row).
All models were trained with incremental data, starting with age, gender and Body Mass Index, then adding comorbidity information, temporal features (e.g. vital signs) and finally by adding hospital laboratory tests where applicable. Please see supplementary tables S1 and S2 for data definitions.
Performance metrics are presented as the Receiver Operating Characteristics Area Under the Curve (ROC-AUC) for True/False positive rates (TPR/FPR) and Precision/Recall (Pre/Rec).
*Model is significantly (p < 0.01) better than the base prediction model (Age + gender + Body Mass Index, BMI).
#Model is significantly (p < 0.01) better than the comorbidities model.
§Model is significantly (p < 0.01) better than the temporal model.
--: Insufficient data available at the time point, or prediction irrelevant (e.g. predicting hospital admission for patients already in the ICU).