Table 1.
Main study characteristics.
| Author [Reference] |
Study Type | Country of Study Population |
Relevant Setting of Collected Data (ED, ICU, or Prehospital) |
Outcome Predicted | Sample Size of Training Dataset |
Sample Size of Test Dataset |
Model Performance | TRIPOD Classification |
|---|---|---|---|---|---|---|---|---|
| Diagnostic | ||||||||
| Brinati, et al. [38] | Retrospective | Italy | ED | Positive COVID-19 status | 279 | N/A (cross-validation) |
Random forest (C-index = 0.84) | 1b |
| Fraser, et al. [39] | Prospective | Canada | ICU | Positive COVID-19 status | 20 | N/A (cross-validation) |
Decision tree (accuracy = 98%) | 1b |
| Vasse, et al. [40] | Retrospective | France | ED | Positive COVID-19 status | 744 | 2390 | Decision tree (Sensitivity = 60.5%, Specificity = 89.7%) | 2b |
| Prognostic | ||||||||
| Abdulaal, et al. [30] | Retrospective | United Kingdom | ED | In-patient mortality | 318 | 80 | Neural network (C-index = 0.901) | 2a |
| Assaf, et al. [31] | Retrospective | Israel | ED; ICU | Critical illness (mechanical ventilation, ICU admission, multi-organ failure, and/or death) | 162 | N/A (cross-validation) |
Random forest (C-index = 0.93) | 1b |
| Burdick, et al. [32] | Prospective | United States | ICU | Decompensation leading to mechanical ventilation within 24 h | 49,623 | 197 | Gradient boosting machine (C-index = 0.866) | 3 |
| Burian, et al. [33] | Prospective | Germany | ICU | ICU admission | 65 | N/A (cross-validation) |
Random forest (C-index = 0.79) | 1b |
| Cheng, et al. [34] | Retrospective | United States | ICU | ICU admission within 24 h | 401 | 521 | Random forest (C-index = 0.799) | 2a |
| Durhan, et al. [41] | Retrospective | Turkey | ICU | ICU admission (software evaluates the extent of normal lung parenchyma) | 90 | N/A | Deep learning software (C-index = 0.944) | N/A |
| Jackson, et al. [35] | Retrospective | United States | ICU | Invasive mechanical ventilation | 297 | N/A | Fast-and-frugal decision tree (accuracy = 70%) | 1a |
| Liang, et al. [36] | Retrospective | China | ICU | Critical illness (ICU admission, invasive ventilation, death) | 1590 | 710 | Deep learning survival Cox model (C-index = 0.852–0.967) |
2b |
| Mushtaq, et al. [42] | Prospective | Italy | ICU | ICU admission (software evaluates the extent of lung opacity and consolidation) | 697 | N/A | Deep learning software based on convolutional neural networks (C-index = 0.77) |
N/A |
| Schwab, et al. [37] | Retrospective | Brazil | ICU | ICU admission | 391 | 167 | Support vector machine (C-index = 0.98) | 2a |
| Resource optimisation | ||||||||
| Belciug, et al. [43] | Retrospective | Italy | ICU | Developed a model for simulating ICU bed occupancy | N/A | N/A | Artificial immune system algorithm (no accuracy measure estimated) | N/A |
COVID-19: coronavirus disease 2019, ED: Emergency Department, N/A: Not applicable, ICU: Intensive Care Unit; a: Performance of the best performing model is reported if multiple models were constructed. Only the performance on the strictest form of validation is reported. A range is given if the model was validated on multiple datasets. b: TRIPOD classification according to strictest validation used (higher values indicate stricter classification, i.e., type 3 is the strictest amongst included studies). 1a: Performance is evaluated directly on the same data; 1b: Performance and optimism of the model are evaluated using re-sampling techniques, such as bootstrapping or k-fold cross-validation; 2a: Model development and performance evaluation are done separately on a random split of the data, such as a train-test split; 2b: Model development and performance evaluation is done separately on a non-random split of the data by time, location, or both; 3: Model development and performance evaluation are conducted on separate data sets, for example, from different studies.