Table 2.
The summary of the included studies. The most important contents of the 21 studies are summarized.
| Study | Population | Study designs | Predicted outcome(s) | Data type(s) | Method(s) | Main contribution(s) | Identified challenge(s) towards integration of AI in practice |
|---|---|---|---|---|---|---|---|
| Zhao et al. (65) | 16,189 adult (age > 18) patients from MIMIC-IV | Retrospective training, prospective validation | Extubation failure | Clinical time series (MIMIC-IV and domestic) | Categorical boosting with SHAP and RFE | Well-performing AI model (up to 0.83 AUROC), increased interpretability, open access UI for model validation | Interpretability, dataset shift problem |
| Jentzer et al. (66) | 11,266 adult (Mean age 68 ± 15 years) patients from Mayo Clinic ICU | Retrospective data analysis | Mortality risk | Numerical clinical data extracted from ECGs (domestic) | Multivariate logistic regression | Well-performing AI model (up to 0.83 AUROC) | Interpretability |
| Gandin et al. (74) | 10,616 patients from MIMIC III | Retrospective data analysis | Mortality risk | EHR (MIMIC-III) | RNN (LSTM with attention layer) | Well-performing AI model (up to 0.79 AUROC), attention layer to increase the interpretability of LSTM | Interpretability and reliability |
| Andersson et al. (67) | 932 adult (age ≥ 18) patients from 36 ICUs across Europe and Australia | Retrospective data analysis | Neurological outcome following out-of-hospital cardiac arrest (OHCA) | Clinical variables and biomarkers (domestic-multicenter) | ANN with SHAP | Reliable AI model (up to 0.94 AUROC) using cumulative clinical data from first 3 days of ICU stay | Generalizability, effect of outliers |
| Parsi et al. (39) | 53 patients from PhysioNet | Retrospective data analysis | Paroxysmal atrial fibrillation | ECG (PhysioNet) | SVM, k-NN, RF, MLP | High performance AI (up to 0.79 accuracy) on implantable defibrillator with low computation power | Low computational power on wearable and implantable devices |
| Yu et al. (40) | 7,368 adult (age > 18) patients from MIMIC-III | Retrospective data analysis | 4-year mortality risk after cardiac surgery | Clinical time series (MIMIC-III) | LR, ANN, Ada, NB, RF, etc. with RFE | Well-performing AI model (up to 0.80 AUROC), open access UI for model validation | Generalizability |
| Wang et al. (75) | 929 adult (age > 18) patients from eICU-CRD | Retrospective training, prospective validation | Noninvasive ventilation (NIV) failure | Clinical time series (eICU-CRD and domestic) | Categorical boosting with RFE and SHAP | Well-performing AI model (up to 0.87 AUROC) applied to easily available clinical variables, open access UI for model validation | Generalizability, low specificity of AI predictions |
| Chen et al. (41) | 1,439 adult (mean age 65.05 ± 12.53 years) patients from Cheng Hsin General Hospital | Retrospective data analysis | Ventilator weaning time | Non-time series clinical data (domestic) | LR, SVM, RF, ANN, XGBoost | Well-performing AI model (up to 0.88 AUROC), identify most simplified key parameters | Generalizability |
| Dutra et al. (76) | 519 adult (age > 18, mean age, 74.87 ± 13.56 years) patients admitted to a Brazilian cardiac ICU | Ambispective data analysis | Mortality risk from heart failure with mid-range ejection fraction (EF) | Non-time series clinical data (domestic) | Cox, Kaplan–Meier, ElasticNet, survival tree | EF is not significantly correlated with mortality | Generalizability |
| Bodenes et al. (42) | 540 adult patients admitted to Brest University Hospital’s cardiac ICU | Prospective data analysis | Mortality risk and heart rate variability (HRV) | Clinical time series (domestic) | k-NN, SVM, LR, decision trees | Low cost and efficient AI model for HRV analysis | Generalizability, interpretability, lack of standardized HRV measurement methods |
| Moazemi et al. (25) | 11,513 patients from MIMIC-III and 502 from University Hospital Düsseldorf’s cardiac ICU (age ≥ 17) | Retrospective data analysis | ICU readmission | Clinical time series (MIMIC-III and domestic) | RNN (LSTM) | Well perforing AI (up to 0.82 AUROC), data-driven approach, validation with external cohort | Interpretability, dataset shift problem |
| Baral et al. (44) | 7,611 patients (age > 15) from MIMIC-III cardiac ICUs | Retrospective data analysis | Cardiac arrest | Clinical time series (MIMIC-III) | Multi-layer perceptron (MLP), RNN (bidirectional LSTM) | Well-performing AI model (up to 0.94 AUROC) to reduce false alarm for cardiac arrest, improved model compared to normal LSTM | Generalizability |
| Qin et al. (43) | 49,168 patients from MIMIC-III | Retrospective data analysis | Sepsis | Textual and structured clinical data (MIMIC-III) | NLP (BERT), Amazon Comprehend Medical for data processing, XGBoost (for classification) | Outperform PhysioNet’s sepsis prediction challenge winner (up to 0.89 AUROC) | Generalizability |
| Nanayakkara et al. (77) | Adult (age ≥ 17) septic patients from MIMIC-III | Retrospective data analysis | Sepsis treatment planning | Clinical time series (MIMIC-III) | RL | Introducing a novel physiology-driven recurrent autoencoder, highly interpretable, uncertainty quantification | Lack of standardization, how/when AI is considered safe enough for clinical routine |
| Zheng et al. (78) | 1,362 critically ill COVID patients (mean age 69.7) from New York University Langone Health | Retrospective data analysis | Managing oxygen flow rate to reduce mortality risk | EHR (domestic) | RL | AI model to identify optimal personalized oxygen flow rate to reduce mortality rate | Generalizability |
| Peine et al. (79) | 61,532 and 200,859 ICU stays of adult patients from MIMIC-III and eICU datasets | Retrospective data analysis | Optimization of mechanical ventilation to reduce mortality risk | Clinical time series (MIMIC-III and eICU) | RL | Introduce VentAI to dynamically optimize mechanical ventilation for individual patients | Generalizability, algorithm bias, missing/false data |
| Akrivos et al. (80) | 162 adult patients (18 < age < 90 on) from MIMIC-II | Retrospective data analysis | Cardiac arrest | Transformed clinical time series (MIMIC-II) | integrated model of sequential contrast patterns using Multichannel Hidden Markov Model | High sensitivity (with the average of 0.78) and specificity to identify high risk patients | False positive rate in classification results |
| Aushev et al. (81) | 75 adult (age > 18)patients from ShockOmics European database | Retrospective data analysis | Mortality due to septic and cardiogenic shock | ECG (ShockOmics Dataset) | SVM, Random Forest, RFE, Bayesian networks | Apply feature selection to identify the most relevant predictors of mortality due to septic and cardiogenic shock using ECG with high certainty (up to 0.84 AUROC) | – |
| Kim et al. (82) | 29,181 adult (age > 18) ICU patients from Yonsei Health System (Severance and Gangnam Severance Hospitals) | Retrospective data analysis | Acute respiratory failure and cardiac arrest | Time series (domestic) | Deep Learning (LSTM) | Introduce FAST-PACE for preparing immediate intervention in emergency situations, outperforming some established scoring systems (e.g., SOFA) (up to 0.88 AUROC) | Lack of relevant input data to AI models, lack of external validation, imbalanced datasets, lack of real time measurements of vital signs |
| Meyer et al. (83) | 11,492 ICU stays from 9,269 adult (age ≥ 18) patients from a German cardiovascular tertiary care center | Retrospective data analysis | Mortality, renal failure, postoperative bleeding leading to operative revision | Time series (domestic) | Deep learning (RNN) | Predict severe complications after cardiothoracic surgery with a higher certainty (up to 0.96 AUROC), validation against MIMIC-III dataset | Dataset shift, biased data, generalizability, transparency and interpretability of AI decision making |
| Yoon et al. (84) | 2,809 Adult (age > 18) patients from MIMIC-II | Retrospective data analysis | Tachycardia as a surrogate for cardiorespiratory instability (CRI) | Vital signs time series (MIMIC-II) | Regularized logistic regression (LR), Random Forest | Developed a risk score for predicting tachycardia episodes, AI model with high accuracy (up to 0.86 AUROC) | Timestamp mismatching and data sparsity, specificity of predictions, lack of external validation |