In general, the greater the number of data sets to which a machine learning model is applied, the stronger its generalization capabilities. Therefore, we investigated the inclusion cohorts and distribution centers of each study and classified these studies into single-site or multisite studies accordingly. We also classified studies by c, the sample size of studies: c<500, 500<c<2000, 2000<c<5000, 5000<c<10,000, 10,000<c<50,000, and c>50,000.

The model methods or algorithms used in each paper were reviewed for analysis, and model methods were categorized as supervised learning, unsupervised learning, or reinforcement learning.

We reviewed variables or features used for modeling in each study. According to routine intensive care unit practices, we classified these variables into 12 groups: demographic variables, vital signs, symptoms, laboratory values, ventilation parameters, medications, nonmedicine therapy, comorbidities, fluid balance, scores, medical history, and outcome. Given the wide range of variable expressions in papers, such as formal medical terms, abbreviations, acronyms, and capitalizations, variable name normalization was implemented using text processing and manual annotation methods. As some studies used self-defined features or derived data for their special study purpose, variables used in only 1 study were excluded.

To determine the applicability and potential impact of various machine learning models for clinicians and patients (ie, in applications), model evaluation methods are important components of model development. We reviewed evaluation metrics used for measuring model performance.

In addition to overall quantitative analysis, which included all studies, selected papers were divided into 4 topics for detailed analysis: detection and monitoring for diagnosis, early identification of clinical events, patient outcome prediction, and treatment decisions.

Among 13 studies, 4 (30.8%) studies [12-15] focused on monitoring or detection of physiological indicators, 3 studies (23.1%) [16-18] focused on the of mechanical ventilation abnormalities (in particular, patient-ventilator asynchrony), 4 studies (30.8%) [19-22] used electroencephalography (EEG) to diagnose brain diseases, and 2 studies (15.4%) [11,23] studies focused on infections. Variables used included demographic variables (n=5), vital signs (n=6), laboratory values (n=5), ventilation parameters (n=5), comorbidities (n=1), and outcome (n=1).

Most data were obtained from a single center (11/13, 84.6%), and only 2 studies (2/13, 15.4%) used multicenter data sets. Some studies (3/13, 23.1%) used data from public databases, such as the MIMIC database, the public NIH Chest-XRay14, and PLCO data sets (Multimedia Appendix 1).

The top 3 models used were neural network (n=4), tree (n=3), and random forest (n=3) models. Support vector machine models were used twice (n=2). Other models, such as logistic regression, and linear regression were only used in 1 study each.

Model performance was mainly evaluated with sensitivity (n=7), specificity (n=8), area under the receiver operating characteristic curve (n=3), and accuracy (n=3), whereas other evaluation methods such as equal error rates, F1 score, recall, and κ coefficients were each used only once.

Quinn et al [13] provided a general model for inferring hidden factors from clinical data and was successfully applied to the major task of monitoring premature infants in the intensive care unit. Eshelman et al [12] described an algorithm consisting of a set of rules for identifying intensive care unit patients who may become hemodynamically unstable. Taking into account the individual differences of intensive care unit patients, Zhang and Szolovits [15] developed an algorithm based on personalized vital signs data to improve the accuracy of alarms. Charbonnier [14] extracted online temporal episodes from the high-frequency physiological parameters of intensive care unit patients to visually support signal interpretation.

Mechanical ventilation is widely used in intensive care units, during which a series of parameters need to be monitored. Kwok et al [16] established a nonlinear adaptive neuro-fuzzy inference system model for fractional inspired oxygen estimation, which reduced the need for invasive inspections. Two groups of researchers discussed the problem of patient-ventilator asynchrony, and developed a classifier based on machine learning to detect abnormal waveforms [17,18].

EEG monitoring plays an important role in the detection of brain function and the diagnosis of brain disease. Koolen et al [19] developed a method for the automated classification of neonatal sleep states via EEG. Golmohammadi et al [21] presented a system that can achieve high-performance classification of EEG events that might correlate with epilepsy, metabolic encephalopathy, cerebral hypoxia, and ischemia. Farzaneh et al [20] developed a machine learning framework to automatically segment and assess the severity of patients with subdural hematoma during traumatic brain injuries [20].

Infections are an important clinical issue in intensive care. Sepsis is a common and serious condition in the intensive care unit that results from an overreaction to infection that damages tissues and organs and can lead to complications, making it one of the leading causes of hospital-related deaths [24]. A high-performance algorithm, InSight, was demonstrated to be superior to the commonly used Modified Early Warning Score, Simplified Acute Physiology Score, and Systemic Inflammatory Response Syndrome score for the diagnosis of patients with alcohol use disorder combined with sepsis shock [23]. In addition, it is still challenging to explain lung opacity in radiography of the supine chest of patients with lung infection in the intensive care unit—Rueckel et al [11] evaluated a prototype artificial intelligence algorithm that could classify underlying lung opacity, which might suggest a diagnosis pneumonia.

Clinical event prediction, the use of data from electronic health records to predict the occurrence of certain events or the best time to give treatment, is one of the most important aspects of intensive care unit clinical decision support system. Among 32 clinical event prediction studies, 3 (9.4%) were related to acute kidney injury, 11 (34.4%) were related to infection prediction, 8 (25%) were related to respiratory diseases, and 10 (31.3%) were related to other predictions and evaluations (Multimedia Appendix 1).

In intensive care unit clinical prediction and evaluation studies, up to 87 variables were used in a single paper. Categories of variables, in order of frequency, were laboratory values (n=25), demographic variables (n=25), vital signs (n=20), scores (n=18), ventilation parameters (n=14), fluid balance (n=8), medications (n=7), comorbidities (n=7), outcome (n=4), nonmedicine therapy (n=3), symptoms (n=3), and medical history (n=1).

More than three-quarters of the studies (25/32, 78%) were based on data from a single center, 10 of which were from the freely available public database Medical Information Mart for Intensive Care II or III. Multi-institutional data were used in the other studies (7/32, 22%).

Logistic regression was the most commonly used method (11/32, 34%), followed by neural networks (7/32, 21%), and random forest (6, 19%). Support vector machine and decision tree models were each used in 5 (15.6%) studies. Naive Bayes, gradient boosting tree model, extreme gradient boosting, fuzzy model, and Insight each appeared twice (6.3%).

Sensitivity (n=16) and area under receiver operating characteristic curve (n=17) were the most commonly used evaluation metrics, followed by specificity (n=12) and accuracy (n=12). The following metrics appeared in fewer than 10 papers: positive predictive value (n=3), F1 score (n=4), and mean absolute error (n=2).

Early prediction of acute kidney injury has a high value for the long-term survival and quality of life of critically ill patients. Acute kidney injury is often associated with high morbidity and mortality rates in intensive care units. The status of other vital organs, initiation of therapy, patient response, and preexisting comorbidities can all contribute to the development of acute kidney injury [25]. Multiple machine learning methods have been utilized and compared to analyze unstructured clinical records and structured physiological measurements to identify early episodes of acute kidney injury [26]. Soliman et al [25] studied the prognostic impact of early acute kidney injury predicted by data from the first day of admission. One study [27] focused on patients younger than 21 years, who are more likely to recover from disease.

Early identification and treatment is the key to survival for many sepsis and infection patients [28], but it is difficult for clinicians to predict before it occurs, because it is extremely complex and each patient is different. Early prediction of sepsis using interpretable or uninterpretable machine learning models can help clinicians enhance the accuracy of fever workup [28] to identify and intervene in a timely manner [29-33]. One research aim is to make accurate predictions with as little electronic health record data as possible [34]. Mao et al achieved early prediction of sepsis using only vital signs validated in multiple centers [35]. The prediction of neonatal sepsis has also received substantial research attention in recent years [36,37]. One paper [38] focuses on predicting infections caused by a specific microorganism—invasive fungal disease due to Candida species—in intensive care unit patients.

Respiratory management in the intensive care unit is an important aspect of critical care and treatment. Early diagnosis of respiratory critical illness has a significant impact on patient prognosis [39]. In addition, maintenance of cardiopulmonary function is required in patients admitted to the intensive care unit due to acute symptoms such as direct trauma, pulmonary infection, heart failure, and sepsis. Machine learning methods can help predict the onset of acute respiratory disease in patients, especially in pediatric patients. Sauthier et al [40] used random forest and logistic regression to predict the time of acute hypoxic respiratory failure in critically ill children with severe influenza. Messinger et al [39] applied a cascaded artificial neural network to design new respiratory scores for early identification of asthma in young children. In addition, early prediction of acute respiratory distress syndrome was studied because of its high morbidity and mortality [41].

Furthermore, ventilator weaning and reintubation after weaning are currently well studied [42,43] in intensive care unit clinical decision support system literature, as well as the effect of drugs on intubation [44]. Moreover, predicting patient oxygen saturation after ventilation [45] and risk factors for failure of mechanical ventilation [46] can help health care professionals respond in a time manner.

There were 10 papers that could not be classified; we simply put them into one class separately. There were forecasts for detection and monitoring indicators, such as urine output after fluid administration [47], glucose [48], lactic acid [49], and activated partial thromboplastin time [50]. Lin [47] established a gradient tree-based machine learning model implemented with extreme gradient boosting algorithms to predict urine output in sepsis patients after fluid resuscitation to prevent fluid overload-related complications. Pappada et al [48,49] developed a neural network–based model to obtain a complete trajectory of glucose values up to 135 minutes in advance. Mamandipoor et al [49] combined least absolute shrinkage and selection operator regression, random forest, and long short-term memory to predict blood lactate concentration in patients in the intensive care unit. Our previous study also compared multiple machine learning approaches to guide clinical heparin administration by predicting the range of activated partial thromboplastin time values [50]. There were also studies that aimed to reduce unnecessary laboratory tests to streamline the process and reduce the burden on patients [51,52]. Predicted clinical events also included acute traumatic coagulopathy [53], delirium [54], advanced anemia [55], and fluid resuscitation therapy [56].

Of 46 papers that used machine learning for outcome evaluation for patients who were critically ill, 11 papers (23.9%) predicted overall mortality and survival, 23 papers (50%) predicted the outcomes of patients with certain diseases, and 12 papers (26.1%) included treatment prognosis, length of stay in the intensive care unit, and other outcome evaluations (Multimedia Appendix 1).

Categories of variables, in order of frequency, were demographic variables (n=39), scores (n=24), laboratory values (n=23), ventilation parameters (n=20), vital signs (n=18), comorbidities (n=17), medications (n=10), outcome (n=8), nonmedicine therapy (n=7), fluid balance (n=4), symptoms (n=4), and medical history (n=3).

Of the 46 outcome prediction studies, 25 (54.3%) were based on single-center data, 6 of which used data from MIMIC II and III, and the other 21 studies (45.7%) made use of multicenter data.

Logistic regression was the most commonly used method (27/46, 59%), followed by random forest (9/46, 20%), random forest (8/46, 17%), support vector machine (7/46, 15.2%) and decision tree model (5/46, 11%) studies. The gradient boosting tree model appeared in 4 (9%) studies, and adaptive boosting and linear regression each appeared twice (4.3%). Other models that appeared only once are not discussed here.

Area under receiver operating characteristic curve (n=37) was the evaluation metric used most often, followed by sensitivity (n=14), specificity (n=11), positive predictive value (n=4), accuracy (n=8), negative predictive value (n=6),F1 score (n=2), Matthews correlation coefficient (n=2), and Brier score (n=2).

Typical outcomes were overall mortality [57-62], survival [63], and long-term quality of life [64]. Mortality [65,66] and survival status at 1 year [67] in critically ill patients aged 80 years and older were also studied using machine learning methods.

Patients with sepsis and infection remain one of the most studied populations in terms of mortality (generally 28 days) [68-72], followed by acute kidney injury [72-75]. There is an increasing trend in outcome prediction studies in critically ill patients with liver disease—acute liver injury [76,77], cirrhosis [77], and advanced liver disease [78] have been studied using machine learning. In patients with severe cancer, 30- [79] and 120-day [80] survival rates were studied retrospectively with logistic regression models.

For cardiac disease, Lee et al [81] used EEG data to predict the outcome of children with cardiac arrest and Murtuza et al [82] found that arterial blood lactate levels can be associated with mortality in children who have undergone cardiac surgery. For brain diseases, the outcomes of patients with subarachnoid hemorrhage [83] and severe traumatic brain injury [84] have been analyzed. Wildman et al [85] predicted the impact of chronic obstructive pulmonary disease and asthma on mortality in critically ill patients. Daly et al [86] used logistic regression to study the relationship between early discharge and mortality with the intention of reducing mortality in this group of intensive care unit patients. Other papers [87-89] examined patient outcomes and factors influencing them after deterioration. Ebadollahi et al [90] predicted the temporal trajectory of physiological data with patient similarity, with the aim to identify universal patterns of disease progression from a large amount of clinical practice data, to establish a generalized computer-aided clinical decision support framework for personalized treatment.

Evaluating the outcome of certain treatments through machine learning can help medical professionals refine their treatments to achieve better therapeutic effects. Evaluation of outcomes after extubation based on continuous vital sign information and static characteristics of children can help adjust the timing of extubation to reduce mortality [91-93]. Evaluation of prolonged mechanical ventilation [94] and 1-year and 5-year functional survival [95] after cardiac surgery was used to help adjust and optimize postsurgical care practices. Evaluating the length of stay in the intensive care unit [96,97] and the risk of readmission after discharge from the intensive care unit [98] to effectively forecast the trend of the disease could improve treatment and care. In addition, designing and improving critical illness scores to indicate disease severity [99-101] was studied. For example, McRae et al [102] designed a score to quickly determine the severity of COVID-19 and achieved optimistic results in 160 individuals.