Critical care medicine is an ideal area for the use of artificial intelligence (AI) and machine learning (ML) (Crit Care 2022;26:75). Rounds in the ICU frequently include debates on diagnosis, treatment, prognosis, and the likelihood of patient deterioration. Critical care patients generate massive amounts of data, including vital signs, laboratory results, medication infusion rates, and ventilatory parameters. Continuous waveforms from arterial, central, and pulmonary artery catheters, ventilator waveforms of pressure, flow, and volume versus time, and EEG recordings require data acquisition at frequencies as high as 500 Hz. Imaging studies (X-rays, CT scans, MRI, and ultrasound examination) contain complex data far beyond what is described in a typical radiology report. Critically ill patients continuously change over time, often in response to interventions, and this evolution has implications for diagnosis, treatment, and prognosis. As a result, data must be acquired constantly throughout the ICU stay. AI systems can utilize natural language processing to include data such as progress notes from physicians, nurses, and allied health personnel. Patients in the ICU are heterogenous in diagnoses, comorbidities, disease presentation, and response to therapy, so guidelines and algorithms are often not relevant to an individual patient. Data in the ICU often have complex nonlinear relationships to patient diagnosis and outcome. The complexity of the above factors is challenging for individuals but is ideal for AI/ML approaches, which can identify complex interactions among diverse sources of data (N Engl J Med 2023;388:1201-8). There are multiple existing, large, detailed databases, such as MIMIC-III, that allow AI/ML models to be developed and then verified in clinical practice (Healthcare 2023;11:710).

Critical care needs AI/ML. Critical care accounts for almost half of all hospital costs and a large percentage of morbidity and mortality among hospitalized patients. Critical care outcomes impact hospital length of stay, discharge to home versus facility, return to work, and long-term health care costs. Patients in the ICU can rapidly deteriorate, either due to their underlying disease or to acute changes such as new sepsis or a pneumothorax. Critical care providers must have extensive knowledge and experience across a broad range of topics and rapidly intervene at the first sign of patient deterioration. The heterogeneity of critical care patients makes accurate diagnosis and prognosis difficult for even experienced clinicians, especially when combined with fatigue, high workload, and burnout. AI/ML models hold the promise of providing improved diagnosis and prognosis with continuous vigilance and no fatigue. Applications of AI to critical care have included diagnosis, phenotyping, prediction and prognosis, clinical decision support, and automated treatment.

Multiple diseases can have similar presentations in critically ill patients or can present in unusual ways. ICU rounds often debate the underlying diagnosis, such as deciding whether pulmonary edema is cardiogenic or noncardiogenic or whether a patient does or does not have sepsis. A correct diagnosis is essential to providing the correct therapy. AI/ML can integrate multidimensional data to determine the correct diagnosis with a high degree of accuracy. The models often use data that may not be obvious, such as diagnosing left ventricular dysfunction by the electrocardiogram. When patients have clinical deterioration, AI/ML can determine the reason despite an extensive list of possible etiologies. A major role for AI in critical care is the identification of specific phenotypes that share a similar clinical presentation but may have different underlying mechanisms and differential responses to a specific therapy. This area has been extensively explored in ARDS and in severe sepsis, identifying hypoinflammatory and hyperinflammatory subtypes of ARDS and four phenotypes in sepsis (Lancet Respir Med 2022;10:367-77; JAMA 2019;321:2003-17).

Accurate prognosis allows the correct choice of treatment and its intensity. For example, patients who otherwise will develop shock can be started on fluids and vasopressors, patients who will develop sepsis can be started on antibiotic therapy, patients who will develop respiratory failure can undergo elective intubation and mechanical ventilation, and patients who will develop delirium can have targeted interventions. Multiple AI/ML models have demonstrated the ability to predict clinical deterioration many hours before it occurs (Crit Care Explor 2022;4:e0744). Current ICU scoring systems such as APACHE, MPM, and SOFA can predict average prognosis for group data but often are not accurate enough for individual patient decisions. Accurate prediction of clinical outcome can result in more effective goals of care discussions. AI/ML models have predicted the need for mechanical ventilation, the development of acute kidney injury, outcome from stroke, and ICU or hospital mortality (Adv Chronic Kidney Dis 2022;29:431-8; Anesthesiology 2022;137:586-601). In addition, the models continuously update prognosis in real time. Prognosis prediction may allow earlier discharge from the ICU for patients determined to be at low risk for deterioration and continued observation in the ICU for those deemed to be at high risk.

Critical care requires timely and accurate decisions based on integration of multiple sources of information. AI/ML can provide support to help the provider meet this challenge, including recommendations for adjusting ventilator settings and for fluid and vasopressor administration in sepsis (Minerva Anestesiol 2022;88:1066-72; Br J Anaesth 2022;128:343-51). Systems already exist for the next step, which is for AI/ML systems to make these changes independently, such as adjusting ventilatory parameters or drug infusion rates.

AI/ML has been used in neurocritical care where patients have diverse diagnoses and data can include hemodynamics, multimodal intracranial monitoring, EEG waveforms, and complex imaging. AI can determine the etiology of the neurologic dysfunction and the likelihood of response to different therapies (Crit Care Clin 2023;39:235-42). Applications of AI in neonatal and pediatric critical care are similar to those for adults and can improve diagnosis and predict shock, cardiac arrest, and mortality. During the pandemic, AI has been used for COVID screening, diagnosis, prognosis, and prediction of developing ARDS, mortality, and hospital length of stay.

A theme of this article is that critically ill patients are complex and generate massive amounts of data that evolve over time. Human decision-making is limited by complexity, by lack of knowledge, and by multiple biases. For example, health care professionals are reluctant to change diagnoses even when contradictory evidence develops. Even if “expert” clinicians are able to perform at the level of AI/ML, such providers are not available 24/7 and are subject to fatigue. In contrast, AI can be continuously vigilant and repeatedly evaluate the patient without bias.

The limitations of AI in critical care are similar to those in other applications (Crit Care 2022;26:75; Clin Exp Emerg Med 2023;10:132-7; J R Soc Med 2022;115:236-8). Studies have been hindered by the use of single and often narrow datasets, the number of patients studied, the completeness of the data, and limited use and timing issues with narrative data entered into the medical record. When accurate models have been developed on a given dataset, they often fail to be replicated in a larger study or are not generalizable to different institutions and patient populations. In addition, even effective models may not result in improved outcomes due to cultural barriers, workflow challenges, and alarm fatigue so that the AI recommendations are often not followed. The dataset itself may have biases, which are automatically incorporated into the AI model. Some may not accept AI recommendations because they are unable to understand how the AI model develops its conclusions. Although there have been advances in this area, such as in the use of SHAP (SHapley Additive exPlanations) plots and user-friendly graphic interfaces, models based on complex multidimensional data may not be understood by many health care professionals. Finally, AI applications in critical care have ethical concerns such as data privacy and patient autonomy.

At the current time, there are few AI/ML applications that have been effective outside of the institution in which they were developed and almost no studies demonstrating improved outcomes. However, the field of AI/ML is rapidly advancing, and many of the problems will be solved in the next few years. The Society of Critical Care Medicine and the European Society of Intensive Care Medicine are collaborating in a data science initiative to speed the development of AI/ML in critical care, and we will see an increasing number of effective applications in the future (Crit Care Med 2021;49:e563-77).