Artificial Intelligence in Cardiac Critical Care: Current Insights and Future Prospects

Disease identification

Identifying the criticality of a disease from an extensive list of differential diagnoses can become easier with the help of AI. For example, pulmonary edema secondary to heart failure can be discriminated from pulmonary pathology using an ML model^[6], and moreover, the amount of edema in the lungs can also be quantified using a semi-supervised variational autoencoder.^[7] Similarly, during the COVID-19 pandemic, severe acute respiratory syndrome (SARS) was detected by processing imaging data from patients using an AI model.^[8]

Disease evolution and risk prediction

In the critical care unit, expeditious clinical deterioration is not uncommon and may be irreversible and life-threatening if diagnosed late. Thus, early prediction and stratification of hemodynamic decompensation are necessary for timely action. Tachycardia is an early sign of observed disparity from the normal range before shock, and it can be predicted 75 min before shock using a random forest model-based normalized dynamic risk score trajectory.^[9]

AI-driven models can forecast the evolution of SARS due to COVID-19 by utilizing various clinical parameters and imaging modalities.^[10] Even cardiopulmonary arrest can be predicted with an electronic Cardiac Arrest Risk Triage score, which was shown to be non-inferior as compared to traditional scoring systems.^[11] Even the Weibull-Cox proportional hazards model, on the basis of multifarious clinical databases and collective vital parameters, can predict sepsis.^[12]

Disease heterogeneity and decision-making

Cardiovascular criticality is complex and rarely has typical manifestation. Rather it has inherent heterogeneity, which afterward can complicate the underlying disease course and may incite multiorgan dysfunction. Furthermore, the composite critical state transits so rapidly that clinicians cannot make decisions on the basis of recent investigations. AI could delineate heterogeneity and criticality and, therefore, trigger individualized management in accordance with current guidelines.

A parsimonious model was developed and validated for latent stage analysis of acute respiratory distress syndrome (ARDS) using various classifier variables (interleukin-6, 8, tumor necrosis factor receptor, protein C, bicarbonate levels, and vasopressors requirement).^[13] This ML-based ARDS classification has extended and integrated the knowledge for assessing and managing such complex situations. Furthermore, for patients on ventilatory support, timed data with almost 44 characteristics were retrieved with the help of the reinforcement learning model (Markov Decision Process) to provide better patient outcomes as compared to clinician conventional care with regard to 90-day and intensive care unit (ICU) mortality.^[14] AI helps in each step of diagnosis and treatment of sepsis, such as early detection by predictive analytics, AI-driven real-time monitoring, risk stratification, personalized treatment plans, antibiotics stewardship, fluid responsiveness, outcome prediction, and integrated electronic health records. Applying reinforcement learning-based algorithms for sepsis management, AI can provide timely optimal decision-making and generate appropriate individualized solutions.^[15]

Coronary artery disease (CAD)

Acute coronary syndrome (ACS) is the most common type of cardiac illness and the major cause of mortality nowadays. ML-based algorithms play a pivotal role in the prompt diagnosis, risk stratification, disease prognostication, and treatment protocol for ACS. Recently, a CNN algorithm using 40 variables such as ECG, biomarkers, and history was validated to predict the requirement of angiography in ACS presenting to the emergency department.^[16] Similarly, another CNN developed from 42 demographic and clinical variables that predict non-ST elevation myocardial infarct.^[17] AI algorithms using the support vector machine (SVM) model can also derive short-term as well as long-term in hospital mortality in patients with ACS.^[18] A random forest ML-based technique was introduced to measure coronary arterial calcium, which can be derived automatically from computed tomography (CT) imaging. This algorithm has a better correlation coefficient and importantly, shortened the processing times in comparison to manual interpretation.^[19] Another novel study used a boosted ensemble ML algorithm (LogitBoost)-based technique to evaluate myocardial perfusion single-photon emission CT scan of suspected and low likelihood of CAD for risk stratification.^[20]

Valvular heart disease

Valvular pathologies are well diagnosed by their characteristic murmur, and therefore, AI-based algorithms were incorporated into phonocardiograms for clinical decision-making. SVM classifier was developed to classify abnormal heart sounds, which is an AI technology.^[21] AI algorithms can even identify minimal discrimination and pattern in pixel density and thereby, widely applied to echocardiography. An ML algorithm was introduced in patients with aortic stenosis undergoing transcatheter aortic valve replacement (TAVR) to measure the aortic valve annulus. Furthermore, in hospital mortality after TAVR can be predicted using artificial neural network (ANN), logistic regression technique, and random forest ML algorithms.^[22] Similarly, another ML-based algorithm was developed for measuring the mitral annular dimensions, and it is significantly less time consuming.^[23]

Electrophysiology

Arrhythmias are common with underlying cardiac pathology, especially atrial fibrillation, which is the most prevalent arrhythmia. Nowadays, AI adds new insights for diagnosis and risk stratification to the key diagnostic tool ECG. Several recent trials incorporate ANN on single, 3, and 12 lead ECGs to accurately distinguish and diagnose arrhythmias. Hannun et al. and Raghunath et al. used deep neural networks (DNN) for the classification of ECG into rhythm classes and to predict all-cause mortality.^[24,25] ML algorithms can be used to detect electrolyte derangements and predict the future incidence of arrhythmias.

Cardiac failure

Cardiac failure is a complex entity where cardiac output does not meet metabolic demand. AI-based predictive modeling and automated cardiac function measurements reinforce diagnostic capacity and decision-making as compared to conventional echocardiography. Moreover, advanced AI like “HeartModel A.I.” uses AI segmentation and 3D echocardiographic algorithm to quantify cardiac volumes automatically and it was validated by Medvedofsky et al.^[26] Son et al. and Kang et al., developed various AI-based algorithms, SVM-based models, and ANN in patients with HF for risk assessment, medication adherence, predicting morbidity and mortality, and repeated hospitalizations.^[27,28] The LINK-HF study and Akbilgic et al., developed a CNN-based ECG-AI model were examples of the advancement of AI technology in the management of heart failure.^[29,30]