In a groundbreaking study published in Npj Digital Medicine, researchers have unveiled a new machine learning (ML) model that demonstrates exceptional accuracy in predicting cardiac arrest in intensive care unit (ICU) patients using electrocardiogram (ECG) data.
Unraveling the need for improved predictive models
Despite advancements in critical care medicine, the incidence of unexpected and sudden cardiac arrests remains a challenge for patients admitted to ICUs. The diverse reasons behind in-hospital cardiac arrest necessitate the development of accurate prediction models for real-time detection, enabling prompt and life-saving interventions such as cardiopulmonary resuscitation (CPR) and early defibrillation.
While earlier models relied on clinical features from electronic medical records (EMRs), limitations in collecting various EMR variables prompted a shift toward ML algorithms based on ECG data. Electrocardiogram-based prediction models ensure constant and real-time monitoring, offering a potential breakthrough for in-hospital cardiac arrest predictions.
The research team meticulously constructed a structured ECG dataset from 5,679 ICU stays involving 4,821 patients. Using electrocardiogram signals, they calculated a comprehensive set of 74 heart rate variability (HRV) measures, including time-domain, frequency-domain, and nonlinear measures. The BorutaShap algorithm selected 33 of these measures as input features for the prediction model.
Employing a light gradient boosting machine (LGBM) algorithm, the researchers optimized model hyperparameters using Bayesian optimization. The primary outcome was the occurrence of cardiac arrest within 0.5–24 hours, with secondary outcomes considered at various time intervals.
Impressive results and comparative analysis
The ML model achieved an outstanding area under the receiver operating curve (AUROC) of 0.881 and an area under the precision-recall curve (AUPRC) of 0.104 for the primary outcome. The model’s discrimination performance remained robust across secondary outcomes, showcasing its reliability in predicting cardiac arrest in ICU patients.
Comparing the ML model with a clinical parameter-based model, which utilized 43 features derived from vital signs, revealed a significantly higher AUROC (0.881 vs. 0.735, p < 0.001) for the study model using HRV measures. This highlights the superiority of the ECG-based ML model in predicting sudden cardiac arrest.
Unveiling key HRV measures and future implications
Feature importance analysis identified the top six measures, including the 20th percentile of the RR intervals (Prc20NN), triangular interpolation of the RR interval histogram (TINN), and the interquartile range of the RR intervals (IQRNN). These measures exhibited dynamic changes approximately six hours before a sudden cardiac arrest, offering valuable insights into the real-time evaluation of a patient’s cardiac condition.
The study’s results suggest that higher IALS (acceleration/deceleration segments) can be associated with compromised cardiac conditions, emphasizing the potential of the ML model in identifying patients at risk of sudden cardiac arrest.
One of the significant strengths of this study is the model’s high accessibility and transferability to diverse healthcare settings. Continuous ECG monitoring, a standard practice in ICU settings, makes the application of this predictive model straightforward in clinical practice.
While the study provides crucial real-time insights, the researchers acknowledge the need for further investigation to establish causality between HRV measures and cardiac arrest.
