Personalized Predictions of Therapeutic Hypothermia Outcomes in Cardiac Arrest Patients with Shockable Rhythms Using Explainable Machine Learning

• Published in: Diagnostics (Basel) Vol. 15; no. 3; p. 267
• Main Authors: Hong, Chien-Tai, Bamodu, Oluwaseun Adebayo, Chiu, Hung-Wen, Chiu, Wei-Ting, Chan, Lung, Chung, Chen-Chih
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 23.01.2025; MDPI
• Subjects: artificial neural network; Cardiac arrest; Cardiac arrhythmia; Coma; Cooling; Drug dosages; Hypothermia; Machine learning; Medical prognosis; Medical records; Patients; Performance evaluation; Shapley Additive exPlanations; shockable rhythms; Support vector machines; therapeutic hypothermia; Variables; Taiwan; artificial neural network; clinical outcome; Shapley Additive exPlanations; machine learning; therapeutic hypothermia; cardiac arrest; shockable rhythms
• ISSN: 2075-4418; 2075-4418
• DOI: 10.3390/diagnostics15030267

Background: Therapeutic hypothermia (TH) represents a critical therapeutic intervention for patients with cardiac arrest, although treatment efficacy and prognostic factors may vary between individuals. Precise, personalized outcome predictions can empower better clinical decisions. Methods: In this multi-center retrospective cohort study involving nine medical centers in Taiwan, we developed machine learning algorithms to predict neurological outcomes in patients who experienced cardiac arrest with shockable rhythms and underwent TH. The study cohort comprised 209 patients treated between January 2014 and September 2019. The models were trained on patients’ pre-treatment characteristics collected during this study period. The optimal artificial neural network (ANN) model was interpretable using the SHapley Additive exPlanations (SHAP) method. Results: Among the 209 enrolled patients, 79 (37.80%) demonstrated favorable neurological outcomes at discharge. The ANN model achieved an area under the curve value of 0.9089 (accuracy = 0.8330, precision = 0.7984, recall = 0.7492, specificity = 0.8846) for outcome prediction. SHAP analysis identified vital predictive features, including the dose of epinephrine during resuscitation, diabetes status, body temperature at return of spontaneous circulation (ROSC), whether the cardiac arrest was witnessed, and diastolic blood pressure at ROSC. Using real-life case examples, we demonstrated how the ANN model provides personalized prognostic predictions tailored to individuals’ distinct profiles. Conclusion: Our machine learning approach delivers personalized forecasts of TH outcomes in cardiac arrest patients with shockable rhythms. By accounting for each patient’s unique health history and cardiac arrest event details, the ANN model empowers more precise risk stratification, tailoring clinical decision-making regarding TH prognostication and optimizing personalized treatment planning.