AI fares better than doctors at predicting deadly complications after surgery (hub.jhu.edu)

Johns Hopkins researchers announced a deep-learning approach that extracts previously hidden signals from routine preoperative electrocardiograms (ECGs) to predict 30‑day postoperative major complications (heart attack, stroke, or death). Using ECGs from 37,000 surgical patients at Beth Israel Deaconess, they trained two models: an ECG‑only network that already beat conventional clinical risk scores (≈60% accuracy), and a "fusion" model that combined ECG waveforms with basic EHR data (age, sex, comorbidities) to reach about 85% accuracy. The work, published in the British Journal of Anaesthesia, suggests a 10‑second, inexpensive test can supply strong prognostic information not visible to clinicians. Technically, the study applies deep learning to waveform morphology to capture systemic physiological signals (inflammation, metabolic or fluid status, electrolytes) encoded in the ECG, and includes an explainability method to highlight ECG features linked to adverse outcomes. If validated externally and prospectively, this tool could transform preoperative risk stratification, patient counseling, and surgical decision-making by augmenting or replacing less accurate risk scores and enabling earlier interventions or monitoring for high‑risk patients. Next steps are broader dataset testing and prospective clinical trials.