Comment on “Multimodal deep learning to predict postoperative major adverse cardiac and cerebrovascular events after noncardiac surgery”

To the Editor, Yoon et al present a pragmatic multimodal deep-learning approach to predict 30-day major adverse cardiovascular and cerebrovascular events (MACCEs) after noncardiac surgery, combining raw preoperative 12-lead ECG waveforms with only age, sex, and simplified ICD-10 procedure codes[1]. This low-burden input design directly addresses a key barrier to perioperative AI adoption – models that rely on extensive chart-derived features often fail to translate across workflows. This commentary was prepared in line with the TITAN 2025 guideline to promote transparent and responsible reporting[2]. The scale and validation strategy are notable (165 577 cases; MACCE 0.6%), including cross-validation and temporal validation[1]. The multimodal gradient-boosting model achieved excellent discrimination (AUROC 0.902) and outperformed a clinical-only baseline and conventional comparators (RCRI and ASA class)[1,3]. Importantly for rare-event prediction, improved AUPRC suggests gains beyond AUROC inflation in imbalanced outcomes[1]. Clinically, the work is compelling because ECGs are ubiquitous and inexpensive, yet their prognostic information is underused in perioperative decision-making. The finding that ECG-derived latent features add substantial signal echoes broader evidence that deep learning can extract clinically meaningful patterns from routine 12-lead ECGs at scale[4]. In parallel, contemporary perioperative guidance emphasizes structured cardiovascular risk assessment and targeted testing – an area where reliable, workflow-embedded prediction tools could reduce both missed risk and unnecessary investigations[5]. Several issues deserve emphasis before deployment. First, clinical utility will hinge on calibration and actionable thresholds aligned to care pathways (e.g., step-up monitoring, biomarker surveillance, cardiology review), not discrimination alone; prospective “decision impact” evaluation is therefore essential[1,5]. Second, the study is single-center without external validation; ECG acquisition, device vendors, coding practices, and perioperative management vary widely across institutions. Third, outcome ascertainment depends on EHR capture, which may miss events occurring outside the system – particularly relevant for 30-day cerebrovascular outcomes[1]. Overall, Yoon et al. advance scalable perioperative risk stratification by showing that raw ECG waveforms can materially enhance MACCE prediction using minimal additional inputs[1]. The next milestone is multicenter external validation and prospective implementation studies demonstrating that acting on model outputs improves patient-important outcomes without unintended over-intervention.