<p>Zhao et al. present machine-learning models to predict intraoperative hemodynamic instability in hypertensive pheochromocytoma and paraganglioma surgery. The clinical motivation is sound and the reported discrimination and decision-curve metrics indicate promise. Yet the work illustrates two recurring challenges for clinical AI. First, models trained and tested within a single institution risk overestimation without external validation that reflects the intended deployment. Second, calibration and timing of prediction must be explicit to avoid information leakage and to ensure actionable probabilities. Attention to these methodological points will enhance interpretability and safe translation into peri-operative practice.</p>