Time-aware machine learning for prospective tachycardia prediction in cardiovascular disease

Tachycardia can accompany acute deterioration in hospitalized patients with cardiovascular disease, yet most machine-learning studies emphasize detection rather than prospective prediction. Using routinely collected physiological monitoring data from a hospital-wide database of 512,764 patients, we identified 968 patients with myocardial infarction (MI; n = 260), pulmonary thromboembolism (PTE; n = 48), or aortic stenosis (AS; n = 660) and developed models for prospective tachycardia prediction. Tachycardia was defined as heart rate >100 bpm. Historical physiological features were separated temporally from the future target, population-level autocorrelation analysis informed lag selection, and Random Forest (RF), Naïve Bayes (NB), and a feedforward neural network (NN) were evaluated using patient-blocked, time-aware validation. At the selected 90-step lag configuration, RF showed the strongest performance, including an F1-score of 0.89 ± 0.02, PR-AUC of 0.72 ± 0.03, and AUROC of 0.93 ± 0.02 for PTE. These findings provide internal feasibility evidence for time-aware tachycardia risk prediction from routine hospital monitoring data; external and prospective validation is required before clinical deployment.