Joao C Ferreira, Luis B Elvas
Logistic Regression was the best-performing model. Under the IQR-based definition: AUROC 0.78, recall 85.4%, precision 66.1%, F1-score 74.5%. Under the clinically standard threshold-based definition: AUROC 0.81, recall 91.7%, precision 68.8%, F1-score 78.6%. Expanded metrics (event prevalence, false-positive rate, alert burden, calibration, confidence intervals) are reported. A prototype real-time e-mail alert system was demonstrated as a feasibility proof-of-concept and reviewed informally by two cardiovascular specialists.
INTRODUCTION: Cardiovascular diseases (CVD) are a leading global health challenge, particularly in resource-constrained settings where delayed diagnosis worsens outcomes. Abnormal heart rate (tachycardia/bradycardia) is a modifiable, independent risk factor that benefits from early, automated detection.
METHODS: We used a large real-world dataset from Hospital Santa Maria, Lisbon (23,122 patients, January 2019-November 2021, spanning the COVID-19 period). Two operational outcome definitions were evaluated: (i) an interquartile-range (IQR)-based, patient-specific definition, used to compare four classifiers (Logistic Regression, K-Nearest Neighbors, Random Forest, Naive Bayes), and (ii) standard clinical thresholds for tachycardia/bradycardia, applied to the best-performing model. Following CRISP-DM, models were trained and evaluated with patient-grouped, walk-forward time-series cross-validation to prevent patient-level data leakage.
RESULTS: Logistic Regression was the best-performing model. Under the IQR-based definition: AUROC 0.78, recall 85.4%, precision 66.1%, F1-score 74.5%. Under the clinically standard threshold-based definition: AUROC 0.81, recall 91.7%, precision 68.8%, F1-score 78.6%. Expanded metrics (event prevalence, false-positive rate, alert burden, calibration, confidence intervals) are reported. A prototype real-time e-mail alert system was demonstrated as a feasibility proof-of-concept and reviewed informally by two cardiovascular specialists.
DISCUSSION: Recall and AUROC alone do not establish clinical utility. Key limitations include the single-centre design, substantial missingness in ethnicity/diagnosis variables, the small clinician-evaluation sample (n = 2), and pandemic-period data collection. This work is presented as a methodologically strengthened proof-of-concept, with prospective, multi-site, multi-clinician external validation identified as the necessary next step.