Missing Data Governance in Real-World Cardiovascular Risk Prediction: Effects on Machine Learning Performance and Interpretability

Abstract Background Machine learning has become central to data driven cardiovascular risk prediction, yet the empirical foundation of the field rests heavily on a small set of curated, open benchmark datasets that are clean, complete, and drawn from populations unlike those in which models are deployed. Reported performance therefore often reflects idealized conditions rather than the irregular, incomplete data of routine clinical practice. This study reframes the heart disease prediction problem around the data themselves: the reliability of a clinical prediction model is governed not only by the choice of algorithm but also by how the incompleteness of real-world data is handled. Methods An original dataset was assembled directly from the cardiology records of a private hospital in Denizli, Türkiye, covering patients seen between 2019 and 2024. The records were retrieved without any prior check of whether individual fields were populated, and nineteen parameters were selected with cardiology specialists. Eleven supervised learning algorithms were evaluated under four contrasting strategies for treating missing data: imputation, retention of missing values, hybrid handling with an explicit unknown category, and complete case analysis. Models were tuned by grid search with five-fold cross validation, assessed on a held-out test partition using accuracy, precision, recall, and F1-score, and benchmarked against a majority class classifier. Explainable artificial intelligence analysis was used to examine whether the pattern of missingness itself carries information. Results Data quality governance proved decisive. Imputation and the hybrid strategy preserved predictive capability and stability, whereas deleting every incomplete record degraded performance markedly across all algorithms. Boosting ensembles were the most resilient to incomplete data, while instance based and simple probabilistic methods were the most adversely affected. Interpretability analysis showed that an indicator of unrecorded blood pressure ranked among the most influential features, so that the absence of a measurement itself acted as an informative signal. Conclusions Appropriate handling of incomplete records preserves model reliability, whereas uncritical deletion erodes it. The results argue for a shift from marginal accuracy gains on familiar benchmarks toward locally grounded, transparent models that can be trusted in the clinical environments they are meant to serve.