Health Sci Rep. 2026 Apr 2;9(4):e72070. doi: 10.1002/hsr2.72070. eCollection 2026 Apr.
ABSTRACT
BACKGROUND AND AIM: With the transition of the COVID-19 outbreak from a pandemic to an endemic state, people are now experiencing COVID-19. COVID-19 is associated with severe symptoms in infected populations with chronic diseases, and CVD patients are no exception to this rule. Early prediction of COVID-19 mortality risk using AI-based models is crucial for improving prognosis and survival rates by enabling more effective surveillance among CVD populations. This study aims to develop predictive models of COVID-19 mortality risk and to provide insights into this topic.
METHODS: This retrospective study used a single-center database of 1255 CVD participants admitted to Shariati Hospital, affiliated with Tehran University of Medical Sciences (TUMS) in Tehran City, from February 2020 to November 2021. Several demographics, patients' clinical histories, clinical features, and laboratory findings were leveraged and assessed using univariate and adjusted correlation analyses to train AI algorithms to predict the risk of COVID-19 death among CVD patients as the outcome variable. We leveraged the SHAP summary plot to introduce the AI's explainability (XAI).
RESULTS: Our empirical results indicated that XGB achieved a PPV of 0.931, an NPV of 0.926, a sensitivity of 0.917, a specificity of 0.939, an accuracy of 0.929, an F-score of 0.924, and an AUC of 0.864, outperforming other models. SHAP values indicated that age, pneumonia, ICU admission, type of surgery, and d-dimer were the most significant predictors.
CONCLUSION: The XGB model demonstrated greater potential to stratify at-risk CVD patients on admission, particularly for COVID-19 mortality, by better allocating clinical resources and improving the prognosis of COVID-19 patients with this chronic condition, thereby achieving greater predictive performance and clinical usability.
PMID:41948633 | PMC:PMC13051999 | DOI:10.1002/hsr2.72070