Scholarlink

Background <p>Hypertension serves as a prevalent health issue, particularly in South Asia, where it is also a risk factor and comorbidity that affects the prognosis of Acute Coronary Syndrome (ACS) by increasing the likelihood of complications such as heart failure and arrhythmias. Conventional tools for predicting ACS mortality risk, such as the TIMI and GRACE scores, often fail to adequately represent the Asian population and do not specifically address the needs of ACS patients with comorbid hypertension. Our study focuses on developing a tailored and interpretable risk prediction model to assess mortality rate of Asian ACS-hypertensive patients.</p> Methods <p>We used in-hospital data from the NCVD Registry, which includes ACS patients with hypertension from 2006 to 2019 and selected key factors such as demographics, medications, and clinical details using Sequential Backward Elimination. We then used various ML techniques and combined them with ensemble meta-learners to predict mortality. We measured the model’s accuracy using the area under the curve (AUC) and fine-tuned the best model with Platt Scaling. Our model was compared with the TIMI score to calculate the net reclassification index (NRI). Finally, we applied LIME and SHAP to understand feature importance and interpret model performance.</p> Results <p>The RF model with selected 28 features outperforms other prediction models with an AUC of 0.95. The calibrated RF model has raised Brier score (ACS: 0.046; STEMI: 0.051; NSTEMI: 0.041), and Chi-squared (ACS: 10.253; STEMI: 6.023; NSTEMI: 23.030). NRI verified our calibrated model’s better reclassification than the conventional TIMI risk score (improved accuracy by 15%-78%). The model’s robustness was proven by its greater AUC (0.841–0.898) than the TIMI risk score (0.641–0.83). SHAP and LIME analyses revealed that the Killip classification and peak CK levels significantly influenced mortality. Conversely, ACE medications were found to be more strongly associated with survival.</p> Conclusion <p>Our calibrated RF model demonstrated superior predictive accuracy in mortality risk for Asian hypertensive ACS patients, outperforming traditional TIMI scores. It efficiently identified key mortality predictors, which is further interpreted by SHAP and LIME, effectively underscore the significance of selected features for ACS mortality risk.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting mortality risk in hospitalized ACS patients with hypertensive comorbidity: an interpretable machine learning approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Sazzli Kasim,
  • Kiew Xue Ning,
  • Sorayya Malek,
  • Putri Nur Fatin Amir Rudin,
  • Chien Wen Koh,
  • Khairul Shafiq Ibrahim,
  • Alan Yean Yip Fong,
  • Wan Azman Wan Ahmad

摘要

Background

Hypertension serves as a prevalent health issue, particularly in South Asia, where it is also a risk factor and comorbidity that affects the prognosis of Acute Coronary Syndrome (ACS) by increasing the likelihood of complications such as heart failure and arrhythmias. Conventional tools for predicting ACS mortality risk, such as the TIMI and GRACE scores, often fail to adequately represent the Asian population and do not specifically address the needs of ACS patients with comorbid hypertension. Our study focuses on developing a tailored and interpretable risk prediction model to assess mortality rate of Asian ACS-hypertensive patients.

Methods

We used in-hospital data from the NCVD Registry, which includes ACS patients with hypertension from 2006 to 2019 and selected key factors such as demographics, medications, and clinical details using Sequential Backward Elimination. We then used various ML techniques and combined them with ensemble meta-learners to predict mortality. We measured the model’s accuracy using the area under the curve (AUC) and fine-tuned the best model with Platt Scaling. Our model was compared with the TIMI score to calculate the net reclassification index (NRI). Finally, we applied LIME and SHAP to understand feature importance and interpret model performance.

Results

The RF model with selected 28 features outperforms other prediction models with an AUC of 0.95. The calibrated RF model has raised Brier score (ACS: 0.046; STEMI: 0.051; NSTEMI: 0.041), and Chi-squared (ACS: 10.253; STEMI: 6.023; NSTEMI: 23.030). NRI verified our calibrated model’s better reclassification than the conventional TIMI risk score (improved accuracy by 15%-78%). The model’s robustness was proven by its greater AUC (0.841–0.898) than the TIMI risk score (0.641–0.83). SHAP and LIME analyses revealed that the Killip classification and peak CK levels significantly influenced mortality. Conversely, ACE medications were found to be more strongly associated with survival.

Conclusion

Our calibrated RF model demonstrated superior predictive accuracy in mortality risk for Asian hypertensive ACS patients, outperforming traditional TIMI scores. It efficiently identified key mortality predictors, which is further interpreted by SHAP and LIME, effectively underscore the significance of selected features for ACS mortality risk.