Clinical Feature-Based Machine Learning Model for 1-Year Mortality Risk Prediction of ST-Segment Elevation Myocardial Infarction in Patients with Hyperuricemia: A Retrospective Study

Zhixun Bai, Jing Lu, Ting Li, Yi Ma, Zhijiang Liu, Ranzun Zhao, Zhenglong Wang, Bei ShiComput. Math. Methods Medicine 2021 · DOI

TL;DR

This study developed and validated a machine learning model, specifically CatBoost, to predict 1-year mortality in patients with ST-segment elevation myocardial infarction (STEMI) and hyperuricemia. The CatBoost model, especially after optimization with a hybrid sampling technique (SMOTEENN), significantly outperformed the traditional GRACE risk score, offering a more accurate and robust tool for risk assessment in this high-risk patient population.

What Problem Does This Paper Address?

Accurate risk assessment is crucial for managing patients with ST-segment elevation myocardial infarction (STEMI), a severe form of heart attack. However, for STEMI patients who also have hyperuricemia (elevated uric acid levels), a well-established and practical method for predicting their long-term prognosis (specifically 1-year mortality) is lacking. Hyperuricemia is increasingly recognized as a cardiovascular risk factor, and its presence in STEMI patients is associated with poorer outcomes. Traditional risk scores, like the GRACE score, may not fully capture the complexity of this specific patient subgroup. This paper aims to bridge this gap by exploring the utility of various machine learning models for predicting 1-year mortality in STEMI patients with hyperuricemia.

Methods

The researchers employed a retrospective study design, analyzing data from 656 STEMI patients with hyperuricemia admitted to the Affiliated Hospital of Zunyi Medical University between January 2016 and January 2020. They compared the predictive performance of five machine learning models: logistic regression (LR), k-nearest neighbor (KNN), CatBoost, random forest (RF), and XGBoost. These models were evaluated against the traditional GRACE risk score.

A key methodological challenge addressed was the imbalanced nature of the dataset, where the number of patients who survived was significantly higher than those who died. To mitigate this, they utilized SMOTEENN, a hybrid sampling algorithm that combines synthetic minority oversampling technique (SMOTE) with edited nearest neighbor (ENN). This technique aims to oversample the minority class (death cases) by creating synthetic samples and then cleaning mislabeled instances, thereby improving the model's ability to learn from the rarer outcomes.

All models were validated using 10-fold cross-validation. Feature engineering involved one-hot encoding for categorical features and the missForest method for imputing missing values. Model performance was assessed using several metrics: accuracy, Area Under the Curve (AUC), recall, precision, and F1 value.

Datasets

The study utilized a dataset of 656 patients diagnosed with STEMI and hyperuricemia from the Affiliated Hospital of Zunyi Medical University. This dataset was collected between January 2016 and January 2020, with a follow-up period extending to January 2021. No new benchmark datasets were introduced; the study relied on existing patient data.

Key Insights and Ideas

  • Hyperuricemia as a significant factor: The study reinforces the notion that hyperuricemia is not just a comorbidity but a significant factor influencing the prognosis of STEMI patients, leading to higher mortality.
  • Machine learning's superiority: Machine learning models, particularly ensemble methods like CatBoost, demonstrate a clear advantage over traditional risk scores (GRACE) in predicting 1-year mortality for this specific patient group. This highlights the potential of ML to uncover complex, non-linear relationships in clinical data.
  • The power of hybrid sampling: The SMOTEENN technique proved crucial in improving the predictive performance of machine learning models, especially in the face of imbalanced datasets. This underscores the importance of addressing data imbalance for accurate risk prediction.
  • CatBoost's robustness: The CatBoost algorithm emerged as the top performer, showcasing its ability to handle complex data and achieve high predictive accuracy, even before optimization. Its effectiveness was further amplified by the SMOTEENN technique.

Key Results

  • CatBoost's superior performance: Before optimization, CatBoost achieved an accuracy of 0.89 and an AUC of 0.87. After applying the SMOTEENN technique, its performance dramatically improved, reaching an accuracy of 0.96 and an AUC of 0.99, with a precision of 0.95 and an F1 value of 0.97.
  • Outperforming GRACE: The traditional GRACE risk score achieved an accuracy of 0.84 and an AUC of 0.80, significantly lower than the optimized CatBoost model.
  • Ensemble benefits: While individual models showed promise, ensemble methods also demonstrated good performance, with the "All" ensemble achieving an accuracy of 0.90 and an AUC of 0.86.
  • Mortality rate: The study reported a 1-year mortality rate of 13.8% among the enrolled patients.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective design: The retrospective nature of the study introduces a potential risk of bias in data collection.
  • Single-center study: The findings are based on data from a single center in China, which may limit generalizability to other populations.
  • Imbalanced data challenges: Despite using SMOTEENN, the inherent imbalance in clinical samples means that even optimized models might not achieve perfect prediction.
  • Computer-generated samples: The hybrid sampling technique generates synthetic samples, which are not real patients, necessitating careful interpretation.

Future work could involve:

  • Larger, multi-center studies: Expanding the dataset to include more diverse patient populations and clinical settings.
  • Incorporating more data types: Integrating social factors, environmental parameters, and phenotypic information (e.g., genomic or proteomic data) to further enhance prognostic prediction.
  • External validation: Validating the developed models on independent datasets to confirm their robustness and generalizability.

Summary updated

BibTeX
@article{bai2021clinical,
  author = {Zhixun Bai and Jing Lu and Ting Li and Yi Ma and Zhijiang Liu and Ranzun Zhao and Zhenglong Wang and Bei Shi},
  title = {{Clinical Feature-Based Machine Learning Model for 1-Year Mortality Risk Prediction of ST-Segment Elevation Myocardial Infarction in Patients with Hyperuricemia: A Retrospective Study}},
  journal = {Comput. Math. Methods Medicine},
  year = {2021},
  doi = {10.1155/2021/7252280},
}
Figure