Developing EHR-driven heart failure risk prediction models using CPXR(Log) with the probabilistic loss function.

Area of Science:

• Biomedical informatics
• Machine learning in healthcare
• Prognostic modeling

Background:

• Accurate survival prediction is crucial for effective patient management in healthcare.
• Electronic Health Records (EHRs) offer rich data for developing prognostic models but present challenges due to complexity and high dimensionality.
• Existing classification methods struggle with the intricacies of EHR data for survival prediction.

Purpose of the Study:

• To develop and validate prognostic risk models for predicting 1, 2, and 5-year survival in heart failure (HF) patients.
• To apply a novel classification algorithm, Contrast Pattern Aided Logistic Regression (CPXR(Log)), with a probabilistic loss function to EHR data.
• To assess the impact of incorporating patient comorbidities on model performance.

Main Methods:

• Utilized Contrast Pattern Aided Logistic Regression (CPXR(Log)) with a probabilistic loss function on Mayo Clinic EHR data.
• Developed prognostic models to predict 1, 2, and 5-year survival rates for heart failure patients.
• Incorporated patient comorbidities into the models to evaluate performance improvements.

Main Results:

• The CPXR(Log) model achieved an Area Under the Curve (AUC) of 0.94 and an accuracy of 0.91, outperforming previous prognostic models.
• Incorporating comorbidities improved CPXR(Log) model AUC by 15.9%.
• The proposed probabilistic loss function yielded a 1% AUC improvement over existing functions.

Conclusions:

• CPXR(Log) demonstrates superior performance in predicting heart failure survival using EHR data.
• Heart failure is a heterogeneous disease requiring subgroup-specific prediction models.
• The study highlights the potential and challenges of using EHR data for predictive modeling, advocating for subgroup-based approaches like CPXR(Log).