Using EHRs and Machine Learning for Heart Failure Survival Analysis.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Heart failure (HF) presents a significant global health challenge with high morbidity, mortality, and healthcare costs.
• Existing risk prediction models, like the Seattle Heart Failure Model (SHFM), often rely on clinical trial data and may not fully represent routine community care.
• Incorporating contemporary therapies into models like SHFM has historically involved extrapolation from external trials.

Purpose of the Study:

• To evaluate the performance of the Seattle Heart Failure Model (SHFM) using electronic health records (EHRs) from Mayo Clinic.
• To develop novel machine learning-based risk prediction models utilizing routine clinical care data for heart failure (HF).
• To assess the impact of incorporating new predictive markers, such as comorbidities, on prognostic performance.

Main Methods:

• Utilized patient-specific characteristics from Electronic Health Records (EHRs) at Mayo Clinic.
• Developed and assessed machine learning models for heart failure (HF) survival risk prediction.
• Compared the performance of EHR-based models against traditional models, evaluating improvements in Area Under the Curve (AUC).

Main Results:

• Models developed using EHR data demonstrated an 11% improvement in AUC compared to traditional methods.
• EHR-based models showed greater accuracy and applicability in routine clinical settings.
• Inclusion of additional predictive markers, such as comorbidities, led to an 8% improvement in prognostic performance.

Conclusions:

• Machine learning models leveraging EHR data provide a more accurate and practical approach to heart failure (HF) survival risk prediction.
• The integration of comorbidities and other routine clinical data enhances the prognostic capabilities of HF risk models.
• These findings support the use of EHR data and advanced analytics for personalized HF management and improved patient outcomes.