Cardiovascular Risk Assessment Using Artificial Intelligence-Enabled Event Adjudication and Hematologic Predictors

Area of Science:

• Biostatistics
• Cardiovascular Medicine
• Clinical Pathology

Background:

• Clinical risk models often require novel biomarkers, balancing cost, availability, and deployment ease.
• Ideal population health risk assessment tools utilize widely available patient data.
• Hematologic markers from complete blood counts (CBCs) are common outpatient measures.

Purpose of the Study:

• To evaluate the utility of common hematologic markers in developing risk models for cardiovascular events.
• To assess if hematologic indices can enhance existing risk prediction models based on demographics and diagnostic codes.

Main Methods:

• Developed Cox proportional hazards models to predict major adverse cardiovascular events and all-cause mortality.
• Utilized 10 hematologic indices from routine CBCs, demographic data, and diagnostic codes as predictors.
• Employed automated event adjudication of discharge summaries for outcome ascertainment and validated models on an external cohort.

Main Results:

• Hematologic markers alone achieved a concordance index (CI) of 0.60-0.80 for predicting outcomes, with best performance in heart failure and all-cause mortality.
• Models incorporating hematologic indices demonstrated improved discrimination and calibration compared to those using only demographic and diagnostic data (CI improvement up to 0.072).
• Random survival forests offered minimal additional benefit over Cox proportional hazards models.

Conclusions:

• Low-cost, widely available hematologic markers can serve as valuable inputs for cardiovascular risk prediction models.
• Biologically informative, ubiquitous laboratory data can provide effective population-level risk stratification.
• Routine blood tests offer a feasible and effective strategy for enhancing cardiovascular risk assessment in population health initiatives.