A Genetic Neural Architecture Search Framework to Predict Biomarker Status in Cardiovascular Disease Patients during Pandemics

Background: The COVID-19 pandemic disrupted healthcare systems, especially impacting high-risk groups such as individuals with cardiovascular diseases (CVDs). CVD biomarkers such as LDL cholesterol and HbA1c are critical indicators of patient health, disease progression and treatment efficacy. Prior predictive models for CVD biomarkers during this pandemic predominantly explored traditional machine learning techniques, which had inadequate predictive capacity. Genetic Neural Architecture Search (GANAS), employs a genetic algorithm to iteratively evolve neural network architectures and search a large hyperparameter space to optimize model performance via simulated natural selection. Our work is the first to explore GANAS for CVD biomarker trajectory prediction.Objective: Explore GANAS to create a Feed Forward Neural Network Model for predicting LDL cholesterol and HbA1c biomarker status in CVD patients from retrospective Electronics Health Records (EHR) data during the COVID-19 pandemic.Approach: GANAS was employed to design FFNNs using data from 50,000 CVD patient records across two hospitals in Massachusetts, spanning January 2019 to June 2021. LDL Cholesterol and HbA1c biomarker status (normal vs. abnormal) were the target labels. Model inputs included patient demographic risk factors (age, race, gender, socioeconomic status), clinical measures (BMI, blood pressure), and before and after pandemic onset time periods. Model performance was evaluated and compared to baseline NAS models (Differentiable ARchitecTure Search (DARTS) and Efficient NAS (ENAS)), as well as traditional ML classifiers such as Random Forests and Support Vector Machines.Results: GANAS outperformed both ENAS and DARTS as well as traditional ML models for predicting both HbA1c and LDL cholesterol biomarker status. For HbA1c prediction, GANAS achieved an accuracy of 0.9739, precision of 0.9485, recall of 0.9739, and an F1 score of 0.9610. For LDL cholesterol prediction, GANAS maintained robust performance with an accuracy of 0.9117, precision of 0.9143, recall of 0.9100, and an F1 score of 0.9120. SHAP analysis revealed that socioeconomic factors, particularly middle-class status, and before and after the pandemic period were the strongest predictors.Conclusion: The superior performance of GANAS suggests its potential to improve the identification of CVD patients that need interventions during pandemics, enhancing risk assessment and personalized care. However, our results need to be validated on data gathered from more hospitals to confirm generalizability.CCS Concepts• Computing methodologies → Neural networks.