•  
  •  
 

Corresponding Author ORCID Identifer

Kasthuri Subaramaniam https://orcid.org/0000-0003-0704-923X

Section

Regular Issue

Abstract

Healthcare Machine Learning (ML) applications face significant challenges in accessing high-quality training data due to stringent privacy regulations, institutional data silos, and concerns over patient confidentiality. This paper explores synthetic data generation as a viable and privacy-preserving alternative to real patient data for developing ML models in healthcare settings. We present techniques for generating realistic vital signs data including body temperature, blood pressure, heart rate, respiratory rate, and oxygen saturation according to appropriate statistical distributions. In addition, we demonstrate how synthetic datasets generated can be used to train diagnostic prediction models. The generated datasets were applied to multiple diagnostic prediction tasks such as hypertension, fever, Chronic Obstructive Pulmonary Disease, atrial fibrillation, and diabetes mellitus. Experimental results reveal that ML models trained solely on synthetic data achieved comparable predictive performance to those trained on real datasets for conditions with explicit physiological manifestations. In particular, gradient boosting classifiers attained an Area Under the Curve (AUC) of up to 0.89 in predicting hypertension. We also illustrate that augmenting sparse real patient data with artificial samples preserves model accuracy at the expense of decreased reliance on sensitive data. This method has great potential to satisfy healthcare organizations who are interested in creating stable ML applications without compromising on privacy standards like Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR).

Download

Share

COinS