Community and Facility Health Information System Integration in Malawi: A Comparison of Machine Learning and Probabilistic Record Linkage Methods

Abstract

Accurate and efficient record linkage methods are essential to link patients between community health worker digital health apps and an EMR system, facilitating information flow and improving coordination of care. This study presents the eTrace workflow as an illustrative example, highlighting the benefits of enhanced coordination of care for patients in antiretroviral and non-communicable disease programs in rural Neno district, Malawi. This research focuses on the following major contributions: (1) development of a machine learning-based record linkage model for electronic health information systems, (2) comparison between the machine learning-based and probabilistic approaches to record linkage and (3) a concrete evaluation of our approach on real data for the eTrace workflow. A review of the standard record linkage architecture and its application to health information exchange systems is also presented. An empirical comparison conducted of logistic regression and the Fellegi-Sunter algorithms for this use case reveals comparable results. Both classifiers demonstrate an average precision of 0.86, while logistic regression achieves a higher recall at a fixed 0.90 precision of 0.74.