Implementing Informative-Based Active Learning in Biomedical Record Linkage for the Splink Package in Python

Choosing the similarity threshold at which two records are classified as the same patient is still an open problem in biomedical record linkage. This paper implements an informative-based active learning strategy for the Python package Splink, which measures how useful candidate training pairs are before they are manually labelled. Active learning should always be considered when training data come from manual labelling, and the label frequencies it produces give a quick indication of how hard a linkage problem is. Both are essential in big-data settings, where under- and overfitting are amplified.