Assessing the Application of Machine Learning Techniques to Predict Loss-To-Follow-Up

Abstract

Loss-to-follow-up (LTFU) is a burgeoning problem in the global health and public health community that negatively impacts treatment program evaluation, clinical trial data quality, and most importantly the quality of life of thousands of individuals with diseases and ailments of all kinds. There is considerable potential in applying machine learning techniques to predict on an individual basis whether patients are likely to become LTFU or not based on information gathered from their initial appointments and medical history. Few studies have approached the LTFU problem using such techniques. We applied four commonly used machine learning classifiers to clinical trial data from the Multicenter Osteoarthritis Study (MOST) to assess which algorithm best identified patients who became LTFU. The Naive Bayes algorithm was the most successful on all measures used here, but only achieved an accuracy of at most 78.3% and a true-positive rate of at most 56.9% on data sets that were modified to account for the very small number of patients who became LTFU relative to the entire group. This indicates the need for a more thorough search of possible machine learning algorithms to this problem, along with the need for greater understanding regarding the necessary preprocessing steps to make this approach as effective as possible.