Identification of Patient Prescribing Predicting Cancer Diagnosis Using Boosted Decision Trees

摘要

Machine learning has potential to identify patterns in pre-diagnostic prescribing that act as an early signal of cancer diagnosis. Danish studies using classical regression models have shown that prescribing of particular drugs increases in the months prior to lung and colorectal cancer diagnosis. The aim of this case-control study is to assess the potential for machine learning to extend these findings to identify combinations of prescriptions that might act as pre-cancer signals. We use a boosted trees approach to analyse prescriptions data from NHS Business Services Authority linked to English cancer registry data to classify individuals into two classes: cancer patients and controls. We then identify the drugs that contributed the most to the classification decisions in the models. To the best of our knowledge, this is the first study utilising machine learning to find pre-diagnostic primary-care-prescription-related indicators of cancer diagnosis in England. We assess two feature selection approaches using text categorisation methods alone and in combination with clinical domain knowledge. Matched samples of controls (ten controls for each patient) to control for age are used throughout. We train models for matched cohorts of 6,770 lung cancer patients and 5,869 colorectal cancer patients starting the cancer pathway for the first time between January and March 2016. The models outperform classical methods by AUC, AUC-PR, and F_(0.5) score, showing strong potential forusing machine learning to extract signals from this dataset to aid earlier diagnosis. Our findings confirm the Danish studies.