71629 - Comparison of machine learning tools for identifying non-metastatic and metastatic colorectal cancer in primary care

Introduction Primary healthcare (PHC) often serves as the initial point of contact when diagnosing colorectal cancer (CRC). Identifying patients with cancer symptoms in PHC can pose challenges, but machine learning holds promise in aiding this process. Method Comparison between two case-control studies containing data on PHC visits in the Västra Götaland Region during 2011 for 542 patients diagnosed with non-metastatic CRC, 146 patients diagnosed with metastatic CRC, and 2716 matched controls. Separate models for prediction of non-metastatic and metastatic CRC were constructed using machine learning (stochastic gradient boosting) based on diagnostic codes and number of consultations from PHCs during the year before date of cancer diagnosis. Variable importance was estimated using normalized relative influence (NRI). Odds ratios of marginal effects were calculated for all variables. Result The optimal models for non-metastatic and metastatic CRC had area under curves (AUCs) of 0.832 (0.790–0.874) and 0.765 (0.675–0.856), respectively (Figure 1). In both models, variables having NRI > 1% were considered important. This rendered 16 variables of importance for non-metastatic CRC and ten variables for metastatic CRC. Both models contained anaemia, abdominal pain, other diseases of anus and rectum, and number of consultations among variables entailing higher risks. Only the non-metastatic CRC model included rectal bleeding, haemorrhoids, and change in bowel habit, while only the metastatic CRC model included fatigue. Hypertension had a weak protective effect in both models. Discussion Machine learning is useful for identifying variables of importance for predicting CRC in PHC. Variable importance differs between the two models.