Utilizing Primary Health Care Data for Early Detection of Colorectal Cancer: A Machine Learning Approach

Sammanfattning

Colorectal cancer (CRC) is a health challenge worldwide and early detection of the disease is crucial to improve patient prognosis. It is common for the first contact with care to occur in primary care centers where general practitioners often face the challenge of distinguishing CRC from other diseases with similar symptoms. In this master thesis, patient records from primary care were used to create, optimize, and evaluate a machine learning model that classifies patients with CRC for early detection of the disease. The data used in the project included parts of electronic health records (EHRs) from both public (SLSO) and privately run (Capio and Praktikertjänst) primary care centers in the Stockholm region. The available dataset was cleaned and pre- processed, and then tested on four separate models. After selecting and optimizing the most promising model, LightGBM, a detailed evaluation of the model was performed. To simulate realistic clinical conditions, data from the three months prior to diagnosis were excluded from two of the datasets. The results were then compared with a baseline machine learning model that utilized ICD codes extracted from EHRs in primary care for early detection of CRC.The results showed that the final developed model had a generally good performance with an AUROC score of a maximum of 85.8%, which indicates very good ability to distinguish between the classes. The performance dropped when using the datasets with 3 months of data removed, but the ROC curves still showed a better ability than random classification to distinguish between the classes with a AUROC score of maximum 60,8%. The results also showed that the model developed in this master thesis outperforms the baseline model, which was based on ICD codes, from a performance perspective. For future development and before a possible clinical implementation, a larger data set should be used for training and testing.