244P Identifying metastatic breast cancer in Northern Ireland using machine learning algorithms and cancer registry data: Development and validation

This is the authors' abstract. We don't add key points for this paper.

Identifying Metastatic Breast Cancer using Machine Learning Algorithms – development and validation.Metastatic breast cancers (MBC) which occur sometime after primary breast cancer (BC) are not systematically recorded by population-based cancer registries (PBCR), meaning accurate data on MBC cases is difficult to determine. We developed multiple machine-learning models to identify MBC cases in Northern Ireland (NI) and externally validated the predicted classifications using a gold standard dataset (GSD). Data from the NI Cancer Registry (NICR) Official Statistics (OS) dataset was linked with administrative health datasets, namely Patient Administration System (PAS) and General Register Office's (GRO) death records data. Over two million PAS records were processed, with older ICD-9 converted to current ICD-10 codes. Patients with a non-breast primary cancer prior to their MBC and those with only C77.3 were excluded. Supervised machine learning models (MLM) including Classification and Regression Trees [CART], Random Forest, K-Nearest Neighbour [KNN], Extra-Gradient Boosting [XGBoost], Gaussian Naïve Bayes and Support Vector Machine [SVM] were built using the merged dataset to classify MBC patients. Models were externally validated against a GSD, an NICR research dataset of 1, 028 BC patients diagnosed 2009 and followed up for MBC to 2017, with 184 patients determined as MBC via manual search by experienced NICR staff of radiology, pathology, and Multidisciplinary Team (MDT) databases and oncology electronic software systems. CART was the best performing MLM as assessed by feature importance, with PAS Primary Inpatient Diagnosis and GRO Primary Cause of Death being the most important features of the model. The CART model successfully identified MBC patients with a high degree of accuracy, with sensitivity of 90.6%, specificity of 99%, positive predictive value (PPV) of 93.6% and negative predictive value (NPV) of 98.5%. The machine learning algorithms developed were successfully applied to integrated PBCR, hospital inpatient record and GRO death record datasets to identify MBC cases in NI which will help planning, care and support for MBC patients.