Líkanagerð fyrir afar sjaldgæfa atburði í sannprófunarvillum sjálfvirkra pilluskanna

Medication errors are common in health care and potentially dangerous. Pill scanners act as a second nurse, scanning pills and verifying them against patient-specific prescription information. Verification errors made by devices are extremely rare, making it challenging to evaluate effects and predict outcomes. This thesis examines potential external factors contributing to device verification errors and strategies to address extreme class imbalance, using both linear and non-linear modelling with Logistic Regression (LR) and Random Forest (RF), respectively. Two data sets were analysed: a complete data set and a reduced data set, in which observations were restricted to flagged pills only. Random undersampling and oversampling with SMOTE were compared using a 500-iteration Monte Carlo simulation, in which each iteration split devices 80/20 to mimic unseen data, making the strategies’ intra-run performance comparable. Restricting the data to only flagged pills showed a slight improvement in F1-score, but a drastic decrease in AUROC. At the same time, the effects and importance of variables remained similar. Neither sampling strategy improved the model performance. The sampling strategies dramatically reduced LR´s accuracy while lowering its F1-score variance, whereas the RF performance remained unchanged. RF achieved higher accuracy overall but predominantly lower F1-scores than LR. Based on the F1-score, LR performed better than RF, with LR using the conventional sampling strategy scoring the highest most often, in approximately 40.6% of iterations.