Application of neural network techniques for risk stratification of multiple myeloma patients

Multiple myeloma(MM) is a challenging and incurable type of blood cancer. The global impact ofMM was estimated in 2018, with 160,000 recorded deaths worldwide. Accurate risk stratificationremains critical for personalized treatment, yet current models such as the International StagingSystem (ISS) lack sufficient accuracy and the goal is to build models that can stratify the risk ofpatients better than ISS. There remains a lack of accurate and reliable risk stratification models for MM. In this thesis, weinvestigated neural network (NN)-based predictive models to stratify high-risk MM patients fromlow-risk ones. We implemented four NN architectures including 1D-CNN, CNN-MLP hybrid, Resnet and Tabnet.The effect of using input features without any selection was compared to two feature selectionmethods, variance threshold and NetRank. Furthermore, we investigated the effect of representing input data with and without sorting by chromosomal location of genes. The dataset consistedof 241 patients, with 97 categorised as high-risk patients (death within 3 years of diagnosis) and144 as as low-risk (survived longer than 3 years). Our results show that Tabnet combined with feature selection via variance thresholding outperformed all other methods. It achieved an accuracy of 0.72, F1-score of 0.76, recall of 0.92, precision of 0.60 and area under curve (AUC) of 0.58. The performance on accuracy was better thanthe current clinically used international staging system (ISS), which achieves accuracy of 0.65.