Wong FY
AI-driven Survival Modelling in Breast Cancer

Abstract Disclosures

Abstract
Background: Machine learning (ML) and deep learning (DL) have shown tremendous success in classification and regression tasks, especially for unstructured data such as images, time series data, and texts. The use of ML and DL techniques for survival prediction in time-to-event data in medicine have been relatively modest. Several ML/DL algorithms [1,2] developed for survival modelling include: Random Survival Forest (RSF), Multi-Task Logistic Regression (MTLR), Support vector machines (SVM), and Deep neural network (DeepSurv). The advantage of ML/DL techniques over traditional semi-parametric or parametric techniques such as Cox model or models with defined probability distribution, is the ability to model the non-linear relationship and improve the prediction made. For instance, the DeepSurv network [2] was applied to both simulated treatment data and Rotterdam & German Breast Cancer Study Group, and have outperformed the use of Kaplan-Meier (KM) curve and RSF in providing treatment recommendations.

Aims: There are over 12,000 cohorts and this is a good sample size to start building a decent ML/DL model. The first aim is to compare the ML/DL techniques with conventional Cox model for survival prediction, and quantify the performance via time-dependent ROC curve. The second aim is to compare the techniques in stratifying patients undergoing treatment and looking at the possibility of developing a recommender support system. The third aim is to apply techniques of ML/DL to patients’ investigation, including radiological and laboratory findings in addition to clinical factors, to predict for response to treatment. The fourth aim is to apply techniques of ML/DL to perform economic modelling for genetics in breast cancer patients.

Hypothesis: Techniques of Machine learning (ML) and deep learning (DL) can be applied to predict survival and other clinical outcomes in breast cancer patients from their demographical, clinicopathology, laboratory and imaging –omics data with better precision than traditional regression methods.

Methodology: Retrospective study using information of patients in the Joint Breast Cancer Registry (JBCR) Singapore including using the results from their laboratory and imaging investigations. In addition to traditional semi-parametric or parametric techniques such as Cox model, we will use ML/DL techniques, including but not limited to: Random Survival Forest (RSF), Multi-Task Logistic Regression (MTLR), Support vector machine (SVM), and Deep neural network (DeepSurv), to model the non-linear relationship and improve on prediction accuracy. For imaging-omics data, pyradiomics 3.0 pipeline will be used for feature extraction, followed by dimensionality reduction, and then ML/DL.

Joint Breast Cancer Registry Singapore