Multimodal Integration of Data for Breast Cancer Risk Prediction
Ryan Tan, Tan HQ, Wong FY
Abstract:
Major advances in cancer therapy have come from identification of biomarkers for adjuvant therapies. For example, substantial decreases in breast cancer mortality have come through use of estrogen receptor (ER), progesterone receptor (PR) and ERBB2 (HER2/neu) as biomarkers for adjuvant therapy. Recently, incorporation of gene signatures to predict outcomes of ER-positive breast cancer patients has allowed for some patients to avoid chemotherapy. However, these studies relied on bulk and clonal tumor characteristics. Evidence suggests tumors that have a predisposition towards progression are more heterogenous and tend to carry subclonal alterations that are difficult to detect using conventional histologic, immunohistochemical or genomic sequencing methods.
Artificial intelligence I has emerged as a tool in that can be leveraged to detect cancer, predict some standard pathology biomarkers, genetic mutations and gene expression signatures in high-resolution digital pathology images. Body composition analyses of radiology images also holds potential for better risk prediction and prognostication of various clinical outcomes. However, these features have yet to be incorporated in current clinical models of cancer risk, disease outcome prediction and treatment response.
Thus, we aim to utilize artificial intelligence for integrative multimodal analysis of histologic, genomic, molecular, radiologic, laboratory and clinical data to formulate better composite biomarkers of cancer risk, disease outcome prediction and treatment response. As an example, prior work by Dr. Wong Fuh Yong and Tan Hong Qi has shown that traditional machine learning models are able to predict OncotypeDx risk categories using just text clinical variables with decent area under the receiver operating characteristic scores of between 0.7 - 0.75 with around 300. We will work on improving this with larger and multi-modal datasets together with Dr. Cheng Chee Leong and other breast pathology colleagues here. Other use cases we plan to work on include automated detection of adipose tissue inflammation on histology slides and prediction of subsequent breast cancer risk post-diagnosis of ductal carcinoma in situ.

Joint Breast Cancer Registry Singapore