Recurrent Time-Varying Multi-Graph Convolutional Neural Network for Personalized Cervical Cancer Risk Prediction

Abstract

Cervical cancer screening programs have reduced the incidence of cervical cancer, but suffer from over- and too infrequent screening as women’s risk of developing cervical cancer differs. Personalized risk prediction models contribute toward efficient, personalized cancer screening. This paper presents a personalized time-dependent cervical cancer risk prediction scheme to aid experts in recommending screening intervals. From partially observed screening histories, the proposed approach learns time-varying row-graphs that model the time-varying relations among the screening records of patients and a column-graph that encodes smoothness of an individual screening history. Then, leveraging these geometric structures, we reconstruct the entire latent risk of each individual from scarce screening data. In order to accomplish this, a novel time-varying multi-graph convolution neural network is proposed. These estimated latent risk profiles are used to forecast the cancer risk of new patients. The proposed approach is tested both on synthetic and real-life screening data obtained from the Cancer Registry of Norway.