Chuan Zhou

Ph.D

University of Michigan

Research Professor

PLCO (Learn more about this study)

PLCOI-791

Jul 6, 2021

Developing a general AI decision support framework for early diagnosis and treatment of variety of cancers

Data-driven reinforcement learning (RL) is one of the machine learning artificial intelligent (AI) paradigms together with supervised learning and unsupervised learning. RL has been widely used in various applications such as autonomous driving, financial trading, gaming, and is increasingly used in healthcare. We are developing a general framework based on RL for early diagnosis and treatment of different types of cancers. There are major challenges in applying AI to healthcare, including 1) the decision should be made upon the data from time series and different types of exams, 2) the lack of rigorous predictive models, and 3) the mismatch between existing dataset and patient population. The aim of our study is to develop an AI framework that can be flexibly used to build predictive models for early diagnosis, monitoring the disease progress over the time, and predicting the effectiveness of the treatment and patient outcome. We have developed predictive models using deep learning with RL method (D-RL model) for early diagnosis of lung cancer using the data set (including CT and pathologic images) collected from National Lung Cancer Screening Trial (NLST) project with access permission. In this study, with data access permission from Prostate, Lung, Colorectal and Ovarian cancer screening trial (PLCO), we will rebuild new D-RL models for early diagnosis and predicting the effectiveness of the treatments and patient outcomes for prostate, lung, colorectal and ovarian cancers. We will also evaluate the effectiveness by applying our previously built models to the PLCO data set with and without transfer learning, and evaluate the generalizability of our developed models with large data set from PLCO. Leveraging the “big data” of PLCO, we expect our developed predictive models can be more robust, reliable, and more importantly, be generalizable to a large patient population.