Artificial Intelligence in Health Care Research

Organizers: Xiaofei Wang (Duke)
Chair: Nate Bennett (UCB)
Vice Chair: Shuyen Ho (UCB)

David Carlson (Duke)
Mark Chang (BU)
Rachel Draelos (Duke)


Title: Machine Learning Methods to Learn Improved Electrophysiological Biomarkers in Clinical Trials
Speaker: David Carlson (Duke)

Clinical trials for autism spectrum disorder (ASD) and other conditions frequently utilize electroencephalography (EEG) measurements to track and evaluate neural dynamics. The utility of these measurements is determined by the efficacy of the biomarkers (or features) extracted from the EEG. Novel methods to learn improved biomarkers from clinical trial data can increase our ability to understand how the brain itself is changing during these treatments, enhance statistical power, and enhance clinical relevancy. Towards this end, we first developed a novel deep learning framework that is appropriate for relatively small data size and is interpretable such that it can explain how its predictions are made. Second, a particular challenge in medicine is the “little big data” structure, where we get many total samples, but they only come from a few participants. We model this as a multiple domain generalization problem and introduce a methodological framework to address it To demonstrate the efficacy of the developed approach, we apply the methodology to a recent clinical trial on cord blood infusions in ASD to capture brain differences pre-treatment, 6 months post-treatment, and 12 months post-treatment. We demonstrate that the proposed methodology significantly enhances predictive ability, significantly improving the utility of the measurements. By examining the features, we can visualize how neural changes correlate to treatment stage. These techniques and learned biomarkers are currently being evaluated on a recently completed phase 2 clinical trial.

Title: Artificial Intelligence For Drug Development & Future Health Science
Speaker: Mark Chang (BU)

Artificial intelligence (AI) has been used in drug discovery, molecular design, disease diagnosis and prognosis as discussed in my AI short course for the conference. In this talk, we focus two particular AI application areas: (1) similarity-based AI for clinical trials and precision medicine and (2) artificial general intelligent (AGI) and its implementations in elderly care robots.

We will review the differences among supervised, unsupervised, reinforcement learning, swarms intelligence, and evolutionary learning methods and applications of AI in drug development, precision medicine, and healthcare. Deep learning methods requires big data. However, data from clinical trials are usually small, thus imposes particular challenges in utilizing ANN-based deep learning. We will explain the similarity principle and outline the similarity-based machine learning for both small and big data, and present comparative results.

Unlike narrow AI that accomplishes specific tasks for human, AGI is the study of machines with full scope of human intelligence. We will discuss, from completely new perspectives, key concepts in AGI such as “understanding,” “creativity,” “discovery,” “invention,” and “emotion”. In contract to the mainstream big data approach with built-in languages, the new approach is a sequential, recursive, hierarchical reinforcement learning approach without the requirements of any built-in language or big data. In dealing with population aging issues, several companies have started making elderly care pets; however, the goal of our approach is beyond providing the traditional services, but the emotional companion too.

Title: Machine-Learning-Based Multiple Abnormality Prediction from Large-Scale Volumetric Medical Imaging Data
Speaker: Rachel Draelos (Duke)

Machine learning has the potential to power automated medical image interpretation systems that assist radiologists, accelerate the medical workflow, and benefit patients. Chest computed tomography (CT) scans are volumetric medical images that are widely used to diagnose and manage numerous conditions including cancer, lung diseases, and infections. Although the typical chest CT contains 10±6 different abnormalities, previous work in automated chest CT interpretation has focused on one class of abnormalities at a time, e.g. nodules. Even when successful, these focused models have limited clinical applicability because radiologists are responsible for a multitude of findings in the images. To address this need, we investigate the simultaneous prediction of multiple abnormalities using a single model. We curated and analyzed a chest computed tomography (CT) data set of 36,316 volumes from 19,993 unique patients. To the best of our knowledge this is the largest multiply-annotated volumetric medical imaging data set in the world. To annotate this data set, we developed a rule-based method for automatically extracting high-quality abnormality labels from free-text radiology reports. We use these extracted labels to develop a deep convolutional neural network model for multi-organ, multi-disease classification of whole chest CT volumes. This model reaches a classification performance of AUROC > 0.90 for 18 abnormalities, with an average AUROC of 0.773 for all 83 abnormalities, demonstrating the feasibility of learning from unfiltered whole volume CT data. To facilitate future research in automated chest CT interpretation, we plan to make all 36,316 volumes and labels publicly available pending institutional approval. We hope this work will contribute to the long-term goal of augmented medical image interpretation systems that enhance the radiology workflow and advance patient care.