DEEP LEARNING IN AUTOMATED CLASSIFICATION OF ADVERSE EVENTS IN CLINICAL STUDIES OF ALZHEIMER’S DISEASE

Abstract

The Clinical Data Sciences Initiative (CDSI) by USC's Alzheimer's Therapeutic Research Institute (ATRI) was created to identify opportunities for research and innovation in the conduct of clinical studies and to accelerate the completion of study milestones, improve participant safety review and increase data quality standards. The collection and analysis of adverse event (AE) data during the course of a clinical study plays a critical role in ensuring participant safety as well as proper reporting. The unstructured diagnosis data collected must be post-processed to facilitate downstream analysis and reporting through a manual review process with classification of adverse event in accordance with standard medical terminology code from the Medical Dictionary for Regulatory Activities (MedDRA). Deep Learning with its simple and flexible framework has shown rapid increase in accuracy with big data, along with enormous growth and success when applied to several different areas of research including computer vision and speech recognition. A unique advantage of the deep networks (DNN) is its demonstrated scalability without plateauing of accuracy. Prior work by this team applied a well-established approach for text categorization popularly known as ‘Bag of Words’ to AE data and tested a wide range of classifiers (Ravindranath et. al, AAIC 2017). We showed that K-Nearest Neighbors (KNN) is simple and accurate with smaller training sets and the resource intensive logistic regression and neural network displayed similar performance after balancing categories in smaller training set. Here, we show that a DNN demonstrates consistency in prediction as well as improvement in accuracy over the previously reported KNN classifier. DNN obtained an accuracy of approximately 75% in its top prediction with a promise of improving the accuracy as the AE dataset continue to grow. We expect this method reduce the workload on the medical coders and allow them to focus on evaluating the predicted results.