3D-CNN-SPP: A Patient Risk Prediction System From Electronic Health Records via 3D CNN and Spatial Pyramid Pooling

The problem of extracting useful clinical representations from longitudinal electronic health record (EHR) data, also known as the computational phenotyping problem, is an important yet challenging task in the health-care academia and industry. Recent progress in the design and applications of deep learning methods has shown promising results towards solving this problem. In this paper, we propose 3D-CNN-SPP (3D Convolutional Neural Networks and Spatial Pyramid Pooling), a novel patient risk prediction system, to investigate the application of deep neural networks in modeling longitudinal EHR data. Particularly, we propose a 3D CNN structure, which is featured by SPP. Compared with 2D CNN methods, our proposed method can capture the complex relationships in EHRs more effectively and efficiently. Furthermore, previous works handle the issue of variable length in patient records by padding zeros to all vectors so that they have a fixed length. In our work, the proposed spatial pyramid pooling divides the records into several length sections for respective pooling processing, hence handling the variable length problem easily and naturally. We take heart failure and diabetes as examples to test the performance of the system, and the experiment results demonstrate great effectiveness in patient risk prediction, compared with several strong baselines.