Augmentation Techniques for Sequential Clinical Data to Improve Deep Learning Prediction Techniques

Methods based on neural networks have become more and more attractive in the medical domain as Deep Learning frameworks mature and popularize. One application in this context refers to the use of recurrent networks to predict the most probable clinical conditions of a patient, given his/her history of hospital/medical admissions. The problem is that clinical data to support machine learning is scarce, mainly due to privacy matters and to ill-structured medical databases. In this work, we demonstrate that it is possible to augment clinical data to improve the performance of automatic predictive systems. We introduce two methods to create synthetic clinical histories (trajectories) based on existing data; the first one extracts subsequences of trajectories to emphasize the transition in between hospital admissions; the second method benefits from the hierarchical structure of standard diagnosis codes (like ICD-9) trajectories whose characteristics resemble those of real-world clinics. Our results demonstrate significant improvements in two datasets, demonstrating the feasibility of data augmentation in the clinical scenario. Our results shall inspire the augmentation of data in several other scenarios related to the medical domain.