Continuous Clinical Event Prediction with Recurrent Deep Networks
Continuous prediction of clinical outcomes and early warnings are a constant challenge in intensive care units. Early risk and outcome prediction remain difficult problems due to data size, class imbalance, and noisy heterogeneous data sources that must be integrated across time-scale and granularity before learning. In this paper, we demonstrate a successful deep risk and outcome learning network for a variety of outcomes in a real ICU dataset of 30,000 patients. Our work integrates clinical notes, demographics, vital signs and lab values using multiple recurrent neural networks with attention mechanisms dedicated to generating interpretable features from the generated network. We investigate two tasks: predicting onset of critical events and predicting the need for interventions. In both cases, prediction is done in ``real-time'', and classification errors are propagated through the fully connected layers for weight updates in the learning process. This deep learning-based approach allows for the diverse attributes of each data type to be integrated into a single common representation. We demonstrate that our method improves prediction AUC compared to baseline models and identify key features for each task over time.
Marzyeh Ghassemi is a PhD student in MIT’s Computer Science and Artificial Intelligence Lab supervised by Dr. Peter Szolovits. Her research focuses on machine learning with clinical data to predict and stratify relevant human risks, encompassing unsupervised learning, supervised learning, structured prediction. Marzyeh’s work has been applied to estimating the physiological state of patients during critical illnesses, modelling the need for a clinical intervention, and diagnosing voice disorders from wearable sensor data. While at MIT, Marzyeh co-organized the NIPS 2016 Machine Learning for Healthcare (ML4HC) workshop, and her work has appeared in KDD, AAAI, IEEE TBME, MLHC, JAMIA, and AMIA-CRI.