Uncertainty Characterization for Predictive Analytics with Clinical Time Series Data

Uncertainty Characterization for Predictive Analytics with Clinical Time Series Data

34th AAAI Conference on Artificial Intelligence (AAAI-20) W3PHIAI-20 Workshop

Yang Guo, Zhengyuan Liu, Savitha Ramasamy, Pavitra Krishnaswamy

Bayesian LSTM, Uncertainty Quanti cation, Time Series

07 February 2020

Recurrent neural networks (RNNs) o er state-of-the-art (SOTA) performance on a variety of important prediction tasks with clinical time series data. However, meaningful translation to actionable decisions re- quires capability to quantify con dence in the predictions, and to address the inherent ambiguities in the data and the associated modeling process. We propose a Bayesian LSTM framework using Bayes by Backprop to characterize both modelling (epistemic) and data related (aleatoric) uncertainties in prediction tasks for clinical time series data. We evaluate our approach on mortality prediction tasks with two public Intensive Care Unit (ICU) data sets, namely, the MIMIC-III and the PhysioNet 2012 collections.We demonstrate the potential for improved performance over SOTA methods, and characterize aleatoric uncertainty in the set- ting of noisy features. Importantly, we demonstrate how our uncertainty estimates could be used in realistic prediction scenarios to better interpret the reliability of the data and the model predictions, and improve relevance for decision support.

PublisherCopyrights

This research was supported by grant funding from A*STAR, Singapore (SSF A1818g0044 and IAF H19/01/a0/023).

https://oar.a-star.edu.sg/communities-collections/articles/15292

Institute for Infocomm Research