Ontology-Based Temporal Relation Modeling with MapReduce Latent Dirichlet Allocations for Big EHR Data

In this paper, we propose a model called Temporal & Co reference Topic Modeling (TCTM) to do automatic annotation with respect to the Time Event Ontology (TEO) for the big-size Electronic Health Record (EHR). TCTM, based on Latent Dirichlet Allocations (LDA) and integrated into MapReduce framework, inherently addresses the twin problem of data sparseness and high dimensionality. As a non-parametric Bayesian model, it can flexibly add new attributes or features. Side information associated with corpora, such as section header, timestamp, sentence distance, event distance or disease category in clinical notes makes latent topics more interpretable and more biased toward co referring events. Furthermore, TCTM integrates Hidden Markov Model LDA (HMM-LDA) to obtain the power of both sequential modeling and exchangeability. A MapReduce based variational method is employed to do parameter estimation and inferences, thus enabling TCTM to overcome the bottleneck brought by big data.