Towards an LSTM-based approach for detection of temporally anomalous data in medical datasets

Health-related data is complex, heterogeneous, and frequently temporally discontinuous which makes it difficult to analyze, and even more difficult to assess for quality, detect errors, or address anomalies. Undetected errors and anomalies within medical data, if not properly addressed, can have far reaching humanitarian and/or financial consequences. Many machine learning methodologies applied towards automatic detection of complex errors or anomalies are well known. However, the additional challenge in detection of anomalies in health care data is the so-called temporal or contextual dependency – i.e., the challenge of distinguishing "episodes" of anomalous data that are anomalous either in the context of a particular time window or specific to some particular medical condition. In this paper, we present research in progress on the application to medical data of state-of-the-art anomaly detection techniques based on Long-Short-Term-Memory (LSTM) neural networks. We hypothesize that LSTMs present a robust and flexible approach to the temporal anomaly detection in medical data sets due in part to their ability to retain information about both long-term and short-term dependencies on the model outcome.