We trained different time series models to embed medical time series data from mechanical ventilation episodes, and then we clustered these to uncover hidden patient subtypes in the data.
Our model, Temporal Pointwise Convolution (TPC), is specifically designed to mitigate common challenges with Electronic Health Records, such as skewness, irregular sampling and missing data. We have achieved significant performance benefits of 18-68% over the Long-Short Term Memory (LSTM) network, and the Transformer.
Our model, LSTM-GNN, is designed to take advantage of similarity between patients in the EHR (established using the diagnoses). First, it processes the time series data for each patient with the LSTM component, before sharing information within the neighbourhood of patients via the GNN. This is an alternative way of presenting diagnoses information (the common approach is to use an encoder in the late stages of a model). We found that using both methods together gains the best performance.
Transfer learning with convolutional neural networks on the task of melanoma classification.