Classifying unevenly spaced clinical time series data using forecast error approximation based bottom-up (FeAB) segmented time delay neural network

Clinical time series data contain large set of time-stamped data points that describe the patient's health. The observations of these data points are done at irregular intervals and hence knowledge mining turns challenging. To overcome this, there is a need to reduce the dimension (length) of time series data into smaller representations with minimal loss of information. The objective of this work is to present a forecast-error approximation-based bottom-up (FeAB) segmentation for segmenting and classifying clinical time series data using time delay neural network (TDNN). The proposed approach includes two functionalities namely temporal data summarisation and classification. In temporal data summarisation, clinical time-series data are divided into sequence of temporal interpreted segments using FeAB segmentation. FeAB adopts a double exponential smoothing technique to derive the growth rate, mean and forecast-error for each clinical observation. The obtained forecast-error is used to compute the merge-cost for FeAB segmentation. TDNN classifier builds classification model for the segmented time series. The classifier is trained using backpropagation with Levenberg-Marquardt algorithm. The time series dataset of hepatitis and thrombosis patients are used for experimentation. The results illustrate that the proposed framework has effectively handled the temporal data irregularities and has shown improvement in classification accuracy.