Trend Prediction Methodology Based on Time Series Similarity Analysis and Haar Wavelet Decomposition

This work presents a strategy for the prediction of biosignals' future trend, based on a Haar wavelet transform. The proposed scheme is based on the hypothesis that the future evolution of a given biosignal (template) can be estimated from similar patterns existent in a historic dataset. Thus, the first step consists of a simple and efficient measure to evaluate the similarity between biosignal time series. Then, supported on the similar patterns identified using the similarity process, a predictive scheme is introduced. The proposed approach, which does not use an explicit model, considers the wavelet decomposition of the signals (template and similar patterns) to determine the most representative trend at each of the several decomposition levels. These trends are then aggregated to derive the required biosignal future estimation. A Matlab tool was developed to support the proposed strategy, consisting of two main components: the similarity and the prediction modules. These were applied in the validation task using vital signals (heart rate, blood pressure and weight) daily collected during two tele-monitoring studies: TEN-HMS and My Heart.