In many practical applications, it is important to check whether certain patterns occurred in the past. For example, when performing a medical procedure such as dialysis, it is important to check whether the patient experienced a fast decrease in vital characteristics in the past. Such a decrease can be dangerous, so patients with this past experience should be carefully monitored. Detecting such patterns based on the original time series is time-consuming--and in medical situations, decisions are often needed urgently.
To speed up detection, the authors preprocess the corresponding time series: instead of the original values, they only store the general trends and values corresponding to different time quanta. For example, for each four-hour period, they store whether during this period the characteristic of interest was stationary or increased or decreased (and whether the corresponding increase/decrease was weak or strong). Similar information is stored about each eight-hour period, each two-hour period, and so on. The authors then provide efficient algorithms that use such preprocessed time series to detect patterns. Experiments with the actual dialysis-based time series confirm the efficiency and scalability of this approach.