Matching Multifrequency Clinical Time Series

In this article we consider the problem of matching time sequences in the MIMIC II[1] database. Unlike other time series similarity matching problems, our task is to match time sequences of observations made at different frequencies that have been obtained from the same subject. For each time series in high-frequency, our goal is to find its low-frequency counterpart. Heart rate of each patient in the ICU is recorded by both the patient monitor automatically and by nurses manually. Series of heart rate recorded by nurses(low-frequcency clinical data) is widely applied to analysis and detection of diseases, while the high-frequency numerical data monitored by bedside monitors do not have the information needed to link it directly with clinical data. We want to match those anonymous numerical data with clinical data for future research. We studied various metrics of time series similarity and proposed two efficient metrics with high accuracies. To evaluate the performance of various metrics, synthesized unmatched pairs, along with provided matched pairs of clinical and numerical datasets, compose the training and testing set. Cross-validation is conducted to evaluate the metrics. Experiments of matching multifrequency series are conducted on the testing set. According to our experimental results, accuracy of detecting true matching is 58.80%. Higher accuracy is expected when a more efficient matching algorithm is implemented.