CinC Challenge: Cluster analysis of multi-granular time-series data for mortality rate prediction

The goal of this research is to develop novel cluster analysis techniques to identify similarity between ICU time-series data. The results generated by cluster analysis are further used for ICU mortality prediction. To preprocess multi-granular ICU time-series, we proposed a segmentation-based method to divide time-series into several segments. The minimal and maximal values within each segment were captured to maintain the statistical feature of the segment. A weighted Euclidean distance function was in place to evaluate the similarity between two instances and clustering was later used to convert each time-series into a corresponding cluster number. This way, we turned the high dimensional ICU time series data into a 2-dimensional matrix. A rule-based classification model was developed from this 2-dimensional matrix, and the model was used to predict the in-hospital mortality for test cases. The experiments show that above approach is effective in handling ICU time-series data.