Real-Time Medical Electronic Data Mining Based on Shapelet Pattern Recognition

Data mining of clinical data has big feasibility for refining the quality of hospital remedy and increasing survival probability of patients. Data-driven prediction techniques rely heavily on data gathering and analysis of patient vital signs. Time series shapelets are lately discovered short discriminant subseries that are not only accurate, but also explainable for the classification of time series problems. Similarity between shapelets and the whole series are employed in shapelet-based classification as a discriminatory feature, which are more comprehensible in the problem domain. This paper aims to exploit features in a distance-based classifier by replacing the distance between wireless sensor networks medical electronic data by the distance between shapelet transformed features. By using the cross relationship of different vital signs, we got the shapelets regions. Also, we introduced Multi-metric Large Margin Nearest Neighbor Classification(LMNN) to study Mahalanobis distance matrixes for the purpose of reducing the dimension of the transformed feature while keep the non-Target patients' features are far away from the target patients'. Experiments show that the combination of shapelets and Multi-metric LMNN can significantly improve the classification precision and also gives better explanation to doctors in units.