Temporal case matching with information value maximization for predicting physiological states

With the rapid growth in volume of temporal medical data, predicting physiological states plays an important role in classifying medical cases. In this study, we propose a novel temporal classification framework that aligns multi-granularity modeling with decision-making. Particularly, we present a method for propagating case matching to predict unlabeled temporal cases and optimize the information value, which is the amount gained by the feature reconstruction module in answering queries. The proposed method facilitates the execution of multi-granularity case matching with temporal similarity and provide practitioners with a useful method of understanding temporal case-based reasoning. In the proposed method, the objective functions in a bilevel mixed integer optimization are the size of the reconstructed feature sets and their information values, which trade off the utility of additional information against the cost of feature combination depending on the decision variable. Unlike the conventional case matching method that uses all temporal features, the proposed method establishes effective classification rules based on the optimal temporal feature set. Moreover, this method also adapted multi-scale entropy to extract the dynamic features from multivariate temporal data. The numerical experiment verifies the effectiveness of the modeling framework and the robustness of the classification rules.