Multi-granular Time Series And Its Application In ICU Medical Mortality Prediction

Temporal data with domain background is a typical category of big data. Although Granular Computing is acknowledged as a vital theory for solving complex problems, the researches on temporal granular are still not systematic. As a representative domain problem with temporal characteristics, the mortality prediction for ICU patient is of great importance to the rescue of those with severely illness. The patient records with early stage are thus critically analyzed. Accordingly, the main researches are classified into the following three aspects:Firstly, in order to facilitate the research of temporal data, Granular Computing theory is introduced by applying binary relations on the axis of time so that a series of temporal granular can be induced. The original problem is thus converted to a novel model, named as temporal granular system(TGS). The relations between different temporal granular are further analyzed and the conversion methods are defined meanwhile. Based on the prediction task, the construction of multiple granularity as well as temporal elastic granular is proposed. This work can theoretically support the research of temporal data in real applications.Secondly, considering multiple features of ICU data, a hybrid algorithm framework with two-staged style is proposed. In the first stage, the innate connection within ICU data is analyzed hierarchically so that the high-dimension and uncertainty can be reduced. In the second stage, classical machine learning components are examined on the same data derived in stage one to develop the classifier, the mechanism between mortality prediction and multi-parameter can be thus achieved.Last but not least, in view of optimizing the mortality prediction performance and make best of the properties of uncertain sampling, imbalance and high-dimension, the crucial algorithms integrated in the proposed framework are upgraded. It embodies mainly on the clustering component and methods for temporal feature extractions. By comparing results the validity of aforementioned improvements, the effectiveness of proposed hybrid framework are demonstrated.