Abstract: The mortality prediction of ICU patient has been an active topic in the past decades. Machine learning algorithms have been proved to have preliminary effects in this domain and still have room for improvement. In order to deal with the ICU time series which is both high dimensional and uncertain sampling interval, this paper proposes the idea that the unequal sampling frequency phenomenon in time series can be transferred to the empty value under the regular sampling frequency and corresponding strategies. Then, this paper proposes a two-step hybrid framework which combines the time series clustering and machine learning algorithm. In the first step, the dimension and uncertainty are reduced; in the second step, classical machine learning algorithms are conducted for mortality prediction of ICU patient. The experiments on a public data set show that the results of classifying the minority death patients are more efficient than the traditional solutions and the elastic interval is better. The selection for best time interval is validated by the experiments meanwhile.

Key words: ICU, uncertain time series, prediction, machine learning, hybrid framework

摘要： ICU病人生死预测一直都是医学界的研究热点和难点。数据挖掘的机器学习方法近年来在该领域取得了一定的进展，但依然有很大的发展空间。针对ICU时序数据的高维度和不确定间隔采样特性，提出了不确定间隔采样转化为确定间隔的空采样的思想和相应的处理策略；在此基础上将传统的时间序列聚类与机器学习方法相结合，提出了一个两阶段的混合多机器学习算法框架，使得数据集的高维和不确定性得到了约简，从而可以采用经典的机器学习方法挖掘病人生死知识。在一个公开数据集上的两组实验结果表明，基于该算法框架的ICU病人死亡预测方法对于少数样本的分类效果优于传统方法，弹性时间间隔下的预测效果更好，最优时间间隔的选取可以通过实验效果来验证。

关键词: ICU, 不确定时间序列, 预测, 机器学习, 混合框架