基于电子健康档案中异构时态数据的学习

Learning from Heterogeneous Temporal Da-ta Based on Electronic Health Records

摘要: 电子健康档案包含大量的纵向数据，对于生物医学信息学研究很有价值。然而，由于数据的复杂结构，包括随时间不均匀分布的临床事件，对标准学习算法提出了挑战。时态数据建模的一些方法依赖于从时间序列中提取单一值，导致有潜在价值时序信息的丢失。因此，如何更好地解释临床数据的时效性，仍然是一个重要的研究问题。本文研究了电子健康档案中时态数据新的表示方法，这些表示保留了时序信息，并且可以由标准机器学习算法直接处理。基于时间序列数据符号化表示的研究方法有多种不同的方式。使用电子健康档案真实数据库中临床测量的数据集的实证研究结果表明，相比使用原始序列或聚类序列，对随机子序列使用距离度量显著提高了预测性能。本文提出的表示方法更好地解释了临床事件的时效性，对于生物医学领域的预测任务十分关键。

Abstract: Electronic health records contain a large number of longitudinal data, which is valuable for biomedical informatics research. However, standard learning algorithms present challenges due to the complex structure of the data and clinical events that are unevenly distributed over time. Some methods of temporal data modeling depend on extracting single values from time series, which leads to the loss of potentially valuable sequential information. Therefore, how to better explain the temporality of clinical data is still an important research question. In this paper, a new representation of temporal data in electronic health records are studied, which preserves the sequential information that can be processed directly by the standard machine learning algorithms. The research method based on time-series data symbol representation has many different ways. Empirical studies using clinically measured datasets in the real-life database of electronic health records have shown that using distance metrics for random subsequences significantly improves predictive performance compared to the use of original sequences or clustering sequences. The representation method proposed in this paper better explains the temporality of clinical events and is key to the prediction task in the biomedical domain.