基于健康数据分析的半监督在线学习血糖预报建模算法研究

Diabetes has become the third human health harm, blood glucose prediction makes lots of sense for both diabetes prevention and medicare of diabetes patients. Lots of blood glucose related records have been saved from various data sources, including electronic medical records, health examination records, dietary records, sport and sleep records, etc. However, existing blood glucose predictive algorithms only leverage one or few kinds of aforementioned data sources and widely existed unlabelled samples are unused to improve the model precision. Additionally, owing to the large scale of healthcare data, traditional offline modelling algorithms are too time-consuming for practical use. In this study, by leveraging the high volume, high variety, high sparsity characteristics of healthcare data and based on the ensemble learning theory, a novel semi-supervised online learning blood glucose predictive model will be designed with high precision and online updatable virtue. In order to meet the practical requirements including low space and computation complexity, high accuracy and adaptive online updatable, etc., the model precision and modelling efficiency are improved by effectively use of data from different data sources and unlabelled samples respectively.

糖尿病已成为危害人类健康的第三大疾病，血糖预测对糖尿病预防与患者治疗具有重要意义。随着信息化与网络化及可穿戴设备的发展，血糖相关的健康数据大量的通过多类数据源被收集存储，包括电子病历、体检记录、饮食、运动、作息等。而现有血糖预报算法只针对上述的一种或者少量几种数据进行建模，且没有使用大量存在的无标签样本辅助建模，无法充分利用全部数据。此外，传统的离线建模算法针对大量的健康数据建模速度过慢，严重影响了血糖预报算法的实用性。本研究针对健康数据的数据量大、数据源多样、稀疏度高等特点，基于集成学习理论，充分利用多数据源数据与大量无标签样本数据提升模型精度并加快建模速度提升现有血糖预报算法模型，设计与实现可在线更新、可利用不同数据源有效建模的半监督在线学习血糖预报算法模型，以满足现实应用中往往需要满足的低空间与计算复杂度、高准确率、模型可自适应在线更新等需求。

通过真实体检场景下108386条健康数据收集与分析，发现体检数据具有非常好的分析价值。按照之前的课题设计，通过半监督学习等不同的算法对数据进行了建模分析，得到了预期中的预测模型结果。本研究在提升算法理论基础上，更多的提供了可产业化的实际模型，为通过医疗健康数据分析辅助筛查一些疾病提供了理论和实践基础，包括对糖尿病发病风险的预测（预测血糖未来一年上升或者下降正确率超过80%），包括对精神心理健康指标，如抑郁（AUC曲线下面积0.948）、双相情感障碍（AUC曲线下面积0.921）等疾病及状态的辅助筛查，都有广阔的市场空间和社会价值。