Tsang Kevin C H, Pinnock Hilary, Wilson Andrew M, Ahmar Shah Syed
Annu Int Conf IEEE Eng Med Biol Soc. 2020 Jul;2020:5673-5677. doi: 10.1109/EMBC44109.2020.9175679.
While there have been several efforts to use mHealth technologies to support asthma management, none so far offer personalised algorithms that can provide real-time feedback and tailored advice to patients based on their monitoring. This work employed a publicly available mHealth dataset, the Asthma Mobile Health Study (AMHS), and applied machine learning techniques to develop early warning algorithms to enhance asthma self-management. The AMHS consisted of longitudinal data from 5,875 patients, including 13,614 weekly surveys and 75,795 daily surveys. We applied several well-known supervised learning algorithms (classification) to differentiate stable and unstable periods and found that both logistic regression and naïve Bayes-based classifiers provided high accuracy (AUC > 0.87). We found features related to the use of quick-relief puffs, night symptoms, frequency of data entry, and day symptoms (in descending order of importance) as the most useful features to detect early evidence of loss of control. We found no additional value of using peak flow readings to improve population level early warning algorithms.
虽然已经有多项利用移动健康技术支持哮喘管理的努力，但迄今为止，还没有任何一项能提供个性化算法，根据患者的监测情况为其提供实时反馈和量身定制的建议。这项研究使用了一个公开可用的移动健康数据集——哮喘移动健康研究（AMHS），并应用机器学习技术开发早期预警算法，以加强哮喘自我管理。AMHS包含来自5875名患者的纵向数据，包括13614份每周调查问卷和75795份每日调查问卷。我们应用了几种著名的监督学习算法（分类算法）来区分稳定期和不稳定期，发现逻辑回归和基于朴素贝叶斯的分类器都具有很高的准确率（曲线下面积>0.87）。我们发现与使用缓解症状吸入剂、夜间症状、数据录入频率和日间症状相关的特征（按重要性降序排列）是检测控制丧失早期迹象最有用的特征。我们发现使用峰值流量读数对改进总体水平的早期预警算法没有额外价值。
