Predicting Acute Respiratory Infections from Participatory Data
Predicting Acute Respiratory Infections from Participatory Data
复制标题
从参与数据预测急性呼吸道感染
DOI：
10.5210/ojphi.v9i1.7650
复制
发表时间：
2017
影响因子：
--
通讯作者：
R. Chunara
中科院分区：
文献类型：
--
作者：
Bisakha Ray;R. Chunara
目的应用机器学习模型评价参与性数据对急性呼吸道感染(ARI)实验室诊断的预测价值。引言ARI具有流行和大流行的可能性。现有研究中根据个体体征和症状预测ARI的存在是基于临床来源的数据1。临床数据通常代表最严重的病例，以及那些来自可以进入医疗机构的地区的病例。因此，来自临床样本的病毒信息不足以捕捉普通人群中的疾病发病率，也不足以从症状中预测疾病的发病率。参与性数据--今天的个人可以自己产生的信息--在无处不在的数字工具的推动下，可以通过提供社区自我报告的数据来帮助填补这一空白。基于互联网的参与性努力，如流感2，通过及早和广泛地检测疫情和公共卫生趋势，扩大了现有的ARI监测。方法建立GoViral平台3，从社区获取自我报告的症状和诊断标本(表1汇总参与细节)。参与者包括来自数据最多的州、马萨诸塞州、纽约州、密歇根州、密歇根州和加利福尼亚州的参与者。每个参与者都被要求提供年龄、性别、邮政编码和疫苗接种状况。参与者提交了唾液和鼻拭子样本，并报告了以下症状：发烧、咳嗽、喉咙痛、呼吸急促、寒战、疲劳、身体疼痛、头痛、恶心和腹泻。病原体通过RT-PCR在Genmark呼吸道面板分析上确认(之前报道的完整病毒清单3)。缺失、无效或模棱两可的实验室测试的观察结果被删除。表2总结了二进制特性。年龄组别是：≤20，>；20和<；40，≥40代表年轻、中年和老年。失踪的年龄和性别价值是根据总体分布计算出来的。考虑了三种机器学习算法--支持向量机(SVMs)4、随机森林(RFS)5和Logistic回归(LR)。对个体特征及其组合进行了评估。结果是实验室诊断为ARI(1例)或不存在(0例)。结果10次交叉验证重复10次。使用的评估指标为：阳性预测值(PPV)、阴性预测值(NPV)、敏感性和特异性6。以咳嗽、发热为预测指标，LR和SVMS预测PPV值最高，为0.64(标准差：±0.08)。敏感度为0.59(±0.14)，以咳嗽、发热、咽喉痛为最佳。以发热、咳嗽、咽痛为代表的CDC ILI症状谱，RFS的NPV和特异度分别为0.62(±0.15)和0.83(±0.10)。添加人口统计数据和疫苗接种状态并不能提高分类器的性能。结果与使用临床来源数据的研究一致：咳嗽和发烧一起被发现是流感样疾病的最佳预测因素1。由于我们的数据包括轻度感染和无症状病例，与临床数据相比，分类器的灵敏度和PPV较低。结论发热和咳嗽一起是社区ARI的良好预测指标，但由于抽样偏差，临床数据可能高估了这一点。整合参与性数据不仅可以通过积极参与公众2来改善人口健康，而且还可以扩大仅以临床监测数据为基础的研究范围。表1.所包括参与者的详细情况。表2.二进制特征的编码
Objective To evaluate prediction of laboratory diagnosis of acute respiratory infection (ARI) from participatory data using machine learning models. Introduction ARIs have epidemic and pandemic potential. Prediction of presence of ARIs from individual signs and symptoms in existing studies have been based on clinically-sourced data 1 . Clinical data generally represents the most severe cases, and those from locations with access to healthcare institutions. Thus, the viral information that comes from clinical sampling is insufficient to either capture disease incidence in general populations or its predictability from symptoms. Participatory data — information that individuals today can produce on their own — enabled by the ubiquity of digital tools, can help fill this gap by providing self-reported data from the community. Internet-based participatory efforts such as Flu Near You 2 have augmented existing ARI surveillance through early and widespread detection of outbreaks and public health trends. Methods The GoViral platform 3 was established to obtain self-reported symptoms and diagnostic specimens from the community (Table 1 summarizes participation detail). Participants from states with the most data, MA, NY, CT, NH, and CA were included. Age, gender, zip code, and vaccination status were requested from each participant. Participants submitted saliva and nasal swab specimens and reported symptoms from: fever, cough, sore throat, shortness of breath, chills, fatigue, body aches, headache, nausea, and diarrhea. Pathogens were confirmed via RT-PCR on a GenMark respiratory panel assay (full virus list reported previously 3 ). Observations with missing, invalid or equivocal lab tests were removed. Table 2 summarizes the binary features. Age categories were: ≤ 20, > 20 and < 40, and ≥ 40 to represent young, middle- aged, and old. Missing age and gender values were imputed based on overall distributions. Three machine learning algorithms—Support Vector Machines (SVMs) 4 , Random Forests (RFs) 5 , and Logistic Regression (LR) were considered. Both individual features and their combinations were assessed. Outcome was the presence (1) or absence (0) of laboratory diagnosis of ARI. Results Ten-fold cross validation was repeated ten times. Evaluations metrics used were: positive predictive value (PPV), negative predictive value (NPV), sensitivity, and specificity 6 . LR and SVMs yielded the best PPV of 0.64 (standard deviation: ± 0.08) with cough and fever as predictors. The best sensitivity of 0.59 ( ± 0.14) was from LR using cough, fever, and sore throat. RFs had the best NPV and specificity of 0.62 ( ± 0.15) and 0.83 ( ± 0.10) respectively with the CDC ILI symptom profile of fever and (cough or sore throat). Adding demographics and vaccination status did not improve performance of the classifiers. Results are consistent with studies using clinically- sourced data: cough and fever together were found to be the best predictors of flu-like illness 1 . Because our data include mildly infectious and asymptomatic cases, the classifier sensitivity and PPV are low compared to results from clinical data. Conclusions Evidence of fever and cough together are good predictors of ARI in the community, but clinical data may overestimate this due to sampling bias. Integration of participatory data can not only improve population health by actively engaging the general public 2 but also improve the scope of studies solely based on clinically-sourced surveillance data. Table 1. Details of included participants. Table 2. Coding of binary features