Supplementary Material for: Continuous Sound Collection Using Smartphones and Machine Learning to Measure Cough

Background: Despite the efforts of research groups to develop and implement at least partial automation, cough counting remains impractical. Analysis of 24-h cough frequency is an established regulatory endpoint which, if addressed in an automated manner, has the potential to ease cough symptom evaluation over multiple 24-h periods in a patient-centric way, supporting the development of novel treatments for chronic cough, an unmet clinical need. Objectives: In light of recent technological advancements, we propose a system based on the use of smartphones for objective continuous sound collection, suitable for automated cough detection and analysis. Two capabilities were identified as necessary for naturalistic cough assessment: (1) recording sound in a continuous manner (sound collection), and (2) detection of coughs from the recorded sound (cough detection). Methods: This work did not involve any human subject testing or trials. For sound collection, we designed, built, and verified technical parameters of a smartphone application for sound collection. Our cough detection work describes the development of a mathematical model for sound analysis and cough identification. Performance of the model was compared to previously published results of commercially available solutions and to human raters. The compared solutions use the following methods to automatically or semi-automatically assess cough: 24-h sound recording with an ambulatory device with multiple microphones, automatic silence removal, and manual recording review for cough count. Results: Sound collection: the application demonstrated the ability to continuously record sounds using the phone’s internal microphone; the technical verification informed the configuration of the technical and user experience parameters. Cough detection: our cough recognition sensitivity to cough as determined by human listeners was 90 at 99.5% specificity preset and 75 at 99.9% specificity preset for a dataset created from publicly available data. Conclusions: Sound collection: the application reliably collects sound data and uploads them securely to a remote server for subsequent analysis; the developed sound data collection application is a critical first step toward future incorporation in clinical trials. Cough detection: initial experiments with cough detection techniques yielded encouraging results for application to patient-collected data from future studies.

背景：尽管诸多研究团队已致力于开发并实现部分自动化流程，但咳嗽计数仍难以达到实用化标准。24小时咳嗽频率分析是一项成熟的监管终点指标（regulatory endpoint），若能以自动化方式实现该分析，便有望以患者为中心的模式，简化多周期24小时咳嗽症状评估工作，为慢性咳嗽（chronic cough）的新型治疗方案开发提供支撑——而慢性咳嗽至今仍存在未被满足的临床需求（unmet clinical need）。

研究目标：鉴于近期的技术进展，我们提出一种基于智能手机的系统，用于客观连续采集声音信号，适配自动化咳嗽检测与分析任务。针对自然场景下的咳嗽评估，我们确定了两项核心必要功能：(1) 持续录制声音（声音采集，sound collection）；(2) 从录制的音频中检测咳嗽事件（咳嗽检测，cough detection）。

研究方法：本研究未涉及任何人体受试者测试或临床试验。在声音采集环节，我们设计、开发并验证了一款用于声音采集的智能手机应用程序的技术参数。在咳嗽检测方向，本工作阐述了用于声音分析与咳嗽识别的数学模型的开发流程。将该模型的性能与已发表的商用解决方案及人工评分者的结果进行了对比。本次对比涉及的方案采用以下方式实现自动化或半自动化咳嗽评估：使用搭载多麦克风的移动式便携设备进行24小时声音录制、自动移除静音片段，以及人工审核录音以统计咳嗽次数。

研究结果：声音采集：该应用程序可通过手机内置麦克风实现持续声音录制；技术验证工作明确了技术参数与用户体验参数的最优配置方案。咳嗽检测：针对由公开数据集构建的测试集，在预设99.5%特异性（specificity）的条件下，我们的咳嗽识别模型对人工标注咳嗽的灵敏度（sensitivity）为90%；在预设99.9%特异性的条件下，该模型的灵敏度为75%。

研究结论：声音采集：该应用程序可稳定采集声音数据并将其安全上传至远程服务器以供后续分析；所开发的声音采集应用程序是未来将其整合入临床试验的关键第一步。咳嗽检测：针对咳嗽检测技术的初步实验结果令人振奋，为其应用于未来研究中患者自行采集的数据提供了可行性支撑。