Electromyogram and Sound of Swallowing Events

Dehydration is a prevalent and serious problem among older adults. As people age, the body’s ability to conserve water decreases, and the sensation of thirst diminishes. This reduced sense of thirst makes it easy for older adults to become dehydrated without realizing it. Monitoring fluid intake is therefore essential for maintaining proper hydration for overall health. The main objectives of this study are to investigate whether transfer learning can enhance classification performance compared to less complex CNN models in detecting swallowing and drinking types, to compare the performance of EMG and sounds in detecting swallowing from non-swallowing events, and to explore if the downsampling will affect the classification performance. In the methodology, two transfer learning models ResNet-18 and ResNet-9 along with proposed convolutional neural network (CNN) architectures (from one to nine layers) were trained on spectrogram images derived from swallowing and non-swallowing signals collected via surface electromyography (sEMG) and a microphone. Eleven individuals (eight females and three males) participated in the study. The results showed that the ensemble F-score for classifying swallowing from non-swallowing events using ResNet was 85% and 95% using the EMG and Microphone signals, respectively. <b>Experimental Procedure:</b>The experiment consisted of a 90-minute session for two types of signals, EMG and acoustic signals, which were captured concurrently during the experiment. Two Delsys Trigno sEMG sensors (Natick, MA, USA) were utilised to capture sEMG data. The sEMG signals were analogue filtered between 10 and 850 Hz and sampled at 2.2 kHz. One EMG sensor was placed on one side of the sternohyoid muscles’ belly, part of the infrahyoid group, chosen for their superficial location, and the microphone was placed on the other side. The use of two EMG sensors was informed by our previous studies. To capture the acoustic signals of the swallowing data, we used the RØDE SmartLav+ Smartphone Lavalier Microphone. Acoustic signals were analogue filtered between 100 and 10000 Hz and sampled at 44.1 kHz. The microphone was placed on the right side of the sternohyoid muscles’ belly, part of the infrahyoid group. Participants were seated comfortably, and their neck area was cleaned with alcohol wipes. The sensor’s placement was determined by palpating the relevant swallowing muscles. After proper anatomical positioning of the sensors, participants were instructed to perform nine different tasks in a random order for each session. The first task required the participants to talk while recording. The second task involved coughing, while the third and fourth tasks involved swallowing saliva and solid food. Participants were given chocolate chip cookies and instructed to take one bite at a time for the solid food task. Tasks five through nine involved swallowing water from a cup in single sips, with the water volume increasing by 5 mL for each subsequent task, starting at 5 mL for the fifth task and reaching 25 mL by the ninth task. A needleless syringe with markings was used to ensure accurate measurement. Participants followed verbal instructions to perform these tasks, and their actions were recorded for analysis and further evaluation.

脱水是老年人群中普遍且严重的健康问题。随着年龄增长，人体保水能力下降，口渴感知也会减退。这种口渴感知的减退使得老年人极易在未察觉的情况下陷入脱水状态，因此监测液体摄入对于维持恰当的体液水平以保障整体健康至关重要。

本研究的核心目标如下：其一，探究相较于结构更简洁的卷积神经网络（Convolutional Neural Network, CNN）模型，迁移学习（Transfer Learning）能否提升吞咽与饮食物类别的分类性能；其二，对比肌电（Electromyography, EMG）信号与音频信号在区分吞咽事件与非吞咽事件中的表现；其三，探索降采样操作是否会对分类性能产生影响。

在实验方法层面，本研究针对通过表面肌电（surface electromyography, sEMG）与麦克风采集的吞咽及非吞咽信号生成的语谱图图像，对两种迁移学习模型ResNet-18、ResNet-9，以及本研究提出的1至9层卷积神经网络架构开展训练。本研究共招募11名受试者（8名女性，3名男性）参与。实验结果显示，基于ResNet模型区分吞咽与非吞咽事件的集成F1分数（F-score）中，使用肌电信号时为85%，使用麦克风音频信号时则为95%。

<b>实验流程：</b>
本次实验单次时长为90分钟，同步采集肌电与声学两类信号。本研究采用两台Delsys Trigno表面肌电传感器（产自美国马萨诸塞州内蒂克市）采集肌电数据，信号经10~850 Hz的模拟滤波后，以2.2 kHz的采样率进行数字化采样。一台肌电传感器放置于舌骨下肌群中的胸骨舌骨肌肌腹的一侧（该肌肉位置表浅，故选为采集位点），麦克风则置于胸骨舌骨肌肌腹的右侧（该肌肉属于舌骨下肌群）。采用双肌电传感器的方案参考了本团队此前的研究成果。

本研究使用RØDE SmartLav+智能手机领夹式麦克风采集吞咽相关的声学信号，该信号经100~10000 Hz的模拟滤波后，以44.1 kHz的采样率进行数字化采样。受试者取舒适坐姿，先用酒精棉片清洁颈部区域，通过触诊相关吞咽肌肉确定传感器的放置位点。完成传感器的精准解剖学定位后，受试者需按随机顺序完成9项不同任务，每个实验时段均遵循此流程。第一项任务为边录音边讲话；第二项任务为咳嗽；第三、四项任务分别为吞咽唾液与吞咽固体食物，本次实验提供巧克力曲奇作为固体食物样本，要求受试者每次仅取一口进食。第五至第九项任务为单次小口饮用杯装水，每次饮水量较前一项增加5 mL：第五项任务饮水量为5 mL，至第九项任务时增至25 mL。本研究使用带刻度的无针注射器以保证饮水量的精准计量。受试者按照口头指令完成各项任务，实验过程全程记录，以供后续分析与评估。