Audio-IMU multimodal cough dataset using wearables

Cough detection is essential for long-term respiratory illness monitoring, but clinical methods are not feasible for home use. Wearable devices offer a convenient alternative, but challenges include data limitation and accurately detecting coughs in real-world environments, where audio quality may be compromised by background noise. This multimodal dataset, collected in a controlled lab setting, includes IMU and audio data captured using wearable devices. It was designed to support the development of an accessible and effective cough detection system. The dataset documentation includes details on sensor arrangement, data collection protocol, and processing methods.

Our analysis reveals that integrating transfer learning, multimodal approaches, and out-of-distribution (OOD) detection significantly enhances system performance. Without OOD inputs, the model achieves accuracies of 92.59% in the in-subject setting and 90.79% in the cross-subject setting. Even with OOD inputs, the system maintains high accuracies of 91.97% and 90.31%, respectively, by employing OOD detection techniques, despite the OOD inputs being double the number of in-distribution (ID) inputs. These results are promising for developing a more efficient and user-friendly cough and speech detection system suitable for wearable technology.

Methods
This dataset comprises recordings from 13 participants, collected as part of a study approved by NC State University IRB Protocol 25003. The participants, who are student volunteers of similar age and health condition, engaged in a series of physical activities, including sitting, walking, running, and transitions between these states, each lasting approximately two minutes with 30-second resting intervals between transitions.

Data Collection Overview:

Audio Recording: Two chest-mounted microphones were used to capture audio data. One microphone faced away from the participant (out-microphone), and the other faced toward the participant (in-microphone). The microphones were housed in a custom-designed enclosure and sourced from Tozo T10 Bluetooth earbuds with the speaker circuit disconnected.

IMU Data: The participant's movement was recorded using a MetaMotionS r1 sensor from Mbientlab, mounted on the chest to capture 9-axis Inertial Measurement Unit (IMU) data.

Synchronization:
At the start of each recording session, participants clapped three times, which serves as a synchronization point across different data modalities. These claps are clearly identifiable in both the audio and IMU signals, allowing for precise alignment of the datasets.