A flexible wireless stethoscope built on a 25-element AlN PMUT array captures cardiopulmonary sounds from 10 Hz to 10 kHz and classifies five respiratory states with 98.7% accuracy

The work presents and fabricates a flexible wireless stethoscope based on a 25-element circular AlN PMUT array, achieving a measured packaged sensitivity of −167.5 dB, an operating bandwidth of 10 Hz–10 kHz, a frequency-response flatness of ±0.5 dB, and a signal-to-noise ratio of 40.09 dB. Compared with traditional dynamic or electret microphones, the PMUT is more compact (2 × 2 mm²), lighter (0.88 mg), and offers a broader frequency response range; the authors used COMSOL finite-element simulations to compare PZT versus AlN materials and square versus circular geometries, ultimately selecting the AlN circular structure because of its higher d31/εr ratio and more uniform stress distribution. Device performance was validated through finite-element simulation, impedance analysis (resonant frequency 1.04 MHz), SEM, EDS, and XRD characterization; the 40.09 dB SNR comes from Supplementary Fig. S3.

In human-subject measurements, the flexible stethoscope acquired heart sounds at five standard auscultation sites, with S1 and S2 features temporally corresponding to the R and T peaks of reference ECG, and continuously tracked heart rate during dynamic activities including walking, stair climbing, and standing. Compared with a rigid-encapsulated stethoscope, the flexible device clearly captured S1 and S2 peaks during walking while the rigid device generated significant noise and motion artifacts; the system is 34 times smaller in volume and 18 times lighter than traditional commercial stethoscopes. Tested on six participants; over a 16-s window both heart-sound methods recorded 20 cycles, and chest motion aligned with inhalation and exhalation phases of bronchial respiratory sounds; during dynamic activities heart rate was 102–106 bpm while walking, 116–145 bpm while climbing stairs, and decreased from 130 to 98 bpm while standing.

Coupling the sensing platform with a residual neural network enabled automated classification of five respiratory states (awake, asleep, apnea, rhonchi, and wheeze) with an accuracy of 98.7%. The network comprises 53 convolutional layers and one fully connected layer, built on transfer learning; one-dimensional time-series signals were converted into two-dimensional spectrograms via short-time Fourier transform before classification, and t-SNE visualization demonstrated clear clustering of feature vectors. Data came from a single participant's continuous measurements, with 70% used for training and 30% for validation; test data from a different session were used exclusively for evaluation; the authors explicitly position this in the conclusion as a 'within-participant, session-separated proof-of-concept evaluation.'

The results are intended for researchers and developers who need continuous cardiopulmonary acoustic signal acquisition during daily activities, applicable to scenarios such as home sleep monitoring, exercise heart-rate tracking, and remote healthcare. The system uses standard MEMS fabrication and flexible packaging, and the authors position it as a scalable, low-cost framework that provides a device and algorithm foundation for subsequent validation in larger, subject-independent datasets and clinically diagnosed populations.

The authors explicitly state in the conclusion that further studies using larger subject-independent datasets, clinically diagnosed participants, and systematic evaluations of battery endurance, wireless reliability, and long-term mechanical durability are required before clinical application. The 98.7% respiratory-state classification accuracy comes from a single-participant, session-separated proof-of-concept evaluation, and its performance in broader populations and real clinical settings remains to be validated. Additionally, this is a full-text parse; the specific contents of Supplementary Figs. S1–S14 and Supplementary Tables S1–S5 are not included in the loaded text, so some device details and comparison data cannot be fully presented here.