Page 70 - 2024F

UEC Int’l Mini-Conference No.53 63

Prevention of Aspiration Pneumonia using In Ear-Audio

1
Filip Hugo PETTERSSON and Takuji KOIKE 2
1 UEC Exchange Study Program (JUSST Program)
2 Department of Mechanical and Intelligent Systems
The University of Electro-Communications, Tokyo, Japan

Keywords: Aspiration Pneumonia, Mastication Monitoring, Deep Neural Net, Transfer learning, Time-
series data, Temporal Window

Abstract
Aspiration pneumonia poses a significant health risk for elderly individuals with compromised swal-
lowing reflexes, traditionally managed through texture modification of foods and liquids. While effective,
these interventions often diminish patients’ quality of life and dietary satisfaction. This study explores
an approach using in-ear audio monitoring by deep learning, to assess food state during mastication, po-
tentially enabling safer consumption of regular-textured foods. Audio data collected from four healthy
individuals yielded 382 chewing strokes, each 0.96 seconds in duration at 16kHz. Using YAMNet for
feature extraction, we developed a deep neural network architecture (128 LSTM - 64 Dense - 32 Dense
- 1 Dense) to generate a chewing progression score from 0 to 1. The model achieved a mean absolute
error of 0.061, equivalent to approximately 1.22 chewing strokes, with subjects averaging 21.22 ± 4.64
strokes per food segment. These results demonstrate that in-ear audio monitoring, combined with
transfer learning via YAMNet, can accurately assess food state during mastication while requiring min-
imal training data, potentially offering a less restrictive alternative to traditional aspiration prevention
methods.