Page 71 - 2024F

64 UEC Int’l Mini-Conference No.53

Prevention of Aspiration Pneumonia Using in Ear­Audio

Filip PETTERSSON, Takuji KOIKE
Department of Mechanical and Intelligent Systems
University of Electro­communication

1.0 Introduction 2.2 Network architecture
• Aspiration pneumonia predominantly affects elderly
people with impaired swallowing reflexes, occurring YamNet for feature extraction to reduce trainable
when food particles enter the lungs due to poor parameters.

chewing and swallowing function [1].
1024­Dimensional Input Layer
• Current prevention methods focus on modifying food LSTM (128 units)
textures which reduces aspiration risk but reduces meal Dense (64, ReLu)
Dense (32, ReLu)
enjoyment and quality of life [2].
Dense (1, Sigmoid)
• Can in­ear audio monitoring assess food state
to balance patient safety and dietery experience? 2.3 Training
2.0 Methodology MSE loss function
Optimizer: Adam η=0.0005
The solution must handle the high complexity data, Batch size: 32
understand the temporal patterns within, and between Training validation split:80/20
bites while only training on a small dataset and run in Custom batch generator
real time.
0
.25
.75
1
0­1

2.1 Data collection 3.0 Results
• Collected in ear­audio from four young, healty
individuals chewing brittle foods • 6% mean average error which is also equal to a
• Segmented audio into single bites with labels for mean average error of 1.22 bites (fig 2)
chewing progression using envelope detection • Training time < 60 s
• Identified faulty segmentation using t­SNE visualization
(fig 1)

Number of segments: 18
Bites per segment: 21.22 ± 4.64
Total bites: 382
Bite segment length: 0.96 s

Figure 2: Loss on training- and validation data during
training
4.0 Future work
• Train and validate on broader range of foods
• Validate on unhealty indivduals
• Hyperparameter tuning

• Robust bite detection capable of running in real time in
noisy environment
Figure 1: t-SNE Visualization of Chewing Embeddings
Colored by Position in Sequence

[1] Logemann, J. A. (2007). Swallowing disorders. Best Practice & Research Clinical Gastroenterology, 21(4), 563­573.
[2] Cichero, J. A., et al. (2017). Development of international terminology and definitions for texture­modified foods and thickened fluids used in dysphagia
management: The IDDSI Framework. Dysphagia, 32(2), 293­314.