Silent Aspiration Detection

Welcome and enjoy my video!

The video above provides an explanation of my project.

Quick Note:

I had said 60% in the video, but it has increased to 80%

Introduction:

Silent aspiration is an extreme form of dysphagia where food or liquid enters the airway without triggering a protective cough reflex. Due to the lack of visible symptoms, it often goes undetected until serious complications arise. Although silent aspiration itself is not the same as aspiration pneumonia, it is a major contributing factor to its development and associated morbidity in high-risk populations (Centers for Disease Control and Prevention, n.d.). Silent aspiration affects up to approximately fifty percent of stroke patients and carries mortality rates of roughly twenty to thirty percent in severe cases (National Institutes of Health, n.d.) (Figure 1).

Inspiration:

I created this project because a relative had fibrosis, which caused major throat vibrations, and I wondered if these vibrations could act as an early warning system. Through research, I discovered the deadly condition of silent aspiration (Figure 2).

Benefit:

What makes this issue especially urgent is that current detection methods are reactive rather than preventative. Clinics rely on imaging and short-term observation, resulting in patients being monitored only in controlled environments, rather than in their daily lives (Figure 3).

Question:

How can we detect a condition that produces no visible warning signs?

Value:

I developed a wearable multi-sensor system that captures vibration and acoustic signals from the throat. By analyzing post-swallow signal behaviour and vibrations, the machine learning system identifies patterns associated with potential aspiration risk with 80% accuracy using over 30,000 high-quality sensor samples gathered through experimental data collection.

Biomechanical Basis of Swallowing Signals

The development of this system began by identifying how swallowing produces measurable signals. Medical research shows that during the swallowing process, the larynx elevates rapidly, surrounding muscles contract, and the airway closes. These coordinated movements create short-duration mechanical vibrations at the surface of the neck, which can be captured using contact-based sensors (Matsuo and Palmer, 2008) (Figure 4).

Sensor Design and Placement Strategy

To measure these signals, a wearable sensing system was constructed using piezoelectric transducers placed directly over the laryngeal region. Two sensors were positioned on the same side of the Adam’s apple to capture vertical displacement and primary muscle contraction, while a third sensor was placed laterally to detect asymmetrical or secondary motion. This multi-point configuration was selected to improve sensitivity to irregular swallowing patterns (Sejdic et al., 2013) (Figure 5).

Signal Conditioning and Hardware

The sensor outputs were connected to an analog conditioning stage built around an MCP6002 operational amplifier. This stage stabilized the signal, reduced noise, and improved the clarity of low-amplitude vibrations before sampling. The conditioned signals were read by an Arduino Nano 33 BLE Rev2, which performed analog-to-digital conversion and transmitted the data for analysis (Microchip Technology Inc., n.d.; Arduino, n.d.) (Figure 6).

Data Collection and Experimental Control

Testing was conducted through repeated trials under controlled conditions. A total of 860 samples were recorded from participants aged 16 to 61 years. Sensor placement, posture, and environmental noise were kept constant across trials to ensure signal variation reflected physiological activity.

Feature Extraction and Machine Learning Processing

Recorded signals were segmented into short intervals and transformed into features, including amplitude, duration, and signal energy. These features were used to train and refine a machine learning model, improving classification of normal and abnormal patterns (Breiman, 2001) (Figure 7).

How the Prototype Works

The prototype captures biometric vibration data produced by swallowing. During a normal swallow, the larynx elevates ~2-3 cm, the epiglottis inverts to protect the airway, and pharyngeal muscles contract in a rapid sequence. These events occur within roughly 1-2 seconds and generate distinct mechanical vibrations at the skin surface (Matsuo and Palmer, 2008; Dodds, 1989). Piezoelectric sensors convert these deformations into voltage: rapid, coordinated motion produces short, high-amplitude spikes, while slower or incomplete motion produces lower or irregular signals. The microcontroller digitizes these signals and transmits them wirelessly for classification (Figure 8).

Sensor Placement Strategy

Sensor placement was designed using principles from cervical auscultation, a clinical technique that analyzes neck vibrations to assess swallowing (Cichero and Murdoch, 2006). Two sensors were positioned directly over the laryngeal prominence (Adam’s apple) on the same side to capture vertical laryngeal elevation and airway-closure dynamics. A third sensor was placed laterally (to the right of the larynx) to capture asymmetrical and lateral tissue motion, which can occur in dysphagia when swallowing coordination is impaired.

This multi-point configuration increases sensitivity to both primary swallow mechanics and secondary irregularities, improving detection compared to single-sensor setups. Studies on cervical vibration and swallowing acoustics show that sensor location strongly influences signal quality and diagnostic value, supporting this placement strategy (Sejdic et al., 2013; Lee et al., 2008) (Figure 9).

Key Observations

Breathing: low-amplitude, stable baseline signals

Swallowing: repeatable, short-duration high-amplitude peaks linked to laryngeal elevation

Speech: irregular patterns that overlap with swallowing due to airflow and tissue vibration

Abnormal events: post-swallow irregular signals that do not return to baseline

Clinically, aspiration can occur after the swallow without a cough reflex, which aligns with the observation that post-swallow behaviour is more informative than the swallow peak itself (Marik, 2001) (Figure 10).

Why These Methods Were Chosen

Feature extraction (amplitude, duration, energy) was used because summarizing characteristics improves model reliability. Accuracy was selected as the evaluation metric to compare classification before. The shift from activity classes to broader categories reduced overlap and improved interpretability, reflecting how clinical decisions prioritize risk vs. non-risk activity labelling.

What Was Learned

Swallowing generates consistent, measurable biomechanical signatures at the neck

Sensor placement directly affects signal quality and model performance

Raw signals are insufficient; feature engineering is essential

The most critical indicator of risk is post-swallow signal behaviour, consistent with clinical understanding of silent aspiration (Figure 11)

Conclusion from Results

The system establishes a strong proof of concept: silent, symptomless swallowing abnormalities can be detected through external vibration sensing. These findings support the development of wearable systems capable of continuous, real-world monitoring, extending detection beyond hospital-based imaging methods.

Main Results

The system demonstrated that swallowing-related risk can be detected using non-invasive vibration sensing and machine learning. From 860 samples ages 16-61, model performance improved from 25% to 80% accuracy, after feature engineering and task simplification. This confirms that throat vibration signals contain classifiable information, but require structured processing to separate overlapping patterns (Figure 12).

Conclusion

This research explores the usage of a non-invasive wearable device that uses vibration-based sensors to identify swallowing-related risk patterns. Utilizing a dataset of 860 samples from participants aged 16-61, this project uses raw waveform analysis to refine feature extraction for classifying biomechanical signals to detect symptoms of Silent Aspiration.

Methodology and Iteration

Initial Model: A multi-class approach using activities rather than unique classifications yielded a 25% accuracy. The failure highlighted significant signal overlap, particularly between swallowing and vocalization, confirming that raw biological data is too complex for simple classification (Figure 13).

Refined Model: By restructuring the problem into broader categories and implementing targeted feature extraction (amplitude, duration, and energy), the accuracy rose to 80%. This represents a 220% improvement, validating the necessity of using feature engineering within noisy biomedical environments (Figure 14).

Key Findings

Biomechanical Signatures: Swallowing produces distinct, classifiable signals at the laryngeal surface (Figure 15).

Silent Aspiration Indicator: Post-swallow signal behaviour emerged as a critical marker for abnormal events, providing a potential digital proxy for silent aspiration.

Significance and Clinical Impact

While 80% accuracy is below the threshold for clinical diagnostics, this project successfully establishes a proof of concept for wearable aspiration monitoring.

Proactive Care: The system detects risk patterns without requiring observable symptoms like coughing (Figure 16).

Future Evolution: With improved sensor calibration and larger datasets, this technology could shift dysphagia management from reactive, hospital-based imaging to continuous, real-time remote monitoring, significantly enhancing safety for high-risk populations.

With over 300,000 raw sensor data collected, the next phase focuses on improving reliability and real-world performance, as well as expanding the dataset. Additionally, we will be testing the system across more varied, uncontrolled conditions to evaluate performance during movement, long-term wear, and environmental noise (Figure 17).

Hardware refinement will include miniaturizing the system into a more compact and stable wearable design. Additionally, further improvements to signal consistency and model performance will be explored to increase accuracy beyond 80%.

Long-term development will focus on clinical validation and adapting the system for practical use in continuous monitoring of high-risk individuals.

Special thanks to:

Dr. Ken Wilson, the Dean, and the Faculty of Science, Ontario Tech University, for sponsoring me and providing me with the opportunity to go to the Canada-Wide Science Fair.

Ms. Olaveson, the Coordinator for the Durham Regional Science Fair, and Ms. Jennifer Mackinnon, our CWSF delegate, guided me through the process of completing the project board and registration.

The judges at the Durham Regional Science Fair provided feedback and constructive criticism about my project.

My teachers at Pickering High School provided support, feedback, science knowledge, and assistance with formatting throughout the completion of this project.

My parents for their continuous support, encouragement, and guidance throughout the process.

Silent Aspiration, Dysphagia, and Clinical Impact

American Speech-Language-Hearing Association. (n.d.). Adult dysphagia practice portal.

https://www.asha.org/practice-portal/clinical-topics/adult-dysphagia/

Centers for Disease Control and Prevention. (n.d.). Pneumonia and aspiration-related conditions.

https://www.cdc.gov/

Langmore, S. E., et al. (1998). Predictors of aspiration pneumonia. Dysphagia, 13(2), 69–81.

https://doi.org/10.1007/PL00009559

Marik, P. E. (2001). Aspiration pneumonitis and aspiration pneumonia. New England Journal of Medicine, 344(9), 665–671.

https://doi.org/10.1056/NEJM200103013440908

National Institutes of Health. (n.d.). Dysphagia and swallowing disorders.

https://www.ncbi.nlm.nih.gov/

Swallowing Biomechanics

Dodds, W. J. (1989). The physiology of swallowing. Dysphagia, 3(4), 171–178.

https://doi.org/10.1007/BF02407376

Logemann, J. A. (1998). Evaluation and treatment of swallowing disorders. PRO-ED.

Matsuo, K., & Palmer, J. B. (2008). Anatomy and physiology of feeding and swallowing. Physical Medicine and Rehabilitation Clinics of North America, 19(4), 691–707.

https://doi.org/10.1016/j.pmr.2008.06.001

Cervical Auscultation and Sensor Placement

Cichero, J. A. Y., & Murdoch, B. E. (2006). Dysphagia: Foundation, theory and practice. Wiley.

Lee, J., Blain, S., Casas, M., Kenny, D. J., & Berall, G. (2008). Automatic detection of aspiration using vibration signals. IEEE Transactions on Biomedical Engineering, 55(2), 651–658.

https://doi.org/10.1109/TBME.2007.903518

Sejdic, E., Steele, C. M., & Chau, T. (2013). Segmentation of swallowing accelerometry signals. IEEE Transactions on Biomedical Engineering, 60(10), 2789–2797.

https://doi.org/10.1109/TBME.2013.2263120

Sensors, Electronics, and Wiring

Arduino. (n.d.). Arduino Nano 33 BLE Rev2 documentation.

https://docs.arduino.cc/

Horowitz, P., & Hill, W. (2015). The art of electronics (3rd ed.). Cambridge University Press.

Make: Magazine. (n.d.). Basic electronics: resistors, capacitors, and op-amps explained.

https://makezine.com/

Microchip Technology Inc. (n.d.). MCP6002 operational amplifier datasheet.

https://www.microchip.com/

SparkFun Electronics. (n.d.). Piezo vibration sensor hookup guide.

https://learn.sparkfun.com/

TE Connectivity. (n.d.). Piezo film sensors technical manual.

https://www.te.com/

Signal Processing

Oppenheim, A. V., & Schafer, R. W. (2009). Discrete-time signal processing (3rd ed.). Pearson(106).

Smith, S. W. (1997). The scientist and engineer’s guide to digital signal processing. California Technical Publishing.

Machine Learning

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.

https://doi.org/10.1023/A:1010933404324

Geron, A. (2019). Hands-on machine learning with Scikit-Learn, Keras, and TensorFlow (2nd ed.). O’Reilly.

Pedregosa, F., et al. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830.

Cover Image Generation

Google. (2026). Gemini 3 Flash (April 29 version) [Large language model]. https://gemini.google.com/