Design and Development of A-vent: A Low-Cost Ventilator with Cost-Effective Mobile Cloud Caching and Embedded Machine Learning

• Main Authors: Cabacungan, Paul M., Oppus, Carlos M., Cabacungan, Nerissa G., Marmadlo, John Paul A., Santiago, Paul Ryan A., Mercado, Neil Angelo M, Faustino, E. Vincent S., Tangonan, Gregory L
• Format: text
• Published: Archīum Ateneo 2021
• Subjects: training; vibrations; ventilators; buildings; prototypes; lung; machine learning; COVID-19; breathing device; hospital informatics; critical care; internet of medical things; near cloud; Analytical, Diagnostic and Therapeutic Techniques and Equipment; Artificial Intelligence and Robotics; Computer Sciences; Diagnosis; Health Information Technology; Respiratory Therapy
• Online Access: https://archium.ateneo.edu/ecce-faculty-pubs/95; https://ieeexplore.ieee.org/document/9550920
• Institution: Ateneo De Manila University

We designed and developed a low-cost mechanical ventilator prototype that meets the government's minimum viable standards. We substituted alternative off-the-shelf food-grade for the medical-grade parts and improvised some components for our prototype. We cleaned the air from the oxygen tanks and compressors before going to the lung test bag. We designed a solar-powered battery system that can run electronic components for a fail-safe operation. We demonstrated how the AIC Near Cloud system can store air flow rate and air pressure data which were generated during the prototype's operation. We used Embedded Machine Learning in sensors and data processing by using flow and pressure sensors to provide accumulated data that can be utilized in training the machine learning software. The patient-ventilator asynchrony detection model was tested using data generated from the emulated ventilator waveform events that mimic the patient-ventilator asynchrony. A different compression pattern was applied to the test lung and results showed the training, validation, and model testing that yielded 98.7%, 99.1%, and 97.18 percent accuracy, respectively. Having demonstrated that the Tiny ML can be trained to detect anomalies from several data points, we realized the feasibility of detecting ventilator patient vibration anomaly, and unusual acoustic signatures, among others, for future works.