The Proceedings of The 17th Annual IEEE Systems Conference
19th April, 2023
In our study, we show that machine learning models deployed on microcontrollers can successfully analyze
ECG signals in real-time for Sleep Apnea detection. We have
created TinyML models using TensorFlow Lite which we have
deployed on cost effective and resource constrained devices like
the Raspberry Pi Pico and ESP32. Our setup has given results
comparable to more advanced and expensive devices for the
detection of Sleep Apnea using ECG signals.We have reduced the cost,
complexity and accessibility of Sleep Apnea detection.
The Proceedings of International Conference on Smart Applications, Communications and Networking
25th - 27th July, 2023
In our research, we have successfully developed a comprehensive and self-contained system that captures ECG signals using
ECG electrodes, AD8232 for collection of ECG signals. The system includes signal processing, sleep apnea detection using Deep Learning models, and displaying the
results on an OLED screen. Our study compares the performance of different micro-controllers with varying capacities and
costs: Raspberry Pi Pico, ESP32, and Arduino Nano BLE Sense. We have achieved comparable accuracies to those of
expensive industrial devices at a significantly lower cost. This work lays the foundation for the development of wearable
devices capable of detecting sleep apnea without the need for internet connectivity or data transmission. This work can be extended for any other
body signals like EEG.
Epilepsy, a chronic neurological disorder characterized by abnormal brain electrical activity, affects 50 million people worldwide, necessitating efficient seizure detection methods. Traditional manual EEG signal analysis is time-consuming and error-prone, prompting the adoption of machine learning and deep learning for automation.
Our novel approach combines 1-D Convolutional layers with Bidirectional LSTM and GRU, alongside Average pooling, for feature extraction. Tested on the Bonn dataset with five-fold cross-validation, our model achieves high accuracy, sensitivity, and specificity.
Our model demonstrates 99–100% accuracy for binary seizure classification, 97.2%–99.2% for normal-interictal-seizure classification, and 95.81%–98% for five-class classification.
Our model significantly improves performance metrics for binary and multiclass classifications, showcasing its effectiveness in accurately detecting epileptic seizures across varying EEG signal lengths. This suggests its potential as a reliable tool for automated seizure detection, enhancing epilepsy diagnosis and management.