tinyML Asia 2021 Video Poster: Fixed complexity tiny reservoir heterogeneous network for…

tinyML Asia 2021
Fixed complexity tiny reservoir heterogeneous network for on-device ECG learning of anomalies
Danilo PAU, Technical Director, IEEE and ST Fellow, STMicroelectronics Italia

The electrocardiogram (ECG), being one of the most extensively used signals to monitor cardiovascular diseases (CVDs), captures the heart’s arrhythmias. Patients with such pathology are often monitored for extended periods of time, requiring data storage, and a very time-consuming off-line search of anomalies. This is especially inefficient when indicative patterns in the biological signals are infrequent, requiring more analysis time of medical doctors, and entailing a difficult visual search task for the diagnosis.

We propose an automated hybrid deep learning and machine learning pipeline based on reservoir computing (RC), followed by principal component analysis (PCA) and one-class support vector machine (OC-SVM). This machine learning pipeline can be used to perform on device personalized learning and real-time anomaly detection of pathological conditions and therefore enable an application to raise warnings.

The on-device learning step requires fixed computational complexity, latency and memory to fit into an off-the-shelf low-power microcontroller (MCU). During the learning phase, it uses a very limited amount of normal input data, e.g., 10,000 bytes, which makes this work suitable for fast personalization every time device restart is required also due to a change in carry position. The detection accuracy has been evaluated on the publicly available MIT-BIH arrhythmia dataset. This dataset originally segmented into individual heartbeats, was also modified to mimic temporal sliding of input tensors on the ECG streaming data. Best F1 score and accuracy are 91.5%, 95.4% respectively, with variance over the processed data of 0.05. On MCU, the learning can run within a latency of 83 seconds at 360Hz sampling frequency and at initialization time and achieves: 43 seconds on device learning, 2 inferences per second (estimated through multiply and accumulate operations) on STM32 M4 at 80MHz. While on STM32 M7, 480MHz, the learning takes less than 5 seconds and achieves 19 inferences per second.

We concluded that recurrent reservoir networks combined with PCA, OCSVM machine learning modules achieve adequate performances both in terms of detection accuracy and execution time, that are competitive to those obtained by more complex models like LSTM based via backpropagation methods (requiring up to 13 minutes), while requiring only few seconds to be trained online.

This makes this work suitable for on-device learning with fast personalization for real-time embedded applicat


44 thoughts on “tinyML Asia 2021 Video Poster: Fixed complexity tiny reservoir heterogeneous network for…

Leave a Reply

Your email address will not be published.