Reduced complexity on micro-controller learning of ECG anomalies
Danilo Pietro Pau, Norhen Abdennadher · 2022
An electrocardiograph (ECG) is used as a key diagnostic signal for cardiovascular diseases (CVDs) [1]. The acquired ECG time series (TS) signals are often manually evaluated by qualified medical doctors in order to discover any arrhythmia that the patient may have experienced. Much effort has been put into automatizing the process of evaluating ECG readings to learn the various forms of arrhythmia. In this paper, we propose a machine learning model based on reservoir computing (RC), followed by principal component analysis, and a readout module based on one-class support vector machine (OCSVM) for the anomaly detection task. This approach is used to perform on-device learning and for real-time anomaly detection of pathological conditions executed on a low-power microcontroller unit (MCU). The learning step requires a limited amount of input data eg. 600 Bytes to make it suitable for cheap deployment on a low-power device. We present practical modifications for optimizing the whole pipeline while keeping performances at the expected level and reducing the complexity of the proposed architecture. We also measured the complexity of the pipeline using X-CUBE-AI tool on STM32 MCU's. The detection performances has been evaluated on two different datasets; the publicly available MIT-BIH arrhythmia dataset and a proposed ST dataset collected from a low-cost sensors system. Proposed model achieved 83% in accuracy and 89% in f1 score on average, up to 97% as a maximum.