Template Matching Based Early Exit CNN for Energy-efficient Myocardial Infarction Detection on Low-power Wearable Devices

Abstract

Myocardial Infarction (MI), also known as heart attack, is a life-threatening form of heart disease that is a leading cause of death worldwide. Its recurrent and silent nature emphasizes the need for continuous monitoring through wearable devices. The wearable device solutions should provide adequate performance while being resource-constrained in terms of power and memory. This paper proposes an MI detection methodology using a Convolutional Neural Network (CNN) that outperforms the state-of-the-art works on wearable devices for two datasets - PTB and PTB-XL, while being energy and memory-efficient. Moreover, we also propose a novel Template Matching based Early Exit (TMEX) CNN architecture that further increases the energy efficiency compared to baseline architecture while maintaining similar performance. Our baseline and TMEX architecture achieve 99.33% and 99.24% accuracy on PTB dataset, whereas on PTB-XL dataset they achieve 84.36% and 84.24% accuracy, respectively. Both architectures are suitable for wearable devices requiring only 20 KB of RAM. Evaluation of real hardware shows that our baseline architecture is 0.6x to 53x more energy-efficient than the state-of-the-art works on wearable devices. Moreover, our TMEX architecture further improves the energy efficiency by 8.12% (PTB) and 6.36% (PTB-XL) while maintaining similar performance as the baseline architecture.

Real-Time Myocardial Infarction Detection Approaches with a Microcontroller-Based Edge-AI Device

“…Although exhibiting a slightly lower accuracy (94.76%) compared to the solution delineated in reference [27] (95.22%), this work approach distinguished itself by minimizing computational complexity, rendering it suitable for resource-constrained wearable devices. In contrast, reference [24] outperformed our method in wearable device memory occupancy, requiring only 20 kB, even if there was a diminished accuracy (84.36%) in comparison to this study's solution. Lastly, compared to the study presented in [30], which employed a deep LSTM network, our work demonstrated superior performance in terms of accuracy (84.17%) and feasibility for wearable devices in real-time MI detection.…”

“…The NN operated correctly, with the NN pulse being invoked every 6 s, taking 830 ms for Spectrogram calculation and inference, in line with the algorithm outlined in the block diagram of Figure 6. The literature contributions that adopted a comparable approach to Method 2 were those using II-lead ECGs, i.e., [24,27,30]. Limited studies in the literature have utilized experiments based on II-lead ECGs, with detailed performance comparisons illustrated in Table 11.…”

“…While [22,23] aimed to streamline classifier complexity for energy efficiency, they heavily relied on extensive feature engineering to identify the optimal features. Moreover, the resource-intensive nature of the feature extraction process in terms of time and energy rendered these methods incompatible for real-time MI detection on wearable devices [24].…”

“…In the eBNN approach, the convolution result undergoes a direct transition to pooling, followed by batch normalization and binary activation, thereby eliminating the necessity for intermediate storage. The authors in [24] introduced an innovative Template-Matching-based Early Exit (TMEX) CNN architecture that enhanced the energy efficiency of the baseline architecture while maintaining comparable performance. The evaluation of their methodology was conducted on two datasets, the PTB and PTB-XL [29].…”

Stress Detection Using Context-Aware Sensor Fusion From Wearable Devices

“…Further, SELF-CARE could be applied more broadly in the domain of affective computing to include additional tasks beyond stress detection and emotion recognition. Moreover, it can also be applied to other wearable healthcare applications like human activity recognition [29]- [31], myocardial infarction detection [32]- [34] etc., that involves data from multiple wearable sensors. Lastly, SELF-CARE's use of a specialized set of ensemble classifiers has broad applicability to IoT sensing, including the domains of sensor networks [35], and transportation [36].…”