Enterprise AI Analysis

Insights from TOP-8 Teams

The top-performing teams demonstrated diverse strategies, often refining standard CNN architectures or employing traditional ML methods for efficiency. Here's a look at some leading approaches:

AI Methodology: Deep Learning vs. Traditional ML

The contest showcased a clear trend towards deep learning, particularly Convolutional Neural Networks (CNNs), for achieving high detection accuracy. However, traditional machine learning methods also proved effective for specific optimization goals.

Optimizing Network Architectures

All deep learning teams improved upon the baseline CNN, focusing on reducing parameters and MAdds while maintaining high Fβ scores. Key strategies included fewer convolutional layers, fewer output channels, and employing larger kernel sizes or dilated convolutions for expanded receptive fields.

For example, Gatech-EIC-Lab and MAD-AI focused on reducing model complexity significantly. SEUer optimized for minimal MAdds through a streamlined architecture and initial mean pooling of input data. Most networks achieved receptive fields larger than 60, essential for capturing global features from the 1250-point IEGM input.

Effective Training Strategies

Advanced training strategies were crucial for achieving high accuracy despite variations in network architecture. Teams fine-tuned aspects like dropout, epochs, learning rates, loss functions, and optimizers.

Key takeaways include the efficacy of Stochastic Weight Averaging (SWA) for generalization, mean pooling for input data reduction (SEUer), and Warmup strategies to stabilize early training (VIPS4Lab@UNIVR).

Data Augmentation Techniques

To overcome the limited size and imbalance of the IEGM dataset, most deep learning teams employed various data augmentation strategies. These methods enhance model robustness and accuracy.

While data augmentation significantly improved performance, some methods like "shift" (used by HuskyCS-Deepical) were identified as potentially distorting for IEGM signals if not applied carefully.

Category-Specific Detection Results

Analysis of error rates across different VA and non-VA categories revealed varying performance challenges. Some categories proved significantly harder to detect accurately.

SVT (Supraventricular Tachycardia) presented the highest challenge, with a 98.55% mean error rate across all TOP-8 models, suggesting it often resembles VT morphologically (Figure 4, Page 6). VFb (Ventricular Fibrillation) also had a relatively high mean error rate (16.12%), though VIPS4Lab@UNIVR achieved a notable low of 3.11%.

Hardware Deployment & Optimization

Successful deployment on the NUCLEO-L432KC microcontroller involved careful compilation and hardware optimization. Teams leveraged various compiler flags and techniques to maximize performance within resource constraints.

Optimization options like -Oz (optimize for size) and -O3 (optimize for speed) were common. LTO (Link-Time Optimization) was used by top teams like Gatech-EIC-Lab and SEUer for inter-modular optimization. SEUer also uniquely employed SLSOM (Split Load and Store Multiple) to reduce interrupt latency, showcasing a deep understanding of the hardware-software co-design.

Key Lessons Learned

The TinyML@ICCAD contest provided critical insights into the development and deployment of AI on edge devices for healthcare.

• Dual AI Efficacy: Both deep learning and traditional machine learning can achieve high accuracy for VA detection. Deep learning generally offers higher accuracy and robustness, while traditional ML excels in lower latency and smaller memory footprints.
• Challenging Categories: Some arrhythmia types, particularly SVT (98.55% error rate) and AFt (46.7% error rate), remain highly difficult to detect accurately. Future contests could benefit from more specific data or sub-tasking for these scenarios.
• Hardware-Software Co-optimization: Advanced hardware and compilation optimizations significantly impact deployment performance. Employing strategies for low latency and low memory occupation is crucial for real-world applicability.
• Balanced Performance: Achieving a high Fβ score alone is not sufficient; a balanced approach considering latency and memory is key to a favorable ranking, as demonstrated by SEUer's second-place finish.