TL;DR
This paper proposes a novel hardware-mappable cellular neural network (CNN) architecture for real-time detection of cardiac wavefronts in implantable devices. By leveraging memristor technology and a distributed computing approach, the system aims to enable more precise and less invasive cardiac therapies.
What Problem Does This Paper Address?
Current implantable cardioverter-defibrillators (ICDs) have limited sensing resolution, requiring high-energy shocks to treat life-threatening arrhythmias. This can lead to myocardial damage and patient discomfort. To address this, there's a growing need for high-definition sensing and localized, precise therapeutic interventions. However, the computational demands of processing high-resolution cardiac data in real-time on compact, low-power implantable devices are a significant challenge. Existing hardware is often too bulky or energy-intensive for these applications.
Methods
The authors propose a distributed computing algorithm based on cellular neural networks (CNNs). This approach is designed to be "hardware-mappable," meaning it can be efficiently implemented on specialized hardware. Key aspects include:
- Cellular Neural Networks (CNNs): Each "cell" in the network processes local information from its neighbors, making it suitable for distributed architectures. This contrasts with traditional monolithic neural networks that require a single, powerful processing unit.
- Chiplet-based Architecture: The CNN is mapped onto a network of small, integrated circuits called "chiplets." Each chiplet acts as a CNN cell, processing local sensor data and controlling a local actuator. This allows for scalability and integration into organ-conformal platforms.
- Memristor Technology: Memristors are explored as a compact and energy-efficient hardware fabric for implementing the CNN's "synapses" (templates). Their non-volatile nature and analog behavior are well-suited for this task.
- Hardware-Aware Training: The training process considers hardware limitations, including limited bit precision for template values and potential noise in memristor devices. Techniques like stochastic rounding are employed to maintain performance with reduced precision.
- Binarization: The output of the network is binarized to simplify processing and control actuators. Both "binarization during training" and "binarization after training" approaches are investigated.
- Data Preprocessing: Optical mapping data is preprocessed using the Hilbert transform to extract phase information, which is crucial for identifying wavefronts.
Datasets
The study utilized a dataset of 1000 optical mapping images of epicardium tissue from a deidentified donor human heart. 800 images were used for training and 200 for testing. This dataset captured several fibrillation events.
Key Insights and Ideas
- Distributed CNNs are a viable path for implantable edge computing: The paper demonstrates that a distributed CNN architecture can effectively process complex cardiac signals on-chip, overcoming the limitations of centralized processing.
- Memristors offer a compelling hardware solution: The inherent properties of memristors make them ideal for implementing the dense, low-power computational fabric required for these implantable systems.
- Hardware constraints significantly impact algorithm design: The need for low power, small size, and limited bit precision necessitates specialized training techniques and architectural choices, moving beyond standard deep learning practices.
- Stochastic rounding is crucial for low-bit precision: This technique significantly outperforms rounding-to-nearest when dealing with limited bit precision in memristor-based hardware, maintaining high accuracy and precision.
- Binarization after training offers a stable convergence: While binarization during training can be faster, binarization after training leads to more stable convergence and better overall performance.
Key Results
- High Performance Metrics: The proposed algorithm achieved high performance metrics in ideal conditions: accuracy and specificity above 99.7%, sensitivity of 97.65%, and precision of 93.83%.
- Robustness to Noise and Imperfections: The system demonstrated robustness to moderate levels of read noise and input noise. Even with 1% of devices stuck, precision remained above 60%.
- Impact of Bit Precision: Stochastic rounding maintained high accuracy (above 99.6%) and precision (above 91.2%) even at 6-bit precision, a significant improvement over rounding-to-nearest.
- Batch Training Benefits: Batch training improved inference performance on the test dataset, increasing sensitivity from 83.49% to over 97% when using 800 training images.
- Neighborhood Size Impact: Reducing the neighborhood size from 8 to 4 neighbors had a minimal impact on performance (typically ±1-2 percentage points), suggesting potential for further hardware optimization.
Limitations and Future Work
- Hardware Implementation Challenges: The paper acknowledges that practical implementation on organ-conformal platforms faces challenges related to chiplet design, integration with sensors/actuators, and achieving millisecond response times.
- Dataset Size and Diversity: Future work will focus on expanding the dataset with more labeled optical and electrical mapping data, including variations across sexes and ages, to improve generalizability.
- Device Yield and Reliability: While the system shows some robustness to noise, the impact of device failures (e.g., stuck memristors) needs further investigation, especially for critical medical applications.
- Exploring Advanced CMOS Nodes: For compact chiplets, advanced CMOS nodes (e.g., 22 nm) might be necessary, requiring careful consideration of processing variations.
- Real-time Operation and Energy Harvesting: Future work will aim for a suitable chiplet design that meets real-time operational requirements and is compatible with energy-harvesting systems for implantable devices.
Summary updated