ANN Based Classification for Heart Defibrillators

NeurIPS 1991

TL;DR

This paper explores using Artificial Neural Networks (ANNs) for classifying heart arrhythmias in implantable cardioverter-defibrillators (ICDs). The authors propose novel ANN architectures and micro-electronic implementation strategies to achieve high classification accuracy while adhering to the strict power and area constraints of implantable devices.

What Problem Does This Paper Address?

Current implantable cardioverter-defibrillators (ICDs) rely on simple classification algorithms to detect and treat dangerous heart rhythms (arrhythmias). These simple algorithms are a necessity due to the severe limitations on power dissipation and physical space within these implanted devices. However, this simplicity leads to a significant problem: many arrhythmias share similar characteristics, making accurate classification difficult. This can result in false positives (delivering therapy when not needed) or false negatives (failing to deliver therapy when required), with error rates reaching up to 13%. The paper aims to bridge this gap by investigating more sophisticated ANN-based classifiers that can offer higher performance without compromising the critical power and area constraints.

Methods

The paper investigates several ANN-based classification architectures for both single-chamber and dual-chamber ICDs.

  • Single Chamber Classification:
    • Multi-Layer Perceptrons (MLPs): Explored MLPs trained with various backpropagation algorithms (vanilla, conjugate gradient, and a specialized limited-precision "Combined Search Algorithm" - CSA). Input features were derived from time-domain signal analysis.
  • Decision Trees: Utilized the C4.5 algorithm to induce decision trees, noting the trade-off between performance and tree size.
  • Dual Chamber Classification:
    • Multi-Module Neural Networks (MMNN): A modular approach designed to improve classification, particularly for Supra-Ventricular Tachycardia (SVT). The architecture divides the classification problem into sub-problems.
  • Hybrid Decision Tree/MLP: Mimics cardiologists' diagnostic process by combining a decision tree for rate and timing analysis with an MLP for morphology-based template matching. An arbitrator then combines these outputs.
  • Microelectronic Implementation: The core focus is on low-power VLSI implementation. This includes exploring:
  • Sub-threshold operation: Operating transistors in a mode that significantly reduces power consumption.
  • Analog implementations: Utilizing analog circuits for neural network components like synapses and neurons to minimize power and area. Two specific chip designs, the "Master Perturbator Chip" and the "BOURKE Chip," are presented as examples of these analog implementation efforts.

Datasets

Data was collected from Electro-Physiological Studies (EPS) at hospitals in Australia and the UK, comprising data from over 70 patients. Cardiologists labeled this data. The testing set used for evaluation was separate from the data used for classifier building. Arrhythmias were induced by stimulation, meaning natural transitions between rhythms were not captured.

Key Insights and Ideas

  • ANNs offer a path to higher accuracy: The paper strongly suggests that ANNs, particularly MLPs and more structured architectures like MMNN, can significantly outperform simpler algorithms currently used in ICDs.
  • Limited precision doesn't cripple performance: For single-chamber classification, using limited precision (6 bits + sign bit) for weights and activities in MLPs did not substantially degrade performance, which is crucial for VLSI implementation.
  • Modular architectures can be superior: The MMNN architecture demonstrated higher performance than a single large MLP for dual-chamber classification, highlighting the benefits of breaking down complex problems.
  • Hybrid approaches mimic human expertise: The hybrid Decision Tree/MLP classifier, by combining rate/timing analysis with morphology, aims to replicate the diagnostic process of cardiologists, leading to high performance.
  • Analog VLSI is key for low power: The paper emphasizes that analog implementations, especially those using sub-threshold operation, are critical for achieving the ultra-low power consumption required for implantable devices.

Key Results

  • Single Chamber MLP: Achieved up to 96% average performance with unlimited precision and 94.8% with limited precision (6 bits + sign bit).
  • Single Chamber Decision Tree: Achieved 95% correct classification.
  • Dual Chamber MMNN: Showed superior performance compared to a single large MLP across various rhythms, with an average performance of 96.2% for MMNN versus 89.3% for the single MLP.
  • Hybrid Decision Tree/MLP: Achieved over 99% correct classification on a multi-patient database.

Limitations and Future Work

  • Limited natural transitions: The data used for testing did not capture natural transitions between arrhythmias, which might be important for real-world ICD performance.
  • Complexity of dual-chamber classification: The paper notes that dual-chamber classification is inherently more complex, necessitating more structured classifier architectures.
  • Future Work: The authors suggest that micro-power VLSI ANNs offer a promising technology for enabling powerful classification strategies in implantable defibrillators. Further progress in microelectronic implementations is ongoing, with a focus on achieving low power and small area objectives through various design strategies. The presented analog implementation efforts are a step towards this goal.

Summary updated

BibTeX
@inproceedings{1991ann,
  title = {{ANN Based Classification for Heart Defibrillators}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1991},
}
Figure