In the realm of medical technology, a groundbreaking development has emerged from TTP in Cambridge. They have successfully created a low-power AI system tailored for implantable medical devices. This innovative AI framework is designed to analyze real-time ECG data and detect potential arrhythmias, all while operating within a power budget that allows seamless integration with implantable pacemakers.
The significance of pattern recognition in AI cannot be overstated. When applied to closed-loop therapies like implanted cardioverter defibrillators and neurostimulators, AI can enhance the accuracy of classifying electrical or nerve activity within the body. However, traditional AI systems are not conducive to medical implants due to their high power consumption and reliance on continuous internet connectivity, which are impractical for life-sustaining devices.
To overcome these challenges, the developers at TTP opted for an off-the-shelf microcontroller from Analog Devices equipped with a specialized low-power neural network accelerator. This strategic choice enabled the system to effectively classify real-time ECG data within the power constraints of an implantable pacemaker. Moreover, they introduced novel methodologies for training AI models tailored for signal classification, aligning them with the hardware design modifications.
One key aspect of their approach involved training the AI model to classify ECG data at a reduced resolution compatible with a low-power AI accelerator. By utilizing Quantisation Aware Training, the AI model gains insights during training on how its performance will be impacted by data resolution reduction. This optimization ensured consistent performance even at the 8-bit resolution of the embedded AI accelerator, a critical requirement for implantable medical devices.
Addressing the challenges posed by varying ECG data amplitudes, the developers revamped the analogue front-end of the arrhythmia classification hardware. This redesign allowed for utilizing the full dynamic range and potentially adjusting gain dynamically before digitization. Furthermore, they tackled the timing dilemma inherent in low-power AI systems, where devices are often inactive to conserve power. By pre-processing data in the analogue domain, they optimized sampling and inference timing, preventing data loss and enhancing classification accuracy.