Ultra-low power machine learning for implantable medical devices 

​​Pattern recognition is a fundamental capability of artificial intelligence (AI). When used in closed-loop therapies such as implanted cardioverter defibrillators and neurostimulators, it promises more accurate classification of electrical or nerve activity in the body. This permits the system to provide the optimal electrical stimulation as a therapeutic intervention.

​Conventional AI systems would quickly drain the limited battery power available in a medical implant. In addition, they typically rely on continuous internet connectivity, making them unacceptable as part of a system that is key to sustaining life.

​We didn’t want to let the power appetite of conventional AI stand in the way of developing more accurate closed-loop therapies. We therefore used one of the first available off-the-shelf microcontrollers with a dedicated low-power neural network accelerator to build a system that is able to classify real-time ECG data at the power budget available in an implantable pacemaker.

​Along the way, we learned three valuable lessons about implementing low-power machine learning “at the edge” (i.e., on the device itself). These lessons comprise changes to the way AI models for signal classification are trained in concert with changes in the design of the system hardware.

​First, we needed to train the AI model to successfully classify ECG data at the reduced resolution of a low-power AI accelerator. We used Quantisation Aware Training, a technique that allows the AI model to understand as it is being trained how its performance will change with a reduction in resolution of the data. This helped us to maintain performance at the 8-bit resolution of the low-power embedded AI accelerator, as opposed to the 32- or 64-bit resolution typical of desktop or cloud AI systems.

​In the body, the amplitude of ECG data is influenced by electrical contact, person-to-person variation and heart activity. To eliminate biases due to the amplitude of features, it is common to rescale the data before processing. But at the limited resolution of low-power edge devices, it isn’t always possible to scale the data digitally and still get the same classification performance. For our arrythmia classification hardware, we therefore needed to skilfully design the analogue front-end to be able to use the full dynamic range, potentially even being able to dynamically change gain before the signal is digitised.

​The final challenge in implementing low-power AI was one of timing. To reduce power consumption, edge devices are off most of the time, meaning that sampling and signal classification cannot run continuously. In addition, pre-labelled training datasets are often time-aligned, so that the trained AI model only expects to see data in the middle of the sampling window.

​If sampling and data processing are started at the wrong time, this can either lead to data being discarded (and battery power being wasted) or result in worse classification performance. In a low-power system, it is therefore desirable to pre-process the data still in the analogue domain to enable efficient timing of sampling and inference.

Since identifying and overcoming these technical challenges, we have implemented a system that is capable of classifying real-time ECG data and screening it for potential arrhythmias at a power budget that makes the system suitable for running with an implantable pacemaker. We believe it’s only a matter of time before low-power AI is implemented as part of other closed-loop therapies, and look forward to continuing developments in this fast-moving field.