Medical electronics are a high-tech wonder, intricate machines to check your heart signals, brain functioning, blood vessels ... While you'll mainly find them in hospitals, that is a place where most people would rather not come too much, let alone stay for a longer period. In addition, these machines are expensive and require specialized personnel to operate. This puts them out of reach for all but acute medical situations. For prevention and follow-up of chronic diseases, we need cheaper electronics that are easily accessible – electronics that you can use at home.
Now imagine a 10cm² plaster, a plaster that you attach to the skin over your heart and that monitors your heart. This plaster functions for a week under all conditions (even if you are running, or taking a bath) and sends the results to the hospital wirelessly. After a week of monitoring, you apply a new patch if needed (and return the old one for recycling). Inexpensive, available to everyone, easy and comfortable to use, and packing all the technology of a full-fledged ECG (electrocardiogram) machine.
At imec and Holst Centre, we’ve done a lot of R&D into the electronics needed for this kind of sensors. Recently, we’ve build a demonstrator ECG necklace that goes a long way towards this vision of disposable, wireless plasters for home-based monitoring. This demonstrator includes some custom-built electronics and algorithms and currently represents the state-of-the-art of low-power electronics for healthcare. At the same time, it illustrates the power consumption gap that we still have to close to arrive at the visionary plaster.
How Far Can We Get with Off-the-Shelf Components?
We started out by integrating an ECG sensor into a comfortable necklace using the best available off-the-shelf components and maximally optimizing them. The first challenge was that that we had to create the sensor interface from scratch, because nothing suitable was available. So we made our own analog biopotential chip [1] and integrated it an off-the-shelf radio and DSP chip.
For the best result and to offload computation from the DSP and radio, we reconsidered some conventional divisions between the analog and digital domains.
For example, the ECG monitor should usable under all circumstances, even if you are jumping around on a tennis court. Therefore the monitor has to filter out motion artifacts, i.e. glitches in the signal caused by the monitor moving relative to the body. This filtering can be done by using accelerometers and having the DSP recompute the heart signal on the basis of the accelerometer input. This, however, will necessitate heavy computation and energy use. Moreover, it will still only consider absolute movements of the monitor, not movements of the monitor relative to the body, which is what you really need to compensate. Instead, we chose to implement motion artifact filtering in the analog component via continuous monitoring of the impedance between the skin and the monitor electrode.
After integration and optimization, we arrived at a fully functional ECG monitor in the form of a necklace. Using all available power modes and duty cycling, this sensor consumes 1.1mW. With a 140mAh Li-Ion battery, this ensures an autonomy of 13 days.
Enough autonomy, usable under most circumstances ... but not a plaster.
Now We’re Talking Real Low-Power
To integrate our ECG electronics in a thin patch that can be worn on the skin or woven into clothes, we would have to get rid of the Li-Ion batteries, and revert to a printed thin-film battery. But driving our optimized electronics with a printed thin-film battery leaves us with a mere 8 hours of autonomy (for a 6cm² battery, giving 1mAh/cm² at 1.5V).
If the requirement is to have a lifetime of 7 days for such a patch, we have to improve the state-of-the-art of low-power electronics with a factor of 21, to reach a target where the whole ECG system consumes only 50µW of power.
In addition, this is one of the simplest examples of body sensors imaginable. Once you start capturing and analyzing brain signals, for example, you may need 1,000 times the compute capacity for the same power. And if you think about using sensors in implantable devices, the autonomy should be stretched to years, if not a decade. So what we will need for these sensors at is a real paradigm-shift in low-power electronics.
Throwing in a Dedicated Radio and DSP
For almost all wireless sensor applications, the radio connection takes up 50 to 85 percent of power consumption. In our first demonstrator necklace, and after optimizations, a full 550µW of the 1.1mW went into the radio.
So we built our own radio chip, which is a 2.4GHz ULP OOK single-chip transceiver [2]. Integrating this radio into the necklace sensor resulted in a total power consumption of 362µW, a system power reduction with more than a factor 3.
The next target was to replace the commercial DSP with our own ultra-low-power design. One of the requirements for the DSP is that we want to do more local processing on the ECG patch. Such processing will add diagnostic intelligence to the patch, and at the same time unload the radio, as less data has to be transmitted.
But here is the challenge: we need a DSP that is really low-power most of the time, but that can also handle high-load, energy-intensive processing when needed. We need a processor that uses almost no energy when it collects data, but that has the power to also crunch those data with a complex ECG algorithm and data compression, and a processor that can switch between modes of operating, from high-intensity computation over low-intensity data collection to sleeping.
Recent work on designing ultra-low power systems has focused on running processors at a subthreshold regime, at the lowest possible input voltage. This has resulted in low-power devices with energy efficiencies of a few pJ/cycle. But these devices are very restricted in what they can process, having a clock frequency usually in the kHz range. Such a low frequency, combined with limited processing capacity, small on-chip memory, and low computation precision makes these systems unsuited for complex ambulatory monitoring.
What we need instead, and what we had to develop, was a processor that could run applications with different degrees of complexity in an energy-aware way, switching between modes of operation in an event-driven way. The result is the CoolBio processor, which we designed in collaboration with NXP [3]. CoolBio is a voltage-scalable biomedical signal processor which uses only 13pJ/cycle at 1MHz and 0.4V.
Cruising Along at 13pJ/Cycle
The prototype DSP-chip that we have designed makes use of extreme system partitioning to enable consistent duty cycling. It features power gating, voltage scaling, multi-clock domains, multi-voltage domains, and extensive clock gating. The processor can run at voltage range of 0.4V to 1.2V and supports frequency range of 1MHz to 100MHz. It has sufficient computational power to run a complex ECG algorithm with feature extraction and motion artifact cancelation or multi-channel EEG processing. Yet it only consumes an average of 13pJ/cycle running a CWT-based ECG application at 0.4V, which is the lowest reported energy consumption reported today. Compared to other recent design, this system has more computation capability and a larger available frequency range at a comparable energy/cycle consumption.
The chip is designed in a standard 90nm LP process, has an area of 1.875 × 3.75 mm², and integrates 195K NAND2 equivalent gates and 2Mb memories. It includes an NXP CoolFlux BSP baseband processor core, a memory sub-system with separated program and data memories, memory controller, interrupt controller, timers, JTAG, I2C, and five device interfaces (an 8-channel EEG sensor, ECG sensor, temperature sensor, and 3D accelerometers).
We implemented system partitioning by dividing the system into 15 power domains. There is the data and program memory, which we divided into 13 memory banks, each in a separate power domain that can be individually put in retention mode or turned on/off. The processor is also in a separate power domain. The rest of the system, including the periphery, constitutes the last power domain – the periphery domain – which is always on. The processor and the periphery domains can run in a voltage range between 0.4V and 1.2V. Only the commercial memory banks we included cannot operate at voltages as low as 0.4V. Therefore, we included innovative level shifters capable of up-converting from 0.36V to 1.32V.
The system can run in one of four power modes:
- Data collection mode: only the periphery domain and the required number of data memory banks are on running at 0.4V and 0.7V respectively; the program memory is in retention mode and the rest of the system is off.
- Low performance mode: the processor and periphery are on running at 0.4V, part of the program and data memory are on, running at 0.7V and the unused memory banks are off.
- High performance mode: All components are running at 1.2V and the unused memory banks are off.
- Sleep mode: all power domains are off and no internal clock is running. The system can be woken up from the clock-less sleep mode by an external interrupt.
The memory controller can access the memory subsystem even when the processor is shut off. This architecture allows collecting data from the sensor in the DSP memory while at the same time shutting off the processor and major parts of the memory. The memory controller wakes up the processor when a predefined number of samples are collected.
Intermediate Result: 5x Less Power, Much More Robust
Integrating the new radio and DSP in our ECG patch, we arrive at a total energy consumption that is five times lower than the optimized off-the-shelf solution that we started with and which is much more robust when used in daily life than the original solution.
In the process, we have demonstrated that it is possible to make relevant prototypes that can be compared and validated successfully against commercial equipment that is orders of magnitudes larger and more expensive. In contrast to such equipment, wireless sensors such as our ECG necklace could be used at home, by everyone. Not only in case of emergencies, but also for prevention, diagnosis and revalidation.
We expect applications such as these to appear first in wellness, sports and gaming applications, and later on in healthcare devices, which have much more stringent requirements and approval procedures.
Eventually, with this technology, it will be possible to help more people than is possible today, at a lower price per person, and for a wider range of conditions, even including preventive healthcare.
Jan Provoost is science editor at imec. He obtained his M.A. in Languages in 1989 and his M.Sc. in Computer Science in 1993, both from the Catholic University of Leuven, Belgium. Jan worked as researcher and writer for several companies and research institutes exploring natural language processing and information security before joining imec in 2007.
References
- http://www2.imec.be/be_en/press/imec-news/archive-2010/biopotential.html
- Vidojkovic, M.; Xiongchuan Huang; Harpe, P.; Rampu, S.; Cui Zhou; Li Huang; Imamura, K.; Busze, B.; Bouwens, F.; Konijnenburg, M.; Santana, J.; Breeschoten, A.; Huisken, J.; Dolmans, G.; de Groot, H.; A 2.4GHz ULP OOK Single-Chip Transceiver for Healthcare Applications, Solid-State Circuits Conference Digest of Technical Papers (ISSCC), 2011 IEEE International; pp458-460 – DOI 10.1109/ISSCC.2011.5746396
- Ashouei, M.; Hulzink, J.; Konijnenburg, M.; Jun Zhou; Duarte, F.; Breeschoten, A.; Huisken, J.; Stuyt, J.; de Groot, H.; Barat, F.; David, J.; Van Ginderdeuren, J.; A voltage-scalable biomedical signal processor running ECG using 13pJ/cycle at 1MHz and 0.4V; Solid-State Circuits Conference Digest of Technical Papers (ISSCC), 2011 IEEE International; pp332-334 – DOI 10.1109/ISSCC.2011.5746341
