Opportunities and challenges for ultra low power signal processing in wearable healthcare

Wearable devices are starting to revolutionise healthcare by allowing the unobtrusive and long term monitoring of a range of body parameters. Embedding more advanced signal processing algorithms into the wearable itself can: reduce system power consumption; increase device functionality; and enable closed-loop recording-stimulation with minimal latency; amongst other benefits. The design challenge is in realising algorithms within the very limited power budgets available. Wearable algorithms are now emerging to answer this challenge. Using a new review, and examples from a case study on EEG analysis, this article overviews the state-of-the-art in wearable algorithms. It demonstrates the opportunities and challenges, highlighting the open challenge of performance assessment and measuring variability.


INTRODUCTION
Wearable devices are starting to revolutionise healthcare and mobile healthcare by allowing the easy, unobtrusive and long term monitoring of a range of body parameters. Activity trackers using accelerometers, such as the fitbit [1], have been the most successful initial devices, and emerging units, such as the Samsung Simband [2], can monitor a range of parameters including accelerometry for activity monitoring, electrodermal activity for arousals, and heart rate via photoplethysomgraphy. Wearable algorithms is the name given to the new signal processing approaches that are emerging for wearable devices which embed signal processing into the device hardware [3]. Illustrated in Fig. 1, historically the focus of online signal processing inside a sensor node has been for real-time data reduction. In wireless sensors it is the transmitter that dominates power consumption and if the data rate can be reduced prior to transmission significantly better battery life can be achieved [4]. Today, there are many additional benefits that are being enabled by the use of signal processing embedded in the wearable itself [3]: • Reduced system power consumption. This  • Increased device functionality.
• Reliable, robust operation over unreliable wireless links.
• Reduction in the amount of data to be analysed offline.
• Real-time data redaction for privacy.
As a result there have been rapid developments in wearable algorithms in recent years, and this is opening new opportunities in signal processing techniques, algorithms, and applications for delivering healthcare benefits. Using a new review, and examples from a case study on EEG analysis, this article overviews the 2015 state-of-the-art in wearable algorithms (Section 2). These then highlight (Section 3) the gaps in the research landscape that are emerging, and the corresponding opportunities and challenges.

WEARABLE ALGORITHMS
Wearable algorithms is a new discipline in signal processing distinguished by the requirements for very low power hardware implementations and power consumption aware performance testing [3]. This was a key driver in the development of compressive sensing as a major signal processing area as it provides low power consumption compression with little distortion [4]. Nevertheless, the underlying aim of compressive sensing is data reduction for reduced power consumption. It does not enable the other potential benefits of wearable algorithms highlighted in the introduction. To do this, specif- ically designed algorithms for each application are required.
Although there are open challenges in each stage, the general flow for these algorithms is shown in Fig. 2. The key stages are low power feature extraction and low power classification. To investigate the state-of-the-art in these blocks Table 1 presents a new review of papers published in IEEE Transactions since 2011 that implement some form of algorithm in hardware for use in wearable sensors. Our focus is on full hardware implementations for the lowest power consumption, rather than software implementations with some hardware accelerators. Table 1 thus provides key insights into the main approaches that are currently being used in algorithms for wearables, and the current state-of-the-art.

GAPS IN THE SIGNAL PROCESSING LANDSCAPE
Focusing on power consumption, in 2010 authors in the IEEE Signal Processing magazine discussed the question: What does ultra low power consumption mean? They came to the conclusion that it is where the "power source lasts longer than the useful life of the product" [25]. To operate for 10 years from a miniature 1000 mAh battery, the average current draw needs to be approximately 10 µA or less. The largest current draw in Table 1 is ∼170 µA, and there are a number of algorithms that are in the <10 µA range, making this goal close to being realised. Such low power levels are possible due to the low frequency nature of human physiology. Few parameters need to be sampled at more than 1 kHz and this allows the signal processing electronics to be heavily duty cycled or power scaled. As a case study, Fig. 3 shows a Continuous Wavelet Transform (CWT) circuit which is designed for processing brainwave (EEG) signals in the 2 Hz region. This low frequency allows the average current draw to be scaled to 60 pA, giving signal processing information essentially for free in terms of power consumption. As a result the most substantial opportunities and challenges for wearable algorithms lie in the interface with algorithmic approaches as opposed to in pure circuit design. Focusing on feature extractions, the majority of wearable algorithms created so far (12 out of the 20 in Table 1) are based upon frequency information, with wavelet transforms being particularly popular. Given the results in Fig. 3 this focus is not surprising as it leads to very low power consumptions. However, it indicates a potential over reliance on time-frequency decompositions as the best algorithmic starting point. It seems unlikely that wavelet decompositions Microphotograph and operation of a low frequency, 60 pW, CWT for real-time EEG analysis [7].
would provide the best, or even suitable, feature extraction across all signal types and all potential applications. There is a clear opportunity for creating wearable algorithms that are based on other feature extraction methods, such as the fractal dimension [26] or Empirical Mode Decomposition [27]. Similarly focusing on classifiers, while there has been recent work on low power implementations of Support Vector Machines (SVMs) (e.g. [15]), many current wearable algorithms are based on threshold detection. The wide range of machine learning approaches have not yet been explored. Recent results [28] have suggested that many disparate classifiers actually achieve very similar algorithm performance. If confirmed this is an ideal opportunity for wearable algorithms. It means the classification procedure can be selected for minimum power consumption, with little impact on the classification accuracy.
Investigating this requires studying the three-way tradeoff between algorithm performance (e.g. correct detections), algorithm cost (e.g. false detections), and power consumption. This is a large design space, which leads to difficult decisions for the system designer: is it preferable to maximize performance, or to minimize cost or to minimize power consumption? Is an algorithm with very low power consumption, but comparatively low algorithm performance a better choice than a higher power, higher performance algorithm? Fully, and systematically, exploring this design space is the major challenge facing wearable algorithms. The algorithms in Table 1 are beginning to populate the space, but there is much more to do. It is an open challenge to formally investigate the trade-offs present (as opposed to making individual algorithms and finding where they lie) which could lead to more automated tools for helping designers choose the most appropriate trade-off point for their application.
The above challenge is compounded by the difficulties in assessing algorithm performance and cost in healthcare applications. Many applications in healthcare are highly variable between different people, and within the same person over time. For example, recent compressive sensing results have highlighted that the level of algorithm success is dominated by the variance in performance over time, not the average performance level [4,29]. Similarly, Fig. 4 shows the algorithm performance results of a CWT based spike detection method when multiple records are analysed [30]. Ideal performance would be in the top left hand corner. Performance  results in different records (red lines) show the common pattern of: many records have high performance and acceptably low cost; some records have high performance, but too high cost; and a few records have both performance and cost too poor. Overlaid (blue lines) is the variance in the average performance introduced due to the non-ideal CWT of Fig. 3 being used. The variance in performance between records is much larger than the variance introduced due to the non-ideal, and very low power, circuit implementation. As a result there is potential for designing even lower power algorithms with more variance, without substantially impacting the average algorithm performance. However, just as early wearable algorithms often did not report measures of both algorithm and power performance [3], many current wearable algorithm approaches report only an average performance level. There is no assessment of intra-and inter-person variability. This is because there is a substantial challenge in devising new methods for accurately and compactly summarising and reporting algorithm variances which can be compared between approaches. This is essential for algorithms to be usable in clinical grade healthcare applications (as opposed to consumer grade applications) and to accelerate the creation of future wearable algorithms. Unfortunately at present there are few methodological tools avaiable, and little consensus for how to best measure and quantify this variability.
Finally, Fig. 5 shows the average performance of the CWT based algorithm from Fig. 4 as the amount of input noise is increased. Traditional design approaches always endeavour to minimize the effective noise present in a system, often by trading-off with increased power consumption. However, noise-enhanced algorithms are a branch of signal processing theory where algorithm performance is not only robust in the presence of noise, but up to a certain point it gets better as more noise is introduced [31]. Noise-enhancement is therefore of great interest for simultaneously reducing power consumption and improving signal processing performance in low power wearables. While compressive sensing is the best known recent development in signal processing theory applicable to wearable algorithms, there are undoubtedly major opportunities for the greater use of noise-enhancement, and also in using innovations from other branches of signal processing theory which have not yet been identified.

CONCLUSIONS
Wearable algorithms are an emerging truly multi-disciplinary problem where, to achieve better functionality at the lowest levels of power consumption, innovations are required on multiple levels: in the human-monitoring application design, in the signal-processing design, in the performance-testing design and in the circuit design. This presents a large, fourdimensional, multi-disciplinary design space that has not yet been fully explored by a long way. Many challenges and opportunities are present, and while innovative design at all of the four levels in isolation will be beneficial, for future systems it is critical to exploit the multi-disciplinary factors present and the interactions between the different levels.