Early wearable fitness monitoring devices were designed to perform a set of valuable but straightforward activities: tallying the number of steps we take daily, recording the number of hours we sleep, and monitoring our heart rate.

However, sales of simple fitness trackers have begun to stagnate, perhaps because these monitors didn’t deliver enough of the right kind of information. The data gathered by these devices on their own offer an incomplete picture of an individual’s health, with limited insights into strategies for long-term health and wellness.

This situation has begun to change with the advent of a new generation of wearable monitoring devices that integrate new technologies to reveal more profound and personalized health and wellness insights. Advances in artificial intelligence (AI) and neural network technologies can provide health and medical analysis that goes far beyond the simple charting of physical activities. These powerful cognitive systems enable the analysis of individual health data within the context of much larger data sets, and with consent, draw from the latest medical insights, family medical histories, genomic data, and a vast network of medical databases to reveal highly personalized, proactive health findings and treatments.

However, to enable a new generation of smarter, more capable health monitoring, devices will require new microprocessors with the performance power and energy efficiency to run real-time computation of AI neural networks, algorithms, and analysis of significant data flows in a small portable device. These hardware-based processing systems will need to be designed from the ground up to support the neural network workloads and architectures that enable advanced health monitoring devices to deliver more predictive health insights and successful outcomes.

Next Generation Health Monitors

Based on the fast-evolving technologies associated with machine learning and neural networks, the next generation of wearable health trackers will be designed to focus more closely on the wearer’s holistic and long-term health and wellness. These new wearable apps will not only monitor fitness, activity, and heart rate, they will pass the data they collect on to other applications that will evaluate, interpret, and customize this information so it is useful and relevant to the user.

These devices will have the capacity to learn about users on a deeply personal level, including information about diet, mood, and medical health history. This will provide insights and guide actions and choices that can help encourage healthier lifestyles, promote more informed health decisions, and overall wellness.

AI will change how a user experiences and interacts with health trackers. A smart interconnected device could monitor a wearer’s heart rate, for example, and be able to analyze this data against a dataset comprised of millions of other heart rate readings to generate meaningful insights into what the particular user’s heart rate should be, based on his or her age and history and on data from other patients of similar age or health status or a specific demographic group. Typically, networks take data from a wider study and then it is partitioned into training and test sets. It is trained on a representative part of the data and then is exposed to data it has not seen before, to understand its fit and error rates (accuracy).

Complex data can even be analyzed to measure mood changes. Response to stimuli can change in terms of galvanic skin response (GSR), eye pupil dilation, heartrate, blood pressure, and multiple other inputs. These can be synthesized to understand feeling states such as excitement, fear, arousal, passivity, or lack of interest.

This is where new technologies such as AI and neural networks come in. Comparative analysis of the data could reveal deep insights into the implications of the current readings and offer a suggested course of action based on that evaluation. For instance, detecting an abnormal heart rate involves looking at patterns and identifying irregularities to that pattern.

By evaluating an individual’s health data in connection with the vast medical data ecosystem, AI and neural networks can help health monitoring devices identify areas of potential health vulnerabilities; issue alerts, advice, or warnings; and offer suggestions of what responses or treatments to pursue.

Powered by neural networks, these smart data services can also serve as an early warning system to help identify signs of disease or age-related early onset problems. By recognizing symptoms at the earliest possible stage and enabling faster more accurate treatment, users have the potential for taking steps to prevent illness from taking hold or making other proactive responses.

Building More Powerful Neural Networks

The larger and more complex that the network of health data and medical databases becomes, the more useful and powerful its insights. However, this size and complexity also increases the system’s computational demands, and requires new levels of performance for neural networks, particularly in mobile use cases.

To do their work, neural networks need to be trained, and this is usually done offline on powerful server hardware. However, the recognizing of patterns and objects, known as inferencing, is done in real time.

In many cases, the inferencing is run on hardware in the cloud. However, when fast response times are required — such as real-time identification of health emergencies — latency issues make it generally not practical to run neural networks via the cloud.

To support these high performing neural networks, Imagination Technologies (Kings Langley, Hertfordshire, UK) has developed a new hardware neural network accelerator (NNA). Their PowerVR Series3NX offers a full hardware NNA solution built from the ground up to support leading neural net models and architectures as well as machine learning frameworks, such as Google’s TensorFlow or BAIR Caffe, at a high level of performance and low power consumption.

As an example, an NNA could be used in a chip specifically designed to monitor athletes for maximizing their performance, where their value is extremely high and any indication of performance issues because of fatigue, or say overheating, could be the difference between winning and losing.

A neural network can be used to analyze and recognize particular types of responses from GSR (based around the conductivity of the skin due to sweat) to sensors that can analyze perspiration for the presence of certain compounds. This “signature” can be recognized by a neural network and alerts or actions can be taken as a result. A sensor could be in a smart-watch where the sensor for sweat is on the back of the watch, close to the skin. It could also be in a patch taped to the person or could be worn as a chest band. Indeed, several different signals could all be compared and used to drive the data.

Diabetes monitoring of hypoglycemia and other aspects of diabetes and the result of falling levels of insulin over time could be recognized as “expected” or as potentially dangerous, depending on time of day, time since last insulin intake, etc.

Using the NNA for neural network training enables inferencing to be performed on the edge device itself. Its high performance and efficient power management are extremely important for this new generation of wearable devices. Moving inferencing on-device also addresses security issues that are critical when processing health data.

In addition, to be truly effective, the neural networks used for health monitoring need to be functional in remote areas. Since cellular networks are not always available, dedicated local hardware to operate health-related networks will result in safer and more reliable coverage when “off the grid.”

Benefits of a Full Hardware NNA Solution

The PowerVR Series3NX’s ultra-low power consumption makes it possible to operate full neural networks in small form factors, opening the market for smart, predictive health monitors on mobile platforms. Although it is a very low-area component, it delivers an extremely high inference rate per square millimeter.

Neural networks are commonly trained in 32-bit precision, yet for edge inference, the NNA allows bit-depths of 16-bit and lower, which, depending on the task, means that the model size and bandwidth can be reduced with only a marginal tradeoff in precision. The Series3NX’s flexible bit-depth support allows tuning on a layer-by-layer basis of each layer’s bit-depth and can improve precision and optimize model size. A single core running at 1.2GHz, can offer up to 4,096 MACs per clock (the industry standard performance indicator); meaning it can run over 10 trillion operations a second. And it is a highly scalable solution — through the use of multiple cores, much higher performance can be achieved, if required.

Toward More Powerful Health Monitoring Devices

NNA represents an inflection point in neural network acceleration and performance. Solutions incorporating it can help health monitoring devices become more technically robust and capable of analyzing data faster using sophisticated AI, machine learning, and deep learning techniques.

It is important, that the requirements for deploying neural networks be achieved within the power and performance constraints of mobile hardware. With this kind of processing muscle in a small footprint, the next generation of health monitors will no longer be reactive, passively tracking activities and vital signs, but will act as a smart, proactive and personalized advocate that can constantly evaluate health data parameters to promote wellness, enable more informed medical decisions, and anticipate health risks for more positive outcomes.

This article was written by Andrew Grant, Senior Business Development Director, Imagination Technologies (Kings Langley, Hertfordshire, UK). For more information, visit here .