The rise in wearable technology has been substantial in the past decade. With the rapid introduction of new and improved technology, health data and other lifestyle insights are becoming increasingly accessible. The data streams are increasing in number as devices can house several sensors and collect massive amounts of data over the course of the average day. Most pieces of technology will be accompanied by software which can turn this data into more digestible metrics for the end user. This helps turn hard-to-understand or overwhelming amount of data into relatively intuitive and actionable metrics. There is no doubt some of these metrics are extremely helpful and potentially offer a wide range of benefits to health and lifestyle optimisation. But with all this access to data, a challenge is created in processing and understanding the relationships within and between each strand of data.
Table of Contents
How Wearables Turn Sensor Data into Health Insights
Artificial intelligence (AI) has now become a household term, with many software programmes introducing some form of AI to act as an assistant or support system to the software. What were once fixed algorithms have now advanced to become flexible problem solvers and thinkers, allowing us to look at the data in many ways and make very thorough appraisals in a matter of seconds. Traditionally, pilot studies might take months of data collection and even more to fully explore the meaning of the data, whereas AI can complete some of these processes almost instantly, offering real-time feedback on what the data shows. Simple curiosities can be answered in seconds and complex analysis can be conducted faster than before. This allows us to identify trends and relationships within our data that weren’t possible before.
Wearable technologies have adopted the use of AI for obvious reasons. The data being collected is vast and under highly variable conditions. Relationships between the data streams, and even within them, can be extremely hard to detect over the course of the day. We are biological beings, between the variance in typical physiological function and the potential noise and limitations of biotechnology, the task of sorting through this data and cleaning it is immense. AI has become extremely useful for this.
AI Can Only Interpret the Data It Receives
We have adopted AI with great trust, and in many cases, it has allowed us to become timelier and more efficient in using our own personal data. There is a major concern, however, which rarely seems to be considered. AI is incredibly clever and efficient with the data it receives, and it can handle it perhaps better than any human can in terms of analysis. But it can only assess the information it has. In many cases, the source data is subject to noise and interference, which has already been smoothed and cleaned. AI only sees what it is given, and if prior processing has already occurred, it may wash out some of the detail it receives. In some cases, this may water down the resolution of the data, which can hide subtle pieces that are exactly what you want the AI to be able to detect in the first place.
Sensor Accuracy and Signal Quality Come First
For many novel sensor technologies some time is required to tease out the real-world applications. A reliable contract research organization will do the initial heavy lifting in regard to showing true effectiveness and proof of concept. What is accurate during tested laboratory conditions, can be vastly different when used in daily life. Sensor placement and installation can be critical components of usable data. This is especially evident with continuous glucose monitors (CGM) where interaction with clothing is known to create signal noise. As a result, upper arm and abdomen placement have become popular despite many installation sites being possible.
Another example of real-world learnings came from CGM, where low blood glucose during sleep was noted for being very common. Lying on the arm with the sensor creates compression and restricted blood flow, artificially reducing the glucose in the local tissue giving falsely low readings. This has nothing to do with lab-tested accuracy and validation. It is a product of real-world use that took some time to recognise and explain.
In short, sensors not only need to be accurate and validated in controlled conditions, but they also need to be able to maintain effectiveness when subject to daily life. Factors like heat, moisture and movement must not significantly disrupt the sensor signal or it will become ineffective too easily to be of any real value.
The Missing Context Problem
As mentioned, laboratory based testing is often the main trial a sensor must pass, but laboratory-based validation does not always provide full use context. For example, lifestyle trackers may be effective at monitoring 24 hr heart rate but have no ability to discern workplace stress, from elevated physical activity. So, we might have a visible elevation in heartrate caused by a brisk walk to catch a bus, or an extremely stressful workplace situation. One of these will have little consequence, the other could have downstream impact for the rest of the day, impacting appetite, sleep and cognition. Context is key and there is little way to do this other than an additional log or note taking. Even with an additional note it is difficult to assimilate that information with the hard data inside an app. AI has the potential to address this but it is very hard to design an algorithm without numbers, it leans heavily on the intelligence and problem solving within the Ai system. AI coaching is an area undergoing major development at present.
The Coffee, Heart Rate and Sleep Example
Of bigger concern is what is factored into the decisions. AI can only use what it has available to come to its conclusions. For example, we might look at something as simple as sleep. You might wear a smartwatch throughout the day, which tracks sleep during the night and heart rate over 24 hours. Let’s say you are a coffee enthusiast drinking multiple cups of strong coffee throughout the day. Your heart rate will more than likely be elevated, and your sleep might be recorded as being reasonably poor and fitful. While correct, the action to be taken might not be the right one to solve the issue. AI might interpret your HR variability as a sign of stress, which would be linked to the poor sleep. You might think that stress is the issue when, in reality, the coffee is the source of the problem. AI does not know about the amount of caffeine you consume across the day. It does not have that critical piece of information to include in its assessment of the situation.
When a Correct Analysis Produces the Wrong Explanation
While in the example AI might be doing an excellent job in assessing the information it has, the critical piece that is missing would change the picture considerably, yet it is not factored in. There are many instances where this issue could occur, leading to highly misleading and perhaps incorrect appraisals being made. Subtle relationships may be pulled to the fore while the more substantial interactions may remain hidden.
As we continue on the road to AI integration into our health management, we must consider these issues. First, we must have some acknowledgement of the limitations of the sensors themselves and how these are processed into their end-product metrics. If heavy smoothing or interpolation is used, then we need to give that some consideration. Sensor lags and interpolation techniques can substantially wash out data to the point where it may not actually reflect much of anything. This is more common than many would acknowledge. But the bigger piece, and the one needing far more awareness, concerns what is missing. There are very few pieces of technology that can capture every aspect of your daily life and every physiological metric available. You might even say it is impossible to do so. But we expect very precise and advanced advice to be given by some of the current systems. It may be a case where we are expecting too much in terms of definitive answers, as these are rare in the biological world.
A Health Metric Is Not the Same as a Clinical Conclusion
While the developments in wearable technology and AI integration have increasing levels of clinical validity, the threshold for this has become somewhat of a grey area. Many argue wearable information does not need to be as accurate if it is serving the purpose of general prevention and lifestyle guidance. But what if someone believes a certain metric is of diagnostic level when it is not. An untrained individual can come to all sorts of conclusions with inconclusive data. Interpretation at the clinical level requires some expertise. A reputable human performance lab or clinic is where an individual should follow up to get a more complete picture. While the knowledge gaps are closing and in-app insights improving, it does not match specialist knowledge. Self-diagnosis and treatment comes with great risk, and unfortunately a false sense of confidence, which might delay proper diagnosis and treatments of more serious illnesses.
Many argue that everyone has the right to their own data, and while devices may not be clinical diagnostic tools, that should not prevent the data from being available. Regulatory bodies have historically monitored these areas to try and create a clear boundary between personal use and medical support. However, these distinctions are constantly being challenged by device manufacturers and frameworks have been loosened as a result.
The Future of Wearables Is Context-Aware, Not Just Data-Rich
In recent years, there has been an explosion of data-driven insights from the rapid development of technology and AI. For most of this period there has been nothing but support for the increase in accessible personal data. However, as we get more familiar with our own data and access becomes less of a novelty, we start to ask the question ‘So what?’. What does it really mean to have a fitness age of ‘x’ years. Are these insights actionable? Are they meaningful? Do they have any impact on health? The reality is that a lot of this data is more limited and contextual than we might like to think. Chasing longevity through the perfect lifestyle is going to have little impact on a major accident for example where it can all be undermined in a second.
Tracking all of these metrics can drive severe anxiety, as we become obsessed with having an impact on something which is hard to influence. Some algorithms are quite insensitive to change and while they may indicate a baseline any further changes can be quite hard to see. Understanding how the metric is produced can unveil major contradictions to the theory behind them. Quite often this isn’t discussed and the process of metric generation may be quite crude. The responsibility is very much on the user to investigate what data is collected, how and in what way does that translate into an end value. With that, one must be informed on about the context of the metric and what their goals are. Sometimes these don’t align fully and can cause frustration.
Conclusion
The purpose of this discussion is not anti-AI by any means. AI is incredibly effective and useful, but it is doing very literal assessments, and its capability leads us into a false sense of acceptance. We will take its findings very literally without stepping back and making our own appraisal of what it is that has been assessed. A little bit of awareness and personal interpretation is important as a final sense check of the data. One must ask how much of the picture is being provided and really factor that into how they are going to make use of the data and advice given. AI can absolutely be used to streamline and quicken our ability to learn about ourselves and help us respond in efforts to optimise or improve. We must be mindful of the holes and appreciate that we must also take some responsibility for decision-making. AI will do the heavy lifting, but it is still up to us to read between the lines.