The ever-evolving ML wearables landscape
In the last five years, the number of ML-powered medical devices has skyrocketed. As of October 2022 (the latest update), FDA approved 521 AI/ML-powered medical devices, with the astounding majority belonging to the radiology sector.
In addition to radiological devices, machine learning algorithms are finding their way into healthcare apps for wearables as well. For instance, in 2022 FDA approved the Atrial Fibrillation (AFib) History Feature for Apple Watch that analyzes pulse rate data and identifies episodes of irregular heart rhythms suggestive of AFib.
For sure, Apple Watch is not the only wearable device that uses machine learning to analyze health data. Garmin is well known for its range of advanced fitness trackers and smartwatches that leverage ML models to provide a holistic view of an individual’s health. For instance, for seasoned athletes Garmin offers Training Load and Training Effect features. These features are powered by the embedded Firstbeat Analytics™ engine that monitors the heart rate and transforms it into EPOC (Excess Post-Exercise Oxygen Consumption) values to calculate the impact of each workout session.
How about smart glucose monitoring and predicting high or low glucose concentration values in time to prevent hypo and hyperglycemic events? The Dexcom G6 Continuous Glucose Monitoring (CGM) system does just that – a tiny sensor embedded underneath the skin sends real-time data to a compatible smart device or a receiver. And when your values go too low, the predictive alert will give you a 20-minute advance warning.
Computer Vision is another promising subset of machine learning models that holds great potential for integration into edge devices. Take a look at Skinive. While primarily a mobile application rather than a wearables-specific one, its CV-based skin screening technology is capable of accurately identifying different skin diseases and conditions by using a smartphone camera. Now, imagine how much faster and easier diagnostics can become with such tools at hand.
But there is no fun in just enumerating further examples of ML-powered medical and fitness wearables – a simple Google search will do just fine. Instead, why don’t we dive a bit deeper and see the magic that unfolds within those devices.
What happens under the hood?
First, let’s underline the fact that modern wearable technology has experienced significant advancements. Equipped with sensors, microprocesses and the Internet connection, some of these wearables like Apple Watch have enough power and memory to run pre-trained machine learning models. This means that the heavy lifting of model training happens on a more powerful machine and the model is then deployed to a wearable device, giving it the ability to recognize patterns and act accordingly, i.e. raise an alert.
But even with all the progress happening to wearable technology, running intricate ML models continuously can rapidly deplete the device's battery life. Thus, the second and more practical option is when a wearable collects and pre-processes data and transfers clean and filtered data to a smartphone for edge computing. TensorFlow Lite is designed specifically for that purpose. It enables developers to deploy lightweight machine learning models on smartphones and edge devices, with optimized efficiency and minimal resource consumption.
The third option is to send this data into the cloud. This strategy is particularly advantageous when we require comparative analysis or in-depth examination, as cloud computing provides unparalleled processing power, vast storage capabilities, and the flexibility to scale resources based on demand. This ensures efficient handling of large datasets and complex computations, often delivering more accurate and comprehensive results.
Navigating the labyrinth of challenges and limitations
Just like with any technology, it always comes down to balancing potential and inherent challenges. And implementing machine learning models on wearables is not without its own set of difficulties. We have already touched upon one such problem caused by the compact nature of wearables – limited battery life and processing power. Imagine running a marathon with a backpack full of bricks – that's how an ML model might feel on a wearable, draining power at every computational turn.
Another major hurdle is properly labeled medical datasets. For CV models to work, you need a labeled set of data that the model can learn from to make correct decisions. The challenge is that properly labeling medical data requires specialized domain knowledge and expertise. For accurate and meaningful annotations, medical professionals or trained annotators with a deep understanding of medical terminology and concepts are needed. And that can be neither cheap nor easy.
While we are on the subject of data, it’s necessary to note that even the most accurate ML models won’t be any good if input data is unreliable. And the quality of data being collected hinges largely on the sensors and hardware used in the wearable. Needless to say that medical-grade wearable devices produced in compliance with the strict industry requirements provide more accurate data compared to their consumer-grade counterparts. Seamless data transfer with no data duplication Data duplication
Finally, data synchronization can present a certain challenge as medical wearables often collect data from multiple sensors simultaneously, be it heart rate, oxygen saturation, skin temperature, or movement sensors. The time stamps associated with different data points must align perfectly as even a slight misalignment can cause issues. For instance, if a heart rate spike isn't correctly synchronized with an activity log, an ML model might incorrectly identify it as an anomaly.
We are living in a world where our smartwatches do more than just tell time – they keep an eye on our hearts, our blood sugar, and even predict potential health hiccups. This fusion of wearable tech and machine learning can be truly groundbreaking but the journey is not without its bumps. From power-hungry algorithms to perfectly labeled medical datasets, the challenges are real. But with every challenge we overcome, we are a step closer to making healthcare truly proactive and personalized.