Who owns your wearable data?
The data collected by wearables and medical IOT devices should belong to users. Manufacturer data silos and proprietary AI are stopping wearables from revolutionizing healthcare. User data needs to be liberated from device manufacturers like Apple, Google, Samsung and Withings as closed ecosystems are limiting the value of the collected data and leading to subpar user experience. People should also have the option to securely donate their data to science. Federated machine learning and smart contracts could allow people to confidencially contribute their data to medical research.
My experience of health IOT devices
I love biohacking and tracking my personal data. As a self-confessed geek, I enjoy experimenting with gadgets and using them to discover more about my own body rhythms. I have the following smart IOT personal devices: smartwatch, under-mattress sleep detector, sleep recording snore pillow, scale, body composition tracker, EEG sleeping headphones, blood pressure monitor, peak breath flow meter, thermometer and an over night blood oxygen reader.
All this tech has taught me the following:
- Work stress hugely affects my sleep.
- Bad sleep is the root of all my personal evils – everything from managing diet, to willingness to exercise.
- None of the readings ever match each other – I have multiple devices that track the same metric and the devices rarely agree.
- The apps don’t talk to each other – I cannot get a single view of all my data.
- Apps are not great.
- Syncing is always buggy.
Recently I found the great YouTube channel of Rob ter Horst, a postdoctoral researcher at CeMM (the Research Center for Molecular Medicine). He spends 11 hours a week methodically measuring and detailing everything about his life using high-quality equipment and tests (temperature, sleep patterns, blood work, urine samples, mood diary, food diary, activity tracking).
Firstly, I was astounded that someone would have the dedication to track this amount of information for two years and secondly, a benefit of the meticulous tracking was that Rob could independently assess the quality of wearables from Fitbit, Withings, Oura and Dreem. His research has confirmed my own impression – the quality of wearables data can hugely fluctuate. For example step counters can overestimate by up to 30%. Below is a chart he shared comparing four activity counters:
If the majority of the devices measure the same thing why is the quality so different?
A great benefit of smart wearables is that they are easy to incorporate into everyday life. Except for having to remember to charge them, watches, bands and rings work tirelessly in the background collecting data. This perfectly suits the average user, who in general wants maximum data for minimal effort. This limits the majority of data collection to temperature, motion, location, oxygen level, skin impedance, and pulse rate. The similarity between sensors is why, whenever a new sensor is released, there is such a huge fanfare! One of the main features hyped at the release of the Apple Watch 4 was the inclusion of an ECG sensor. The fact the device received FDA De Novo classification gave the device gravitas.
The other common personal IOT device is the smart scale. The humble bathroom scale first became popular in the 1960s and for many years this was the only biohacking device people could easily access. When the first smart scale launched in 2009 it promised to revolutionize our health by accurately tracking weight and body composition (something I don’t think really happened).
AI is the real differentiator
Raw sensor data is only half of the problem. What people want is actionable insights and not raw tracking data (see above image). The only way a smartwatch, phone or sleep band can tell if you are cycling, running, sleeping, or snoring is by using AI to interpret the raw sensor input. It was only through advancements of machine learning that smart IOT devices were able to make sense of data. Traditional deterministic algorithms would be too complex to write (and not very good). It was only with supervised machine learning that the raw data could be quickly turned into meaningful output.
Due to the need to go to market quickly devices are often released before the AI components have been fully trained. This is why many IOT devices often need regular firmware updates and don’t work very well when first released. In general, the only differentiation between devices is battery life, aesthetics/ui, and the sophistication of the AI.
The cynic would reason Fitbit devices overestimate “steps” because they want their users to feel “fitter” and have bragging rights. This is the only reason I can think of why a $20 USD Xiaomi Mi Band 3 would more accurately detect activity than the Fitbit Charge 3 (but I’m willing to be told otherwise).
Is there something better we can do with all this data?
An issue with all the personal IOT data is it is locked in silos. There is an ecosystem war at play – with Apple (now the owner of Fitbit), Google, Samsung and Withings all competing for our data.
Whilst the data is locked away a user (or their doctor) can’t easily see a holistic overview of their health, nor can the data be used to improve healthcare, enable earlier diagnostics or assist with drug discovery. The elephant in the room is “who owns your data” – the correct answer should be YOU. But this isn’t the case – a lot of IOT and wearable devices don’t let users easily export their own raw data – sometimes the best a user can sometimes do is a lo-res screenshot.
Luckily technology can help us securely unlock our data.
Federated and Gossip machine learning
Federated machine learning turns AI training on its head. Rather than raw data being shipped to a central data warehouse for analysis federated machine learning performs the machine learning on the private device and only transmits the end result. The central server collects the anonymized set of partially trained models and, using smart mathematics, generates a global model. (Google has a great cartoon here explaining federated learning in detail).
Gossip learning extends the principles of federated learning but removes the central server. Devices communicate directly with each other, sharing their trained model, similar to peer-to-peer communication. Whilst this means each device needs to do more computation this should not be a problem. The latest mobile phones include AI chips and the training can be configured to only launch when the mobile phone is being charged.
Combining Open [APIs + Data + Algorithms] could set the data free
Federate and Gossip machine learning can help to secure data but what about if a user wants to share their data? For example, how can a user share their data with their doctor or hospital? Telemedicine is a growing field, but currently a downside is the sharing and collection of patient information is often through dedicated and siloed applications. Governments should promote and enforce standards for the sharing of personal IOT data, similar to PS2 / OpenBanking.
To solve the issue of the lack of standard APIs, start-ups are building APIs that connect to multiple devices and then harmonize the results. For example, Humanapi.co connects to 300+ devices and provides both wearable and clinical APIs.
In 2019, the Open Wearables Initiative (OWEAR) was launched. They describe themselves as follows:
OWEAR is a collaboration designed to promote the effective use of high-quality, sensor-generated measures of health in clinical research through the open sharing of algorithms and data sets.
OWEAR serves as a community hub for the indexing and distribution of open source algorithms. To identify performant algorithms in areas of high interest, OWEAR acts as a neutral broker to conduct formal and objective bench marking of algorithms in selected domains.
We create searchable databases of bench marked algorithms and source code that can be freely used by all, thereby streamlining drug development and enabling digital medicine.
The main innovation of OWEAR is the sharing of algorithms. By sharing algorithms, they can drive the price of wearables and IOT devices down (by reducing the investment required to develop bespoke high-quality algorithms) and therefore promote their adoption. For wearables and IOT to provide maximum benefit they have to be globally available and not just available to a few wealthy tech geeks; for this to happen they need to be cheaper to make.
If I can donate my body to science why can’t I donate my data?
A new interesting area that is evolving is the ability for people to securely donate their data to medical research.
The Data Donor Movement initiative is promoting the ability for patients to donate their data to medical research. The initiative is petitioning governments to make it mandatory for healthcare organizations to share their data. By sharing data the movement hopes to:
- Push healthcare toward prevention
- Aid doctors with early and better diagnoses
- Provide personalized treatment plans for patients
- Fuel medical research and lead to groundbraking discoveries in healthcare
In order to promote the donation of data the Data Donor Movement is firmly committed to only selectively sharing data:
|Will share with||Will NOT share with|
|Research Institutions||Insurance Companies|
|Hospitals||Advertising & Marketing Companies|
|Government-funded Health Organizations, Like Public Health in Canada||Social Media Platforms, like Facebook|
|Pharma||Unauthorized 3rd parties|
To facilitate the securing of data research has been conducted into using blockchain smart contracts as the enforcement layer. Due to the amount of data being collected the raw data cannot be stored on the blockchain but smart contractors could facilitate the access and decryption of anonymized data.
“Smart contracts allow the creation of agreements in any IoT devices which is executed when given conditions are met. Consider we set the condition for the highest and lowest level of patient blood pressure. Once readings are received from the wearable device that do not follow the indicated range, the smart contract will send an alert message to the authorized person or healthcare provider and also store the abnormal data into the cloud so that healthcare providers can receive the patient blood pressure readings as well later on if needed.”
A Decentralized Privacy-Preserving Healthcare Blockchain for IoT – https://www.mdpi.com/1424-8220/19/2/326/htm
New sensors + new data sources
In general it’s only when someone has an underlying medical issue that they will invest in additional data collection. For example, the most widely used personal medical device is a diabetes glucose monitor. Device manufacturers are aware of the effort barrier and the lack of differentiation of sensors, therefore they are putting increased effort into more advanced sensors. Here are some devices that are scheduled to be released:
- Smart Body Scanner – by using computer vision a body scanner can help you to better track your weight and track your body shape. Two examples are NakedLabs and ShapeScale.
- Smart Toilet – could detect a range of disease markers in stool and urine for example colorectal or urologic cancers. Likely to be especially arrtactive to those who are genetically predisposed to certain conditions, such as irritable bowel syndrome, prostate cancer or kidney failure, and want to keep on top of their health.
- Glutrac – have developed a watch that can track glucose levels with optical sensors. This could assist with diabetes management as well as potentially tracking diet.
- Dreem-Headband – can detect brain activity to track your sleep.
- Neuralink – Elon Musk’s brain and computer interface. Besides letting you control your surroundings it could be used to collect information on your health and brain activity.
The majority of product development has been into non-invasive sensors. The issue with these is they limit the type of insight that can be gathered. Companies like Thriva have stepped in to close this space. They provide monthly or quarterly home blood testing and deliver insights via an app. It would be great to have this data mixed in with the personal IOT data.
Links and References