Running clinical trials with behavioural biomarkers as digital endpoints in the age of AI regulation
As a professional in the clinical trial space, whether you’re working for a large pharmaceutical company, a digital therapeutics company, a clinical research organisation, or as a clinician or academic, you are faced with a wide variety of problems every day of the week. Planning, ethical approval, recruiting participants, working with technology suppliers, they all bring their fair share of headaches.
One of the harder problems to overcome that you may well have been facing if you work in brain and neurology-related areas is that of unreliable and noisy endpoint data from your participants. Studies of brain health, neuro-developmental/degenerative or neuro-muscular conditions often rely on questionnaire-based self-assessments completed by the participants, known as patient-reported outcome measures or PROMs in the trade. These can be noisy for a range of reasons, including issues with memory, perceived stigma, desired outcome of a test, comprehension of the questions etc. There can also be problems with protocol adherence caused by boring tasks and barriers to attending a clinic. Participants may also perform or behave differently (usually making themselves look better than they really are) when they know they are being observed (aka the Hawthorne effect).
Ultimately, this results in lower amounts of more noisy outcome data. That is, an unknown difference between a participant’s unknowable real state and their self-reported state. This lowers the signal-to-noise ratio, which makes your statistical analysis far more difficult: you will have to power your protocol to adjust for this noise and the lack of adherence, meaning that you must recruit more participants. Sometimes, if you work on a rare disease, this means that you need more participants than there are available and your study is at risk of failure. Sadly, too often this is exactly what happens for studies of breakthrough treatments of e.g. rare neurodegenerative diseases .
Digital behavioural biomarkers to the rescue!
Recently, you may have been reading about a new type of biomarker that can be used as a digital endpoint and provides a solution to this noisy data problem: digital behavioural biomarkers. These replace the boring subjective error-prone questionnaires with objective measures such as facial muscle activations, gaze patterns, pupil dilation, head actions, tone of voice, and others. Sometimes these are combined with strong indications of disease-specific behaviour, for example gaze avoidance during dyadic interactions or motor retardation of the voice in depression or anxiety. They can be administered on people’s mobile phones and integrated in engaging interactive designs meaning that they also address the ecological validity issue (it’s better to measure someone in their natural habitat than in a hospital) and the Hawthorne effect. All of which should result in better adherence. Like other decentralised clinical trial approaches, it also cuts staff and trial site costs. My, you should really try this for your next study!
But what about the EU AI Act and other AI regulations? Isn’t this prohibited now?
To answer the most impactful question first: no, using AI-based face and voice biomarkers is not prohibited.
It is a high-risk application of AI though. This means that whoever supplies you with digital biomarker services or software should be EU AI Act compliant as evidenced by new standards certification, such as ISO 42001 or adherence to the NIST AI Risk Management Framework.
What you want to use is a solution that respects your participants' privacy, that is designed with ethics, security, and data protection from the outset.
But how can you privately analyse data that is impossible to anonymise, such as face and voice data?
There is a solution to this paradox! If you take the algorithm to the user and run the face and voice analysis on a participant’s own mobile phone, you can solve this problem.
This is exactly what BLUESKEYE AI’s B-Healthy Platform solution offers. Applying our unique two-stage approach to medically relevant behaviour analysis, the original face and voice data is turned into a couple of hundred anonymous parameters for every video frame, that together describe the behaviour. These parameters include things like facial muscle actions, gaze direction, or the tone of voice. These are then sent over a secure, encrypted channel to a secure backend where the anonymous data remains encrypted at rest. You as the clinical trial manager can then inspect this data, run data science projects on it, or combine it with other anonymised data that you have available through back-to-back server integration for example. All your digital biomarker research can thus be done on entirely private data. A very useful side-effect of BLUESKEYE’s two-stage approach is that all appearance-based biases and invariances are solved at the first stage - it simply HAS to because the image and the audio waveform are discarded afterwards. BLUESKEYE has been proven to be unbiased to apparent gender, age, and ethnicity at this first stage!
Two-stage approach to face and voice analysis
To make it even easier for you to deploy a clinical trial that includes face and voice behavioural biomarkers, BLUESKEYE AI’s B-Healthy Platform provides you with configurable mobile phone modules that you can use to create a data collection app by taking a pick-and-mix approach. By selecting from a set of configurable modules from a library our team can create an app for you in a matter of days. Example modules are participant information and consent forms, video or audio stimuli, and data feedback screens. Most importantly, you can select the interactive task for the participant to perform whilst being recorded. Existing modules in the library include a journalling task, an instruction-following task, a read-out-loud task, an object-tracking task, and others. If there’s a task or a module that you need and that doesn’t exist in the library yet, speak to the team to see if they can fast-track that, we’re always keen to learn what our customers need!
The B-Healthy Platform is already used by a number of pharmaceutical companies in their studies and is now being rolled out to the wider clinical trials audience. As an ethical AI company, we are well underway with being compliant with the EU AI Act and other AI regulations and you can safely use our services now and in the future!
Interested to learn more about the B-Healthy Platform? Read more about it on the BLUESKEYE AI website, or get in touch with info@blueskeye.com.