Validating AI-driven digital biomarkers

The convergence of artificial intelligence and digital biomarkers isn’t speculative anymore; it is a present-day reality, and it’s changing how we measure the impact of new therapies.

As we’ve discussed before, we are firmly in ‘The Age of Health Biomarkers as Clinical Endpoints‘, with wearables and sensors generating real-world, real-time data that extend clinical insight far beyond clinician’s offices and the hospital walls.

AI, and more recently GenAI, have become the catalyst to revolutionise this transformation, capable of identifying subtle patterns in high-dimensional data – from gait and speech to sleep and heart rate variability – that humans cannot detect alone.

However, with this new power comes a profound challenge that the life sciences sector is now facing: how do we validate systems that don’t just execute instructions, but learn and evolve?

The validation conundrum: From deterministic to probabilistic

Traditional software validation is built on a deterministic foundation, to test if specific inputs consistently produce expected outputs. 

An AI algorithm, however, is probabilistic. Its behaviour is not fixed; it shifts and adapts as it processes more data. This fundamental difference means that simply verifying code and testing against a static dataset is no longer sufficient.

Validating an AI system, therefore, requires a new approach that acknowledges its inherent adaptability and non-deterministic nature. Validation must focus on evaluating the model’s performance and robustness across a broad spectrum of data, including edge cases and adversarial examples.

This shift involves assessing the AI’s ability to generalize to new, unseen data and its stability when faced with unexpected or manipulated inputs. The goal is to ensure the AI’s predictions are not only accurate but also reliable and fair, even as its internal state evolves with new information.

Traditional validation methods like unit and integration tests are still useful, but they must be complemented by techniques that continuously monitor and re-evaluate the model’s behavior in dynamic, real-world environments

Explainable AI (XAI): Opening the black box

For regulators and clinicians alike, trust is paramount.

If an AI model flags a digital biomarker as a potential safety signal, the immediate question is not what, but why. In a clinical context, an unexplainable recommendation is an unactionable one.

This is where Explainable AI (XAI) should be considered. Now enshrined in the EU AI Act, methodologies like SHAP and LIME, which provide insights into which features most influenced a model’s output, are becoming standard.

For clinical development leads, this means demanding solutions that offer transparent, visual dashboards that can trace a conclusion back to the underlying data, making the AI’s reasoning auditable.

The human-in-the-loop: Augmentation, not automation

The current consensus, and one we wholeheartedly support, is that AI should augment, not replace, the clinical expert. As we’ve explored in ‘AI in 2025: Are We Exploring, Executing, or Stuck?‘, the most mature AI strategies focus on execution where human oversight is a given.

In practice, this means building systems where the AI provides a first-pass analysis, flagging anomalies or patterns of interest. This leaves the final diagnostic conclusion or clinical decision firmly in the hands of a qualified professional who can apply context, experience, and empathy – qualities no algorithm can replicate.

The power of partnership in a specialised field

Navigating this complex intersection of data science, regulatory affairs, and clinical practice is not to be done in a silo. The industry’s most successful initiatives are born from collaboration, pairing the deep clinical expertise of pharmaceutical companies with the specialised technical knowledge of AI validation partners.

We recently partnered with a pioneering biotech firm to validate an AI model designed to analyse speech patterns from a mobile app as a novel endpoint for a Parkinson’s disease trial.

By co-designing the validation protocol from the outset, incorporating robust statistical monitoring for model drift, and building an XAI dashboard for their clinical team, we were able to provide the evidence package needed for regulatory submission with confidence. This collaborative approach can reduce protocol amendments by as much as 30%, according to a recent Cytel survey, saving both time and significant cost.

As AI’s role in clinical development continues to grow, robust, transparent, and collaborative validation will be the foundation that supports scalable innovation. Done right, this ensures we can harness the full potential of digital biomarkers to bring safer, more effective treatments to patients faster.

If you’d like to connect with our digital health team, reach out today and schedule a call.