What makes a biomarker explainable — and why it matters clinically

Suppose a system tells a clinician that a patient's cognitive signal has worsened. The obvious next question is: why? Which part of the signal moved, by how much, and does that pattern make clinical sense? If the model can't answer, its number is hard to trust and harder to act on — no matter how accurate it is on a benchmark.

This is the tension at the heart of applied clinical AI. Many high-performing speech-based models are effectively black boxes. A recent systematic review of explainable-AI methods for speech-based cognitive-decline detection identifies opacity as a real barrier to clinical adoption, and notes that regulatory frameworks increasingly expect explanations for automated decisions.

Accuracy and explainability are different properties

It's tempting to treat a high accuracy number as the whole story. But accuracy tells you how often a model is right on a test set; it says nothing about why it reached a given conclusion for a given person, or whether that reasoning is clinically sensible. The broader field of explainable AI (XAI) exists precisely because performance and interpretability don't come bundled — and in high-stakes domains, you need both.

What "explainable" should mean in practice

Explainability isn't a single thing. Useful distinctions include:

• Interpretable-by-design vs. post-hoc. A model built from meaningful, inspectable features is different from a black box you try to explain after the fact.
• Local vs. global. Explaining one person's result is different from explaining the model's overall behaviour. Clinicians usually need the local answer.
• Feature attribution. Showing which signals contributed to a result — and in which direction — so a human can sanity-check it.

The point of all of these is the same: let a qualified human see the reasoning, catch when it's wrong, and stay in control of the decision.

Why it matters clinically

Three reasons, beyond good manners. First, trust and oversight: a clinician who can see the contributing signals can integrate a tool into their judgement rather than defer to it — reducing the risk of automation bias. Second, error-catching: transparent contributions make it easier to notice when a result is being driven by an artefact (background noise, a recording glitch) rather than genuine change. Third, regulation: explainability is moving from best practice toward expectation in health AI.

How Recalla approaches it

Recalla is explainable by design. Rather than emitting a single opaque number, every composite decomposes into the signal families that produced it — so a clinician can see which family moved and by how much. The system is non-diagnostic and human-in-the-loop: it's built to inform a professional's judgement, not to replace it.

One honest caveat worth stating plainly: explainability is not validation. An explainable score can still be wrong, and interpretability methods carry their own caveats. Being able to see the reasoning is necessary for clinical trust — but the reasoning still has to be proven right, which is the work our Evidence Hub tracks.