Validated Endpoints Exist. Why Trials Still Don't Use Them

In most central nervous system trials, the primary endpoint is still a number a human assigns after watching and interviewing a patient. That judgment varies from rater to rater, and the variation is costly: rater variability is one of the documented contributors to the high, unstable placebo response that makes CNS trials hard to power and easy to fail. The counterintuitive part is that a fix already exists. Objective, video-derived measures have been validated against trained raters in peer-reviewed studies. The open question is no longer whether they work. It is why most trials still do not use them.

Why do subjective rating scales create noise in CNS trials?

Subjective scales create noise because the score depends on who is rating, and in CNS indications that variability compounds a placebo response that is already high and difficult to predict, especially because subjective endpoints are more likely than objective ones to produce amplified effects when blinding is imperfect.

One mechanism behind this is functional unblinding, which is simpler than it sounds: when a rater can guess, from side effects or demeanor, whether a patient is on the active drug or placebo, their scoring can drift in the expected direction. Multiply that drift across sites and raters, and the trial has to enroll more patients to detect a real effect through the noise.

None of this means rating scales are worthless. They capture the lived experience of symptoms in a way nothing else does, and patient-reported outcomes remain essential; Videra Health administers them natively alongside objective measures. The problem is narrower: relying on human-scored scales alone for signal detection builds avoidable variance into the primary endpoint.

Are objective, video-derived endpoints actually validated?

Yes. In peer-reviewed studies, video-based machine-learning models have matched or outperformed trained human raters on clinically meaningful measures. In a 2025 study, a video-based model detecting tardive dyskinesia reached an AUC of 0.89 in a combined validation cohort, measured against a trained-rater standard, with higher sensitivity and specificity than the raters themselves.

A separate cross-dataset validation of a video-based model reported a Cohen’s kappa of 0.61, above both calibrated human raters (0.57) and uncalibrated ones (0.37), with an AUC of 0.85, sensitivity of 0.79, and specificity of 0.82. A kappa in that range reflects substantial agreement, and the point worth sitting with is that the model agreed with the reference standard more consistently than the trained humans did.

The approach is not limited to movement disorders. A 2025 study in npj Digital Medicine found that a single open-ended video question, analyzed for both language and voice, could predict validated scores for depression, anxiety, and trauma while cutting assessment time by about 65 percent. Fewer than two percent of participants declined to use video-based screening.

If the measures work, why don’t more trials use them?

The barriers are about trust and fit, not validity. A sponsor adopting an objective endpoint needs peer-reviewed evidence they can defend to reviewers, a measure that slots into the existing evidence package and current regulatory expectations, and an approach that does not add burden at sites that are already stretched. Each of those is a real hurdle, and none of them is the question of whether the measure is accurate.

Operational fit is often the deciding factor. An objective measure that requires new hardware, long training, or a parallel workflow will lose to the familiar scale even when it performs better, because the people who have to run it every day will not use it. The adoption question is won or lost at the site.

What does it take for a sponsor to trust an objective measure?

It takes three things: peer-reviewed validation against the current standard, a clean fit into the trial’s evidence package and regulatory expectations, and low enough site burden that investigators will actually use it.

1. Validation that holds up. The measure should be published, tested across datasets, and benchmarked against the raters it is meant to supplement, so a reviewer can see it meets or beats the current standard.
2. Evidence-package fit. The endpoint has to make sense within the trial’s broader evidence and align with regulatory expectations, rather than sitting beside the package as a novelty.
3. Site burden low enough to adopt. If capturing the measure is as simple as a short video on existing devices, sites will use it; if it is not, they will not.

Adoption in practice tends to follow that pattern. In one Phase IV study, an AI-powered eCOA approach was deployed mid-study across 45 sites, restored investigator confidence in the data, and made room for novel AI-driven endpoints, precisely because it fit the existing workflow rather than replacing it. Videra Health works in this role as a neutral, evidence-generation partner, supporting objective measurement without touching prescribing or commercial decisions.

The evidence that objective measures work is already in the literature. The next move for a sponsor is narrower and more practical: pick one endpoint where rater variance is hurting signal detection, and ask what a validated objective measure would do to the sample size and the placebo arm. That is a question the data can already answer.