Speaker: Helen Brooker, Senior Product Manager – Advisory, Signant Health
Key Takeaways
• Traditional monitoring reviews only 20% of follow-up visits — leaving the majority of rater-patient interactions unexamined
• ReviewAssist uses AI algorithms to flag administration and scoring errors for mandatory human review, ensuring 100% coverage
• Early deployments achieve that 100% coverage at just 15–20% human review rates
• Fixed, scale-specific algorithms prevent unpredictable AI behaviour — the model is trained, validated, and deployed without autonomous drift
• Visual scale components (drawing tasks, motor assessments) require continued human-only review — a deliberate design boundary, not a gap
The pharma industry faces an uncomfortable truth: financial constraints force systematic compromises in monitoring quality that directly put trial outcomes at risk. Traditional central review approaches examine 100% of screening and baseline assessments but that scrutiny can reduce to 20% random sample at follow up visits, leaving the majority of rater patient interactions completely unreviewed.
Helen Brooker, Senior Product Manager, Advisory at Signant Health, opened her Pharma Clinical Innovation USA 2025 session by naming this tension plainly. In an ideal world, sponsors would review every interaction at every time point to protect endpoint reliability, to defend data quality before regulators, and to ensure nothing is missed. The reality is that comprehensive human review across all visits remains financially out of reach for most programmes.
"Trial outcomes have been missed not because the treatments didn't work, but because data inconsistencies and gaps in oversight obscured the signal. We should all be honest about that."
The consequences extend well beyond compliance risk. When a pivotal trial fails to demonstrate efficacy not because the therapy lacked effect but because measurement error went undetected, an effective treatment may never reach patients. That is the true cost of systematic monitoring gaps.
The subjectivity problem in clinical endpoints
Clinical trials for cognitive, psychiatric, and neurological indications depend heavily on subjective assessments administered by trained raters using validated scales: the MMSE for cognitive function, MADRS for depression severity, PANSS for schizophrenia symptoms, ADAS-Cog for Alzheimer's cognition. Despite rigorous training and certification, these interactions carry inherent variability.
Brooker framed the challenge through a simple but precise observation: two administrations of the same scale by the same rater on different days can yield different results not because the patient has changed, but because human performance varies. The core of the assessment may be consistent; the execution rarely is perfectly uniform.
Central review processes exist to address exactly this. By having independent reviewers evaluate recorded rater patient interactions, sponsors can identify administration errors improper question phrasing, over coaching, skipped items as well as scoring inconsistencies such as misinterpreting responses or making calculation errors, before these issues compromise efficacy analyses. The approach is established, effective, and trusted by regulators.
But it operates under severe economic constraints. Brooker posed the question that had driven Signant Health's innovation effort: could sponsors get more coverage without a proportional increase in cost?
Algorithm-driven triage with human oversight
ReviewAssist is Signant Health's answer to the monitoring coverage gap. The system uses AI trained on scale specific administration and scoring guidelines to triage rater patient interactions flagging potential issues for mandatory human review while clearing low risk sessions from the review queue.
The technical process begins with transcription of audio recordings. Trained algorithms then assess whether the administration followed protocol and whether scoring reflected guideline standards. Each interaction receives a disposition: either no issues detected, or potential errors identified requiring human review.
"AI is good with guardrails and with keeping the human at the end of the journey. That's not a limitation of the technology. It's a deliberate design choice."
Critically, human reviewers evaluate flagged cases without prior knowledge of the AI's specific findings. This preserves independent clinical judgement while using AI to determine where that expertise is most needed. The algorithm expands coverage; the expert makes the call.
Brooker was explicit about what this architecture is and what it is not. ReviewAssist is a triage mechanism. It is not a decision maker. The AI reviews 100% of interactions and identifies which sessions warrant expert scrutiny. The human reviewer does the rest.
Scale-specific validation and transparent limitations
Brooker gave significant attention to ReviewAssist's constraints, a degree of transparency that is unusual in vendor presentations, and that reflects a deliberate strategic choice.
The system can only assess what audio transcription captures: verbal questions, spoken responses, audible coaching or prompting. Visual components of clinical scales drawing tasks within the MMSE, motor assessments in other instruments sit outside AI evaluation and require continued human only review. This is a defined boundary, not an oversight.
ReviewAssist's applicability is therefore scale dependent for scoring but evolving everyday. Current validated implementations cover MMSE, MADRS, PANSS, HAM-A and ADAS-Cog, all scales with substantial verbal content. Expanding to additional instruments requires dedicated training and validation for each; Signant continues that work, but deploys only what has been properly validated.
"This isn't about unleashing AI into our systems and letting it evolve past what we've taught it. It's precisely the opposite. We train for a specific scale, validate the model, deploy it and it stays there with continuous monitoring."
This design directly addresses regulatory concerns about AI reliability. ReviewAssist does not use continuously learning models that could drift from validated performance over time. When guideline updates occur or performance issues emerge, Signant retrains and revalidates rather than allowing autonomous evolution during deployment.
Early performance data shows the AI detects issues at a sensitivity rate comparable to trained human reviewers, with a deliberate bias toward over-flagging rather than under-flagging. Some administrative nuances are picked up by the algorithm more consistently than a central quality reviewer might in routine practice, a characteristic that aligns with quality priorities. Better to review an unnecessary session than to miss a genuine error.
Brooker was candid that ReviewAssist remains an early stage tool, with real-world evidence still accumulating. The performance profile across different endpoint types continues to develop as sponsor deployments generate data. That honesty about maturity may prove to be a competitive advantage: in an environment where AI capability is routinely overstated, sponsors evaluating monitoring tools can make properly informed decisions about where and when to deploy.
Economic and strategic implications
The business case for ReviewAssist operates at two levels: immediate cost efficiency and long-term risk reduction for endpoint reliability.
Early deployment data shows actual human review rates of 15–20%, compared to the traditional 20% random sampling baseline. Sponsors are therefore achieving materially superior coverage 100% AI triage versus 20% random selection.
The more significant value proposition, however, is risk reduction. Random sampling at 20% means that 80% of follow-up interactions receive no quality review whatsoever. ReviewAssist eliminates that exposure: every session receives algorithmic assessment, and human review is deployed where it matters most, something not possible with random sampling.
"For sponsors with nine-figure investments in pivotal programmes, the risk-return case for this kind of coverage doesn't depend on cost savings. The cost of an undetected measurement error is a failed submission. Or a therapy that never reaches the patients who need it."
Brooker addressed the industry's broader hesitation about AI adoption directly. The dominant concern that AI systems will behave unpredictably or evolve beyond their guardrails reflects a misunderstanding of how well designed clinical AI actually works.
ReviewAssist demonstrates a different model: fixed algorithms, mandatory human oversight, scale-specific validation. It addresses the regulatory and quality concerns that make sponsors cautious while delivering the operational value that makes adoption worthwhile. The goal of 100% monitoring coverage long considered an unaffordable ideal becomes achievable not by replacing human expertise, but by deploying it where it can have the greatest impact.
In order to get you the highlights of Pharma Clinical Innovation USA 2025 faster, we are using generative AI technology to summarise the transcripts of the sessions. The conference organiser is checking the summary for accuracy. If you have any feedback about the summary, please contact lucy.fisher@thomsonreuters.com
Discover more on this topic at Pharma USA 2026 (March 17-18, Philadelphia) - North America's largest cross-functional pharma gathering. Visit the website here.
