Improve Patient Outcomes Data with AI-enabled Voice Research
AI voice analytics, guided by expert linguists, helps clinical teams choose and shape outcome measures around what patients say matters most.
Every outcome strategy eventually comes down to a deceptively simple question: "What should we measure?"
As clinical research professionals, we know that answering it means grappling with some inconvenient truths. The PRO that comparable trials use was validated in a population that only partly resembles yours. The clinician-reported scale captures severity well, but patients describe their hardest days in terms it was never built to ask about. And endpoint decisions get made early, often before the team has heard from more than a couple patients.
With FDA guidance now anchoring COA selection to what matters to patients, getting this question right early is what makes the outcome data collected later worth having. So, it's fair to ask a new question: "Can AI voice analytics improve patient outcome data quality?"
We think it can, by strengthening what goes into outcome measurement before and alongside the instrument work. Let's walk through how we do it.
Three Tools, One Team
“AI voice analytics” can mean a lot of things, so let's be specific about what it means for us. Two technologies do the heavy lifting at scale, and our linguists guide how each one is used.
Large language models (LLMs) work with the language of what participants say. They power our AI Moderator, which follows up when needed on the part of a question an answer left incomplete, using that question's own wording. They also power inVibe Agent's expert-built workflows, which organize concepts across interviews and draft findings with the evidence attached. Every workflow is tested against analyses our linguists already understand.
Speech emotion recognition (SER) is a machine learning model that analyzes the sound of the voice rather than the words. It measures roughly 40 acoustic features, such as pitch, loudness, and pacing. It then describes each response's valence, activation, and dominance relative to the speaker's own baseline. Our linguists read those patterns as context for ‘how’ something was said.
Linguists keep humans in the loop at every step. They design the questions, judge whether AI outputs hold up as research, and keep every finding traceable to the real participants behind it, never to synthetic respondents or digital twins. They also hold the line we care about most: no trained linguist would mind-read beyond what is said, and our AI tools are held to that same standard.
How Voice Analytics Strengthens Outcome Measurement
Together, these tools change what a study team can learn before committing to an outcome measure. Here are four places we see that difference most (in roughly the order outcome decisions get made).
1. Find what matters before you choose what to measure
People describe the same experience in many different ways. Our LLM workflows bring those descriptions together across interviews, even where participants never used the expected term, and report how broadly each concept appears. They flag a serious concern raised by only a few participants while keeping it in proportion. They also show when new interviews stop adding new concepts. In one Parkinson's disease project, this kind of listening informed the client's endpoint selection and its conceptual model of symptom impact.
2. Test whether existing measures fit your population
An instrument fits when its concepts match what your participants actually experience. Our workflows compare groups side by side, such as newly diagnosed and longer-term patients, so you can see whether a concept holds across your population or shifts within it. We listen to patients and the care partners who live alongside them. That shows teams which concepts patients describe firsthand (PRO) and where a care partner's view adds to the picture (ObsRO), with each voice kept distinct. In the same Parkinson's project, patients' accounts led to a recommendation that PRO assessments capture emotional as well as physical impact.
3. Speak your participants’ language
Items and instructions work best when participants read them the way you intended. Our linguistic analysis captures the words patients and care partners reach for, the qualifications they add, and the terms that give them pause. That gives teams patient-grounded language to bring into item wording and formal cognitive interviews. In a recent lupus project, adults with lupus reviewed a trial's outcome measures in their own words. Those who struggled with the terminology were more likely to say they would not consider participating.
4. Design for sustainable participation
Outcome data is only as complete as participants’ ability to keep providing it. Our workflows gather patients' reactions to at-home devices, tablet-based tasks, and visit schedules across participants. That shows teams where burden may build while there's still time to adjust. In a Parkinson's disease project for a decentralized trial, most patients and care partners felt the hybrid design struck the right balance of home and clinic visits. inVibe recommended clear references to technical support for the study's digital components. Once a study is running, the same listening can help diagnose what's behind slow recruitment or poor retention.
From Patient Voice to Outcome Data
Think of voice research as the step before measurement and a companion to it. It helps ensure that the right concepts are measured, in language patients understand, and at a level they can sustain. From there, THREAD's Patient-Centered Outcomes Research team develops and validates instruments, and eCOA brings those measures into the trial.
If you're planning endpoints, writing patient-facing materials, or wondering whether your PRO strategy reflects what people living with a condition actually experience, we'd love to help you listen first. Schedule a demo and hear it for yourself.