MindMed is using a custom large language model to audit clinician-rated assessments across all three of its phase three LSD trials. Dan Karlin, Chief Medical Officer at MindMed, presented the approach as a case study at the Panagora Clinical Trial Innovation Summit 2025 in Basel. The system is already running in near real time across active studies set to read out in 2026.
The Problem With Clinician-Rated Outcomes
MindMed’s lead program, MM120, is a pharmaceutically optimized version of LSD being studied in generalized anxiety disorder and major depressive disorder. The primary endpoints in both indications are clinician-rated scales — the Hamilton Anxiety Scale (HAM-A) in GAD and the MADRS in MDD.
These scales require trained raters to conduct structured interviews. Despite reasonable inter-rater reliability in published literature, the MindMed team identified several concrete trial risks. Raters introduce subjectivity. There is no built-in measure of uncertainty in any given score. And patients who screen in with inflated baseline scores can end up looking like placebo responders — distorting the treatment effect.
To protect data integrity, MindMed uses central raters rather than site staff. Participants connect via video call with a remote, trained clinician who conducts the interview. This removes site-level variability but does not eliminate rater inconsistency across hundreds of interviews over a year-long study.
Building an Automated Auditing System
The solution came from MindMed’s internal AI team — a two-person group tasked with identifying problems the company could address with emerging tools. They proposed applying a large language model to the rating challenge.
The team started with an open-source LLM and trained it on HAM-A interviews from MindMed’s phase two study, holding back a subset of data for validation. The model demonstrated strong inter-rater reliability against the held-back scores. That established a baseline, though the team acknowledged it meant the model initially reflected the tendencies of the original rater pool.
From there, the system was refined. Each HAM-A and MADRS interview is now scored by five separate instances of the model. Outlier results are discarded and the remaining outputs converge on a consensus score. The process runs in seconds — compared to the roughly one hour a human reviewer would need to audit the same interview.
The output is a question-by-question comparison between the human rater’s score and the model’s score. When the two diverge, reviewers can click into the relevant portion of the interview transcript and assess whether the discrepancy reflects a rating error. The model score cannot be used to change the official rating — that is not how regulated clinical trials work — but it can be used to provide targeted, near-real-time feedback to raters.
The result is a live dashboard showing alignment between model and rater for every patient and every interview across all three studies.
Why Timing Mattered
The MindMed CMO framed the LLM adoption as a case of a tool meeting a problem at precisely the right moment — not the reverse. The company’s phase two program concluded just as capable open-source language models became available. Had the phase three programs launched earlier, the tool would not have existed. Had they launched later, the window to train the model on phase two data before phase three began would have closed.
This timing argument is relevant to broader industry discussions about AI adoption in trials. The more common failure mode is deploying technology in search of a problem. The MindMed case is notable precisely because the problem — subjective central rating at scale across three simultaneous phase three studies — was clearly defined before the solution was selected.
The company also made a deliberate design choice in phase two to front-load development risk. The double-blind phase of all three phase three studies is nearly identical to the phase two design. Phase three adds a 40-week extension in which participants can receive up to four additional doses — something the FDA did not permit in phase two. Higher-than-expected placebo dropout in phase two likely inflated the measured placebo response, which the team believes sets a conservative bar for phase three comparisons.
What Comes Next
Two of the three phase three studies target GAD, including one enrolling across five European countries. The third targets MDD. All three are on track to read out across the first and second half of 2026.
A successful readout would support the first regulatory submission for LSD as a medicine. If the automated auditing system holds up as a quality control tool through full trial execution, it could become a template for other psychiatry programs facing the same central rating challenges. The more immediate question is whether near-real-time rater feedback demonstrably tightens inter-rater reliability over time — a metric MindMed will be tracking as data accumulates through the year.
Moe Alsumidaie, MBA, MSF, is founder and Chief Editor of Vanguard Publications, which publishes Clinical Trial Vanguard, Pharma Vanguard and BullScope, and Head of Research at CliniBiz. He has two decades in clinical trial operations and data science, with earlier roles at Genentech, Abbott Vascular and Stanford University Medical Center, and is a guest lecturer in clinical trial sciences at Rutgers University.
