Without scientifically sound frameworks, patient-voice AI is just louder guesswork. Why Semalytix co-developed the POMELO protocol, and how Atlas is engineered to it by design.
Regulators like the FDA and EMA have, for some years now, recognised social-media listening as a valid source of patient experience data. What the field has lacked is not interest, and not data. It has lacked a shared methodology. Without a rigorous methodology that allows for reproducibility of studies, results from social media listening can not be regarded as proper evidence. That is the gap POMELO was designed to close. POMELO, the Protocol for sOcial MEdia Listening Online, is the protocol defined by Pistoia Alliance's "Social Media and Real World Evidence" expert group, co-founded by Semalytix. With Boehringer Ingelheim, Chiesi, Roche and Takeda engaged in this pre-competitive activity, the goal has been to provide to industry a defensible, shared bar for what a rigorous study looks like.
POMELO defines four pillars that every credible patient-voice study should be able to evidence: study setup, data collection, algorithm selection, and analysis. Study setup involves defining a well-defined research question that can be answered over social media and defining inclusion and exclusion criteria for the population to be investigated. Data collection requires finding appropriate fora, and opting for suitable ways of collecting the data from a technical perspective, and respecting terms and conditions of data providers. Data collection requires traceable provenance for every signal that enters the corpus. Algorithm selection involves selecting a suitable algorithm to extract key information and understanding the quality and accuracy of the selected algorithm. Data analysis is about choosing a suitable method for analysing the data, either qualitative, quantitative or mixed method. None of these pillars is per se surprising. Taken together, they are simply the bar one would expect of any evidence destined to inform a pharma decision. The reason POMELO is significant is not that the bar is new. It is that the bar is now shared. Without scientifically sound frameworks, patient-voice AI is just louder guesswork.
Atlas was not retrofitted to POMELO. Because the same team helped to define the protocol, Atlas was built against its four pillars from the start. Cohorts are defined and version-controlled. Every retrieved post carries provenance back to its source. The supervised models that authenticate the patient voice are documented, reviewed and evaluated. The path from a raw forum post to an aggregate finding in a brand team's report is reproducible end-to-end. What this means in practice is that an Atlas study is not a black box deliverable. It is a methodology which a regulator, a compliance team or an internal medical reviewer can interrogate and replicate. That is the difference between a finding that informs an opinion and a finding that survives a stage-gate review.