Practical AI in Healthcare S1, E28 - Adam Blum: AI-Powered Clinical Trial Matching
Mar 15, 2026
Adam Blum, a serial AI entrepreneur who built CancerBot after his follicular lymphoma diagnosis, describes building precision clinical-trial matching tools. He explains the Prompt Workbench for high-accuracy extraction. He unpacks transforming messy eligibility into tractable logic using conjunctive normal form. He discusses patient flows, accuracy results, and open-source collaborations.
AI Snips
Chapters
Transcript
Episode notes
Patient Built CancerBot After Failed Trial Searches
- Adam Blum spent three weeks manually reading global trial registries after his follicular lymphoma diagnosis because commercial matchers returned no true matches.
- He found consultants and matchers used only a few questions (age, gender, stage), so none captured complex eligibility, prompting him to build CancerBot.
Eligibility Text Is Unstructured And Ambiguous
- Trial eligibility is written as long, free-form text with inconsistent vocabulary and complex logical expressions that confuse patients and clinicians.
- Examples include varied drug naming (R-benda vs rituximab bendamustine) and nested prior-therapy logic like required PI and IMID combinations.
Find Eligible Trials Then Rank By Patient Priorities
- First perform precision eligibility matching, then rank eligible trials by what matters to the patient (risk, benefit, burden).
- Separating eligibility from patient-specific goodness avoids recommending trials the patient can't join and enables personalized ordering.