Anthropic’s CEO, Dario Amodei, recently claimed that AI could cure most human disease within five to ten years. Set aside whether that is right. The harder question underneath it is the one that decides how fast clinical research can actually change: what happens to the humans in the study, the patients and the physicians, when the machine gets this good?
That is the question Derk Arts, Founder and CEO of Castor, and Matthew Gordon, Vice President of Real-World Evidence at Aptitude Health, took to a live, no-slides debate. Derk argued the deliberately provocative side, that real-world evidence may no longer need the physician at all. Matthew argued that the physician is still where the real value is. Neither fully believed his own corner, and that was the point.
What the debate actually surfaced
Derk opened with the strong version of the case. Real-world evidence is about standard of care, patients increasingly can access and direct their own data, and modern tools can pull it directly. So why keep the physician in the loop at all? Matthew’s answer became the spine of the hour.
Matthew Gordon, Vice President of Real-World Evidence, Aptitude Health
His example was in oncology. The record may show what treatments a patient received from first-line to second-line therapy, or that a genetic test was ordered. It rarely shows why, or why a physician chose to deviate from guidelines for a particular patient. Derk pushed back: shouldn’t the why live in the notes? Often it does not, which is exactly Matthew’s point, and an audience point from Jen Wheeler sharpened it. If the reasoning is not in the record, it exists nowhere in written form, stranded on a single physician’s memory, a knowledge island you then have to reach. With ten-minute visits and AI scribes now standing between the clinician and their own recall, retrieving it after the fact is close to impossible at scale.
Where they converged is the practical takeaway. Technology does not remove the physician, it changes the job. Pulling data straight from the medical record, through Castor Catalyst, Castor’s AI extraction pipeline with auditable, human-review steps, takes the data entry burden off the clinician. As Derk described it, that frees the physician to contribute the context and adjudication only they can, and it lowers the monitoring and budget load enough to bring more community physicians into studies rather than fewer. Matthew’s own model puts the clinician at the front door.
Matthew Gordon, Vice President of Real-World Evidence, Aptitude Health
The resolution both landed on: there is no one-size-fits-all study. Start from the strategic objective. Some questions are answered fastest by pulling the data direct to patient. Others need the clinician in the room. Most need both.
What is in the full conversation
A few exchanges are worth watching in full. Derk makes the case that a new category of study is arriving, physician-light and patient-led, built for the GLP-1 generation of people who do not see themselves as sick but want to be healthier, where patient-experience and wearable data carry the weight. The two then trade ten-year predictions, and they do not match: Derk expects genuinely physician-less trials pulling wearable, patient-reported, and multi-institution record data, while Matthew holds that no robot replaces the clinician because our prehistoric brains still want the human in the room. There is also a candid stretch on why the care setting will adopt this faster than clinical trials will, and Matthew’s unscripted account of what a single unified data platform changed for the community physicians he works with.
Watch the full debate on demand. No slides, two people who build this work arguing the future of real-world evidence. Best for real-world evidence, clinical operations, and HEOR leaders weighing how far to push direct-to-patient and AI-driven data capture.
FAQ
What is direct-to-patient real-world evidence?
Direct-to-patient real-world evidence collects data with the patient at the center rather than routing everything through a site. Patients consent electronically, retrieve their own medical records (in the US via TEFCA and FHIR or a HIPAA release), and contribute patient-reported data, with wearable device data connected via API. It suits questions about lived experience, adherence, and outcomes, and it is well matched to people who want to optimize their health rather than treat an illness.
How does AI extract data from medical records without losing clinical context?
The safeguard is human review, not full automation. Derk described Castor Catalyst as an AI-driven extraction pipeline with auditable steps and human review built in, so human reviewers stay in control, with risk-based review before data enters the record. Both speakers agreed that letting an AI infer missing clinical reasoning on its own is dangerous, because it invites hallucination. The role of AI is to surface and structure what is in the record, with a human keeping control.
How do you keep AI-assisted real-world data trustworthy?
By keeping the provenance clear. Both speakers agreed the evidence stays trustworthy only when it is obvious what came from the patient, what came from the clinician, and what was inferred by the model. Castor Catalyst attaches a confidence score and source evidence to every extracted value, and human reviewers stay in control, with risk-based review before data enters the record. The model is not allowed to fill missing clinical reasoning on its own, which is where trust would break down.
