NEJM AI Grand Rounds: clinician-built previsit tool gathered history before 150,000 visits

On a recent episode of NEJM AI Grand Rounds, Dr. Michal Nedoszytko, an interventional cardiologist now working as a clinician scientist at Abridge, described Previsit, a tool he built while running outpatient clinics in Belgium. Patients received a text message before the visit, answered pertinent questions from a large language model, and the physician received a summary. He said the system was integrated in six countries and completed over 150,000 previsits.

The problem he set out to solve will be familiar to any clinic. He said he would bet that more than half of patients cannot name their medications, describing them instead as "the drug for the heart" or "the red pill," while the accurate list sits with a spouse at home. Previsit was meant to carry that information into the electronic health record before the patient arrived. He was explicit that the goal was not to replace the clinical history but to improve it: patients came prepared, knew what would be asked, and the conversation had "a completely different quality." Time saved was a benefit, not the whole point.

For compliance and security leaders, the design choices matter as much as the outcome. Nedoszytko said Previsit was built as an administrative tool, with the ability to identify the patient separated from the collected data, no stored IDs, and integration with the EHR. When a Belgian hospital announced it as a first, a GDPR inspector arrived three days later; he said the full review found no issue. He contrasted this with a post-visit tool that suggests decisions to patients, which he treated as a different and higher-risk category.

The evidence has limits. The 150,000 figure, the six-country footprint and the estimate about medication recall were stated in conversation on a podcast. No published study, outcome measure or error-rate comparison was cited, so the effect on reconciliation accuracy, visit length or patient experience is not established here.

The move is a contained test. Ask your Chief Medical Officer and clinic operations lead whether patient-reported history collected before the visit could improve medication reconciliation in one clinic. Define the measure first, such as discrepancies caught before the encounter, and have compliance review how identity and clinical content are separated. The return to watch for is a better first ten minutes for patients and less chart reconstruction for the people who see them.

You are an operations advisor to the Chief Medical Officer and executive team of [your organization].

Context: On the NEJM AI Grand Rounds episode "Abridge's Dr. Michal Nedoszytko on Physicians Who Build Software," Dr. Michal Nedoszytko described Previsit, a tool he built after seeing that more than half of patients cannot name their medications. Patients texted with a large language model before the visit, and the physician received a summary. He said it ran in six countries and completed over 150,000 previsits. He said the goal was better visit quality, not only time saved, and that patient identity was separated from the collected data. These figures were stated on a podcast. No published study was cited.

Task: Draft the questions I should bring to my clinic operations lead and compliance lead to decide whether patient-reported history collected before the visit could improve medication reconciliation in one clinic.

Output: a one-page question set for the team under these headings: Clinical fit, Measure defined first (such as discrepancies caught before the encounter), Workflow and owners, Compliance and data separation, Patient experience, and Decision. Give three to five questions per heading, then a short list of what must be verified before any pilot.

Fill these in: [the clinic chosen for the test], [your EHR], [your current medication reconciliation process and who owns it], [approximate weekly visit volume at that clinic], [how patients are contacted before visits today].

Guardrails: Use only the facts above and the details I fill in. Cite NEJM AI Grand Rounds by name when you refer to its claims. Flag anything uncertain or needing verification, especially the podcast figures. Do not request or include any patient data. Nothing here is medical advice. Label the result as a draft for the team to check.