Samsung Medical Center · KOHI Medical-AI Workforce Training, Hands-on 2: From Data to Manuscript
The second hands-on session of the module. Using a synthetic health-checkup dataset, participants built the cleaning record, Table 1, a regression, and ROC figures first, then a manuscript and slides from that evidence, over four hours.
Overview
- Date: Monday, 24 August 2026, 18:30–22:30 (4 hours)
- Venue: Hippocrates Hall, 9th floor, Building B, Samsung Medical Center Ilwon Campus
- Host: Ministry of Health and Welfare · KOHI, "2026 Medical-AI Workforce Training for Health Professionals," Samsung Medical Center track
- Course: Organizing research material and knowledge management with AI (hands-on module 1, part 2): the manuscript pipeline
- Format: Hands-on on participants' own laptops, with a synthetic dataset provided by the instructor or participants' own data
The order of work
Tables and figures came first, and the text was only allowed to say what they showed. I compared it to delegating to a staff member: you set the direction and you check what comes back. The literature library and Introduction from part 1 went straight onto the front of the manuscript.
The practice dataset was synthetic: predicting diabetes from health-checkup data, 320 people and 18 variables (blood pressure, fasting glucose, HbA1c, lipids, smoking, alcohol, exercise), with diabetes status as the outcome. I said up front that it must not be cited as real findings.
Six steps
- Checking the data. Without telling participants what to look for, we ran
/clean-data. Four problems were planted: a fasting glucose of 999, smoking coded as both "NEVER" and "never," a weight of 6.5 kg for a height of 171 cm, and blank cells. People decided how to handle each one, and the AI kept the record. - Table 1. Built with
/analyze-stats, without letting the AI pick the tests on its own. The second checkpoint was picking two numbers from the table and checking them against the raw data. In the multivariable regression we read together how the confidence interval for BMI crossed 1. - Figures.
/make-figuresdrew ROC curves for fasting glucose, HbA1c, and the combined model on one plot. We also covered calibration plots, the participant flow diagram, and common figure mistakes. - Writing.
/write-paperdrafted Methods from the cleaning record and Results using only the numbers in the table, and participants checked each number in the text against the table. We then went through a four-paragraph Discussion, the limitations paragraph, and a/humanizepass that leaves numbers untouched. - Citations and journals.
/verify-refschecked the citations and/check-reportingcompared the draft with the reporting guideline. Zotero inserted the references, and/find-journalsuggested three journals with reasons. - Slides and abstract.
/present-paperturned the same evidence into a 15-minute journal-club deck with speaker notes, and the slide numbers were checked against the table. The abstract came last.
Something that happened while preparing
While preparing the material I verified 13 DOIs. All 13 resolved, but 5 pointed to a different paper. I used this to show why checking that a DOI opens is not enough.
Closing
The last slides were about the difference between using AI well and leaning on it, which comes down to whether you can verify what it produces. For homework I suggested running the same sequence on their own data and growing their library. Thanks to everyone who came to both sessions, and to the organizers.