Synthetic EMR
Evidence & Safeguards
Inspect a Korean or English record, trace each answer to its source, and see what happens when information is missing.
Synthetic data onlyPatient-level splitsRuns in your browserResearch background · Hika Barki
Biomedical signal processing, wearable sensing, transformers, and medical-image analysis. The publication portfolio includes work on multimodal PPG/EEG affective-state classification, transformer-based stress classification, and gastrointestinal and mammographic imaging.
Selected papers and supplied bibliography · Google Scholar · Organization work
The prototype below is separate from those studies and makes no clinical or LLM-validation claim.
Identifier-redacted record
Extractive summary
Only documented fields, with exact source lines.
Answer or refusal
What this demonstrates
24 invented patients, two visits each. Both visits stay in the same partition: 16 train / 4 validation / 4 test patients. This interface shows only the eight test records.
Each record deliberately omits one supported field. Choose the suggested missing-field question to trigger refusal, or select the treatment question to see an unsupported-request refusal.
Show safeguard audit
Limits and reproducibility
The redactor handles explicit identifier formats in the bundled templates. It is not a comprehensive real-EMR de-identification system. Pseudonymous grouping keys are invented. Shared templates across splits mean these fixtures do not measure generalization.
The browser loads only bundled synthetic data. There are no model calls, form submissions, analytics scripts, or file uploads. The hosting service may process access logs.
See the repository for the Python generator, nine safeguard tests, browser contract tests, and evaluation report. No clinical accuracy or LLM improvement is claimed.