Resources

Swiss-Mammo: A physician-written, synthetic dataset of 28 German mammography reports. The dataset is stratified based on BI-RADS categories and available in German and English.

A collection of data from P300-based brain-computer interface studies.

The subset of the MIMIC-IV dataset includes the examination results and diagnostic information of 4,761 cardiac disease patients. The examination results for each patient are listed separately as evidence for the final diagnosis.

FDTooth is a dataset containing intraoral photographs and cone-beam computed tomography (CBCT) images with annotations for automated detection of fenestration and dehiscence in anterior teeth.

MeDiSumQA is a dataset of patient-oriented QA pairs from MIMIC-IV discharge summaries, designed to evaluate LLMs in generating safe, patient-friendly medical responses for clinical QA and healthcare communication.

We created 23 questions resembling eligibility criteria from the apixaban clinical trial and evaluated them on a random sample of 100 patient notes from MIMIC-IV. We release the 2300 total question-answer pairs as a dataset here.

Annotations for unsupported facts in 100 original MIMIC patient summaries (discharge instructions) and hallucinations in 100 Large Language Model (LLM) generated patient summaries labeled by two medical experts.

In our recent study, we used Llama-3.1-70B-Instruct to generate synthetic training examples resembling clinical trial eligibility criteria. We manually reviewed 1000 of these examples and release them here.

Holter ECG database from Japan, containing data from 100 unique patients with paroxysmal AF including expert annotations of Supraventricular arrhythmias at the beat level.

A benchmark dataset offering hourly records over a 90-day period for 50,920 ICU subjects, including dynamic pulmonary function data and a spectrum of covariates for respiratory intervention analyses.