Research,
week by week.
This is where I document what I tried, what I found, and what needs another look. New entries are added as the project develops; planned work stays clearly labelled.
Project: How well does an open ECG model travel?
The record so far
Published notes first, then the current work and planned follow-ups.
Getting the model running
Connected the data loader to the model to ensure traces could pass through successfully. A technical connection check, not a performance evaluation.
Read the original Week 1 announcement ↗Checking a published result
Ran the authors' 12-lead model on 21,799 recordings. 29 of 30 individual finding scores reproduced within 0.001. The original headline tolerance was not met; the reported average remains unexplained.
Read the corrected Week 2 report →Preparing the surgical-data comparison
The planned VitalDB Arrhythmia evaluation uses selected surgical ECG segments reviewed by anaesthesiologists. Preparing the loader and exclusions; no new performance result yet.
Make the comparisons stronger
Keep each patient’s recordings together when estimating uncertainty. Run a matched single-lead PTB-XL baseline before interpreting VitalDB.
From datasets toward devices
Chest-strap and three-pad hardware experiments. Producing a model output from a new device would not, by itself, establish accuracy.
The project overview explains the question, background and evaluation approach.
Back to the project overview →