Machine learning with ventilator data
summary
Processes 100Hz dual-channel ventilator data from neonatal ICU traces, about 160 ventilator days collected (each day ≈600–700 MB) producing roughly 43,000 breaths per night per baby. Builds an annotation GUI (Bokeh + ipywidgets) to create a gold standard, engineers features and trains a random forest F1 ≈ 0.9 for breath-start segmentation, exposes failure modes on conscious patients, and open-sources code and results for NHS collaboration as reproducible tooling.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:24Newborns need ventilation
- 3:10Patient-ventilator interactions
- 6:02Ventilator waveform data
- 9:32Project goals
- 11:17Breath segmentation challenges
- 16:12Data processing steps
- 18:37Gold standard for ML
- 20:16Annotation tool workflow
- 23:06Classifier performance
- 27:22Next steps and validation
- 29:11Class imbalance and features
- 30:36Human annotation difficulty
- 33:04Ventilator tech and EMG signals
- 35:14Predicting extubation readiness
- 36:26Funding and open-source
- 37:28Dataset and segmentation precision
- 40:06Defining breath start
related talks
06mBreathTrack: Detecting Regular Breathing Phases from Unannotated Acoustic Data Captured by a ...
34mWhy AI in the Medical Field Both Sucks and Rocks...
39mBridging health inequalities through machine learning
16mUtilizing The Power of Machine Learning in Healthcare
26mMachine Learning for Personalised Health
09mLightning Talk: Bridging the Gap: Engineering Compliant...
06mListen2Cough: Leveraging End-to-End Deep Learning Cough Detection Model to Enhance Lung Health ...
27mThe Journey from AI to Bedside: Translating Academic Research to Serving Patients Using Ray