Responsible implementation of Machine Learning models for life-and-death decision making
summary
Pacmed Critical claims responsible ICU decisions by combining concrete safeguards: business rules and out‑of‑domain detection block predictions for on‑life‑support cases, a simple ML monitoring stack alerts on data drift and missing mappings from hospital ETL, and a Shapley‑value dashboard shows why a patient’s probability moved. The system runs hourly predictions, monitors that a ~6% baseline readmission/mortality holds, requires manual retraining when drift matters, and reports issues to hospitals rather than silently degrading.
AI-written summary · about
chapters (13)
- 0:00Welcome!
- 0:10Introduction
- 1:06Bridging the implementation gap in healthcare
- 2:34Agenda
- 2:55Pacmed Critical and the Intensive Care
- 6:44Challenges of implementing ML for life-and-death decisions
- 7:36Data Validity
- 14:35Interpretability and Auditability
- 19:10Prediction Validity
- 24:24What happens when there is a problem in production?
- 26:54Next steps
- 28:42Conclusions
- 30:06Q&A
related talks
11mDesigning Governed AI Pipelines for Healthcare Impact
25mLifesaving AI and JavaScript
32mSHAP and Beyond
04mBlack Box: How AI will Amplify the Best and Worst of Humanity
24mDeveloping Innovative Systems Through Responsible AI In Healthcare
39mBest Practices for Prototyping ML Models for Healthcare
23mEngineering AI-Powered Predictive Healthcare Systems
20mMachine Learning in Cardiology