BigMedilytics: MPyC in practice
summary
Demonstrates using secure multi-party computation (MPyC) to train and evaluate clinical prediction models without sharing raw patient records, applied to a heart-failure pilot with Zilveren Kruis and Erasmus MC to predict hospitalization days. Combines the MPyC toolbox, private set intersection protocols, proximal methods, lasso/linear programming implemented with NumPy notebooks, and shows a distributed demo across 3,000 laptops that produces usable model coefficients, noting practical feasibility but communication and protocol-tuning overhead.
AI-written summary · about
chapters (8)
- 0:00BigMedilytics MPyC intro
- 3:13Heart failure pilot Zilveren Kruis Erasmus MC
- 6:27MPyC in practice
- 9:44Empatie toolbox for MPyC
- 13:42Feature selection MPyC training
- 16:40Proximal functions optimization
- 18:50Live demo intersection
- 20:36Scaling MPyC future
related talks
35mResponsible implementation of Machine Learning models for life-and-death decision making
20mIntroduction of MedModels
26mDoes using data mean giving up privacy?
32mSHAP and Beyond
16mPrivacy Preserving AI
39mBest Practices for Prototyping ML Models for Healthcare
21mUsing R to develop production modeling workflows at Mayo Clinic
27mDr Prakash Data Science Manager, ZS Associates at Cypher 2017