Transferring clinical prediction models across hospitals...
summary
Built an ICU readmission predictor called Padma critical on 14,000 Amsterdam admissions with AUC 0.8 then sees performance fall to 0.76 on later local data and 0.69 on a different hospital. Applies concrete domain‑adaptation methods — importance weighting, sequential transfer (augmenting features with target model probabilities), and domain‑specific coefficient models — and recovers much of the lost AUC, with full data sharing giving the largest gain and transfer methods proving most helpful when the target set is small.
AI-written summary · about
chapters (7)
- 0:00External validation in clinical models
- 3:19Padma Critical ICU use case
- 7:38Performance drop in validation
- 9:23Covariate and concept shift
- 12:34Domain adaptation techniques
- 17:50Domain adaptation results on Padma Critical
- 21:42Key takeaways and future
related talks
32mSHAP and Beyond
54mCrossing the Last Mile with Predictable Hospital Readmissions
37mThe Data Addition Dilemma
53mAre We Measuring the Right Thing? Distribution Shift Lessons for Federated Learning
22mAI or Not AI? That's the Answer!
26mPrediction Model for Favorable COVID-19 Patient Outcomes
35mML: Beyond confusion matrix, model performance analysis
35mResponsible implementation of Machine Learning models for life-and-death decision making