Download presentation
Presentation is loading. Please wait.
Published byEdgar Shepherd Modified over 6 years ago
1
VALIDATION AND UPDATING OF MODELS WITH BIOMARKERS
Ewout W. Steyerberg, PhD Center for Medical Decision Making Dept of Public Health Erasmus MC, Rotterdam, the Netherlands Utrecht, March 25, 2009
2
Erasmus MC – University Medical Center Rotterdam
3
Research
4
Contents Praise for Prof. Paul Ridker
Prediction models for better decision making Validation and updating Extension with biomarker(s)
5
Search Google Scholar …
6
Prediction models for better decision making
Identify low and high patients of cardiovascular disease better targeting of preventive interventions Predictions are probabilities No certainty Validation commonly includes calibration and discrimination Systematically too high / too low predictions Poorer performance than hoped for Better predictions with stronger predictors Much interest in biomarkers
7
Case study 1: validity of Framingham risk models
Updating of regression coefficients: refitting Updating to average outcome: re-calibration
8
Validity of Framingham predictions (JAMA 2001)
9
Tab 1
10
Tab 3: Refit
11
Tab 5: Performance Improvement in c e.g 0.70; native Americans even larger gains ‘substantial improvement by using locally updated coefficients’ Recalibration important for better calibration
12
Case study 2: updating with a biomarker (Circulation 2008)
Refitting of Framingham model Extension with CRP Many statistics to quantify improvement
17
Case study 3: updating with a set of biomarkers
Refitting of Framingham model Extension with 4 biomarkers
18
Another example: NEJM 2008
19
Results: focus on c statistic
20
Elderly: mean 71 yrs
21
All markers add significantly, but NT-Pro-BNP is the winner
22
Substantial improvement in c statistics
23
.. and substantial NRI
24
Conclusions Model validation is followed by model updating and model extension Recalibration is a minimum Often new coefficients required Biomarkers need to be strong and beyond discovery phase Incremental contribution to traditional risk factors Be sceptical about new genetic signatures Quantification of model improvement challenging Patterns in C stat and net reclassification index (NRI) coincide No improvement in c no improvement in NRI Improvement in decisions should be quantified by decision-analytic measures, weighing costs of wrong decisions, e.g. ‘Net Benefit’
25
Read more .. PubMed Paper with attendents of Kattan symposium “Accuracy of prediction models” Books
26
26
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.