Evaluation of Prediction models
The reliable prediction of outcomes from disease and treatment is becoming increasingly important in the delivery and organisation of health care. The learning objective of this tutorial is to understand the elements underlying predictive performance and to show how to quantitatively assess the performance of prediction models. In particular, I address different categories of performance measures (including calibration and discrimination) and valid methods (including bootstrapping and cross validation) for obtaining performance assessments. I will also provide a step-wise framework for developing, evaluating, and reporting on prediction models.
The focus of the tutorial is on conceptual frameworks. Attention will be paid to the various choices in the design of model evaluation procedures, and the relationship between model evaluation and the purposes for which a model has been built. All methods are illustrated with real-world examples.
This tutorial is meant for medical informatics and computer science researchers, health care workers, and epidemiologists who are developing, evaluating, or using prediction models. The participants should be acquainted with the general concept of a prediction model and with the concept of probability.
Upon following the tutorial the participants should be able to understand and assess the performance of prediction models; know how to report on them; and be able to critique performance assessment reported in the literature.
Ameen Abu-Hanna is full professor at the department of Medical Informatics at the Academic Medical Center at the University of Amsterdam. He is Principal Investigator in the research area Methodology in Medical Informatics with interest in artificial intelligence, statistical machine learning and decision support systems. He was formerly an associate editor of the Journal of Biomedical Informatics and president of the European Society of AI in Medicine. In 2017 Ameen became a founding Fellow of the International Academy of Health Sciences Informatics.
- Prediction models
- Informing patients, triage, and benchmarking
- Model building and evaluation
- Difference between prediction and etiological (causal) models
- Potential outcome framework
- Average and Individual Treatment Effect (ATE and ITE)
- Relationship to Prognostic Factor Research
Predictive performance measures
- Performance aspects
- Performance measures
- AUC and other ROC-based methods
- The problem of AUC with imbalanced data
- The Area under the Precision-Recall Curve
- Brier (Skill) score
- Net Reclassification Improvement
- Calibration graphs and calibration intercept and slope
- Characteristics of performance measures
- Proper scoring rules
- Strictly proper scoring rules
Model selection and validation
- Model selection
- The bias-variance trade-off
- Information criteria in regression
- Resampling methods for model selection and validation
- Internal, temporal structural and external validation
- A framework for understanding external validation
- Comparing different models
- Steps for developing, evaluating and reporting on prediction models