Related Experiment Videos
An empirical comparison of expert-derived and data-derived classification trees
M Chiogna1, D J Spiegelhalter, R C Franklin
1Department of Statistics, University of Glasgow, U.K.
Statistics in Medicine
|January 30, 1996
Summary
Classification trees aid in diagnosing critical congenital heart disease in infants. Expert-derived trees outperformed data-driven methods, especially for rare diseases, improving diagnostic accuracy.
Area of Science:
- Medical Informatics
- Pediatric Cardiology
- Machine Learning
Background:
- Classification trees offer transparent discrimination methods, usable with expert opinion or data analysis.
- Accurate diagnosis of critical congenital heart disease (CCHD) in infants is crucial.
Purpose of the Study:
- To compare expert-derived and data-derived classification trees for diagnosing CCHD in infants.
- To evaluate the impact of incorporating misclassification costs on diagnostic performance.
Main Methods:
- Developed classification trees from expert opinion and analysis of 571 infant cases.
- Compared tree performance on a 27-class CCHD diagnosis problem and a 6-disease subset.
- Incorporated a clinical loss matrix for misclassifications.
- Performed hand-pruning for rare diseases where automatic methods struggled.
Main Results:
- Data-driven tree creation and pruning faced challenges with rare CCHD types.
- Incorporating misclassification costs significantly improved clinical performance.
- The expert-derived tree demonstrated a unique building strategy not replicable automatically.
- Expert trees generally outperformed data-derived trees, particularly in identifying key composite features.
Conclusions:
- Expert-derived classification trees show promise for complex diagnostic tasks like CCHD.
- Integrating clinical costs into tree construction enhances diagnostic utility.
- Hand-pruning may be necessary for rare disease classification using decision trees.