Related Experiment Videos
Genetic programming outperformed multivariable logistic regression in diagnosing pulmonary embolism.
Cornelis J Biesheuvel1, Ivar Siccama, Diederick E Grobbee
1Julius Center for Health Sciences and Primary Care, University Medical Center, P.O. Box 85500, GA Utrecht 3508, The Netherlands. cbiesheu@umcutrecht.nl
Journal of Clinical Epidemiology
|July 13, 2004
Summary
Genetic programming shows promise for medical prediction, outperforming logistic regression in a pulmonary embolism diagnosis study. This technique can develop effective diagnostic and prognostic prediction rules.
Area of Science:
- Computational intelligence
- Machine learning in medicine
- Predictive modeling
Background:
- Genetic programming (GP) is a search methodology adept at uncovering complex relationships among numerous variables.
- While GP has applications in areas like myoelectrical signal recognition, its utility in medical prediction for diagnostic and prognostic purposes remains underexplored.
Purpose of the Study:
- To compare the efficacy of genetic programming against logistic regression for developing medical prediction models.
- To evaluate the diagnostic predictive performance of GP in a real-world medical scenario.
Main Methods:
- A diagnostic prediction model for pulmonary embolism was developed using both genetic programming and logistic regression.
- Models were trained on 67% of empirical data and internally validated using bootstrapping.
- Predictive performance was assessed on the remaining 33% of the data (validation set).
Main Results:
- The genetic programming model demonstrated a significantly higher area under the receiver operating characteristic (ROC) curve (0.73) compared to the logistic regression model (0.68) in the validation set.
- Both models exhibited similar calibration, indicating comparable levels of overoptimism.
Conclusions:
- Despite challenges in interpretability, genetic programming emerges as a potent technique for creating medical prediction rules.
- This study provides the first empirical evidence quantifying the value of genetic programming in medical prediction, suggesting its potential for diagnostic and prognostic applications.