Related Experiment Videos
Epidemiologic interpretation of artificial neural networks
M S Duh1, A M Walker, J Z Ayanian
1Department of Epidemiology, Harvard School of Public Health, Boston, MA 02115, USA.
American Journal of Epidemiology
|June 30, 1998
Summary
Neural networks can classify individuals by analyzing input factors. Five hidden neurons in neural networks may optimize risk classification but risk over-fitting, impacting epidemiological studies.
Area of Science:
- Epidemiology
- Machine Learning
- Biostatistics
Background:
- Multilayer neural networks are often criticized as "black boxes" and for their inability to assess input factor importance.
- Understanding how neural networks form decision surfaces and regions is crucial for their application in classification tasks.
Purpose of the Study:
- To illustrate how neural networks can be used to classify individuals.
- To investigate the role of weights in forming neural network decision surfaces and regions.
- To explore the effect of varying numbers of hidden neurons on classification accuracy.
Main Methods:
- Utilized data from a case-control study.
- Employed two significant determinants of case status as input neurons.
- Experimented with zero, three, and five hidden neurons to assess their impact on decision boundaries.
Main Results:
- Three hidden neurons were insufficient for complete discrimination between cases and controls.
- Five hidden neurons showed potential for optimal classification but raised concerns about over-fitting.
- More complex neural networks effectively defined uniform risk regions and assigned risk levels.
Conclusions:
- Neural networks show promise for pattern recognition and complex classification in epidemiology.
- Their utility may be limited in problems with distinct effects of easily distinguishable predictors.