Related Experiment Videos
Statistical approach to neural network model building for gentamicin peak predictions
Journal of Pharmaceutical Sciences
|January 1, 1996
Summary
Feed forward neural networks can effectively model complex data, similar to statistical models. Careful training is essential to prevent overfitting and achieve optimal predictive performance, with stopped training yielding the best results.
Area of Science:
- Computational biology
- Machine learning
- Pharmacometrics
Background:
- Feed forward neural networks (FFNNs) are powerful nonlinear modeling tools extending traditional statistical methods.
- The study investigated building FFNN models analogous to statistical model construction.
Purpose of the Study:
- To test the hypothesis that FFNN models can be constructed similarly to statistical models.
- To predict gentamicin serum concentrations using FFNNs.
Main Methods:
- FFNN models were developed using forward and backward variable selection with varying hidden nodes (0-5) and transfer functions (tanh, linear).
- Model selection employed Akaike information criteria (AIC), Bayesian information criteria (BIC), and stopped training.
- Peak gentamicin concentrations from 392 patients were used for training, testing, and validation.
Main Results:
- Models with all 10 inputs and five hidden nodes demonstrated the lowest root mean square (rms) error.
- Stopped training achieved the lowest average rms error (1.46) on the validation set, followed by AIC (1.51) and BIC (1.56).
- Overfitting was observed with excessive hidden nodes (rms=1.49) or uninformative inputs (rms=1.70).
Conclusions:
- Neural networks can be constructed with numerous parameters for robust predictive performance.
- Careful training is crucial to avoid overfitting, with stopped training emerging as the optimal method for minimizing rms error.