Related Experiment Videos
A new method to estimate parameters of linear compartmental models using artificial neural networks
S S Gambhir1, C L Keppenne, P K Banerjee
1The Crump Institute for Biological Imaging, Department of Molecular and Medical Pharmacology, UCLA School of Medicine, Los Angeles, California 90095-1770, USA. sgambhir@mednet.ucla.edu
Physics in Medicine and Biology
|July 3, 1998
Summary
Artificial neural networks offer a faster and more accurate method for estimating parameters in compartmental analysis compared to traditional weighted nonlinear regression. This approach provides unbiased estimates and can outperform regression, especially with noisy data.
Area of Science:
- Pharmacokinetics and Pharmacodynamics
- Computational Biology
- Systems Biology
Background:
- Compartmental analysis is crucial for understanding biological systems, often employing weighted nonlinear regression for parameter estimation.
- This traditional method involves fitting exponential functions to kinetic data, with parameters linked to model rate constants.
- Weighted nonlinear regression can be time-consuming and may converge to suboptimal local minima, limiting its efficiency.
Purpose of the Study:
- To investigate the efficacy of artificial neural networks (ANNs) as an alternative to weighted nonlinear regression for parameter estimation in compartmental models.
- To evaluate the speed, accuracy, and robustness of ANNs in estimating model parameters from kinetic data.
- To compare ANN performance against traditional regression methods under various noise conditions.
Main Methods:
- Simple feed-forward artificial neural networks were trained to predict compartmental model parameters.
- Kinetic data from compartmental models served as input to the neural networks.
- The performance of ANNs was compared to weighted nonlinear regression for mono- and biexponential models.
Main Results:
- Artificial neural networks demonstrated significantly faster computation times compared to regression algorithms.
- ANNs produced unbiased parameter estimates, comparable to or better than weighted nonlinear regression.
- At typical noise levels, ANNs yielded lower variance estimates than weighted nonlinear regression, particularly when regression failed to converge.
- The improved performance of ANNs was attributed to their ability to avoid local minima issues inherent in regression.
Conclusions:
- Artificial neural networks are powerful and efficient tools for parameter estimation in compartmental analysis.
- ANNs offer a viable, faster, and potentially more accurate alternative to weighted nonlinear regression, especially for complex or noisy datasets.
- This approach holds promise for accelerating research in fields relying on compartmental modeling, such as pharmacokinetics.