Dowell J A, Hussain A, Devane J, Young D
In Vitro-In Vivo Relationship Cooperative Working Group, Pharmacokinetics-Biopharmaceutics Laboratory, Department of Pharmaceutical Sciences, University of Maryland at Baltimore, Baltimore, Maryland 21201, USA.
J Pharm Sci. 1999 Jan;88(1):154-60. doi: 10.1021/js970148p.
Artificial neural networks applied to in vitro-in vivo correlations (ANN-IVIVC) have the potential to be a reliable predictive tool that overcomes some of the difficulties associated with classical regression methods, principally, that of providing an a priori specification of the regression equation structure. A number of unique ANN configurations are presented, that have been evaluated for their ability to determine an IVIVC from different formulations of the same product. Configuration variables included a combination of architectural structures, learning algorithms, and input-output association structures. The initial training set consisted of two formulations and included the dissolution from each of the six cells in the dissolution bath as inputs, with associated outputs consisting of 1512 pharmacokinetic time points from nine patients enrolled in a crossover study. A third formulation IVIVC data set was used for predictive validation. Using these data, a total of 29 ANN configurations were evaluated. The ANN structures included the traditional feed forward, recurrent, jump connections, and general regression neural networks, with input-output association types consisting of the direct mapping of the dissolution profiles to the pharmacokinetic observations, mapping the individual dissolution points to the individual observations, and using a "memorative" input-output association. The ANNs were evaluated on the basis of their predictive performance, which was excellent for some of these ANN models. This work provides a basic foundation for ANN-IVIVC modeling and is the basis for continued modeling with other desirable inputs, such as formulation variables and subject demographics.
应用于体外-体内相关性（ANN-IVIVC）的人工神经网络有潜力成为一种可靠的预测工具，它能克服一些与经典回归方法相关的困难，主要是为回归方程结构提供先验规范这一困难。本文介绍了多种独特的人工神经网络配置，并对它们从同一产品的不同剂型确定体外-体内相关性的能力进行了评估。配置变量包括架构结构、学习算法和输入-输出关联结构的组合。初始训练集由两种剂型组成，包括溶出度试验中六个溶出杯各自的溶出度作为输入，相关输出由参与交叉研究的九名患者的1512个药代动力学时间点组成。第三个剂型的体外-体内相关性数据集用于预测验证。利用这些数据，共评估了29种人工神经网络配置。人工神经网络结构包括传统的前馈、递归、跳跃连接和广义回归神经网络，输入-输出关联类型包括将溶出曲线直接映射到药代动力学观测值、将各个溶出点映射到各个观测值以及使用“记忆性”输入-输出关联。基于其预测性能对人工神经网络进行了评估，其中一些人工神经网络模型的预测性能非常出色。这项工作为体外-体内相关性人工神经网络建模提供了基础，也是使用其他理想输入（如剂型变量和受试者人口统计学数据）继续建模的基础。
