Veng-Pedersen P, Modi N B
College of Pharmacy, University of Iowa, Iowa City 52242.
J Pharmacokinet Biopharm. 1992 Aug;20(4):397-412; discussion 413-8. doi: 10.1007/BF01062465.
Neural networks (NN) are computational systems implemented in software or hardware that attempt to simulate the neurological processing abilities of biological systems, in particular the brain. Computational NN are classified as parallel distributed processing systems that for many tasks are recognized to have superior processing capability to the classical sequential Von Neuman computer model. NN are recognized mainly in terms of their adaptive learning and self-organization features and their nonlinear processing capability and are considered most suitable to deal with complex multivariate systems that are poorly understood and difficult to model by classical inductive, logically structured modeling techniques. A NN is applied to demonstrate one of the potentially many applications of NN for modeling complex kinetic systems. The NN was used to predict the effect of alfentanil on the heart rate resulting from a complex infusion scheme applied to six rabbits. Drug input-drug effect data resulting from a repeated, triple infusion rate scheme lasting from 30 to 180 min was used to train the NN to recognize and emulate the input-effect behavior of the system. With the NN memory fixed from the 30- to 180-min learning phase the NN was then tested for its ability to predict the effect resulting from a multiple infusion rate scheme applied in the subsequent 180 to 300 min of the experiment. The NN's ability to emulate the system (30-180 min) was excellent and its predictive extrapolation capability (180-300 min) was very good (mean relative prediction accuracy of 78%). The NN was best in predicting the higher intensity effect and was able to identify and predict an overshoot phenomenon likely caused by a withdrawal effect from acute tolerance. Current modeling philosophy and practice is discussed on the basis of the alternative offered by NN in the modeling of complex kinetic systems. In modeling such systems it is questioned whether traditional modeling practice that insists on structure relevance and conceptually pleasing structures has any practical advantages over the empirical NN approach that largely ignores structure relevance but concentrates on the emulation of the behavior of the kinetic system. The traditional searching for appropriate models of complex kinetic systems is a painstakingly slow process. In contrast, the search for empirical models using NN will continue to improve, limited only by technological advances supporting the very promising NN developments.
神经网络（NN）是通过软件或硬件实现的计算系统，旨在模拟生物系统，尤其是大脑的神经处理能力。计算神经网络被归类为并行分布式处理系统，在许多任务中，它们被认为比经典的顺序冯·诺依曼计算机模型具有更强的处理能力。神经网络主要因其自适应学习和自组织特性以及非线性处理能力而被认可，并且被认为最适合处理复杂的多变量系统，这些系统难以理解且难以用经典的归纳、逻辑结构化建模技术进行建模。应用神经网络来展示其在建模复杂动力学系统方面众多潜在应用中的一个。该神经网络用于预测阿芬太尼对六只兔子采用复杂输注方案后的心率影响。来自持续30至180分钟的重复、三重输注速率方案的药物输入 - 药物效应数据用于训练神经网络，以识别和模拟系统的输入 - 效应行为。在30至180分钟学习阶段固定神经网络的记忆后，然后测试神经网络预测在实验后续180至300分钟应用的多重输注速率方案所产生效应的能力。神经网络模拟系统（30 - 180分钟）的能力非常出色，其预测外推能力（180 - 300分钟）也非常好（平均相对预测准确率为78%）。神经网络在预测高强度效应方面表现最佳，并且能够识别和预测可能由急性耐受性戒断效应引起的过冲现象。基于神经网络在复杂动力学系统建模中提供的替代方法，讨论了当前的建模理念和实践。在对这类系统进行建模时有人质疑，坚持结构相关性和概念上令人满意的结构的传统建模实践，与很大程度上忽略结构相关性但专注于模拟动力学系统行为的经验性神经网络方法相比，是否具有任何实际优势。传统上寻找复杂动力学系统的合适模型是一个极其缓慢的过程。相比之下，使用神经网络寻找经验模型将不断改进，仅受支持非常有前景的神经网络发展的技术进步的限制。
