|
|
|
|
|
|
|
The body is complex and scientists become increasingly focused on one small part without being able to comprehend how it actually functions together. In isolation, a drug may have a particular effect on a single locus of action but when multiple sites are implicated, especially when they intercommunicate, it becomes very difficult to disentangle the separate activities. Physiological and systems analysis may address some of these problems but for these to be useful it is necessary to define a specific model and then to test the assumptions proposed. For more complex systems, which cannot easily be understood, or more importantly to measure the component parts of the system, it is possible to use emerging methods of information analysis called neural networks. These attempt to emulate the brain's neuronal processing capability using artificial intelligence. A neuronal network consists of a network of artificial neurones that receive a certain number of independent input variables. If we use the control of blood pressure as an example, the inputs could be time, plasma drug levels, heart rate, peripheral resistance, diuresis, and cardiac output at rest and after different degrees of exercise. These are processed according to certain characteristics of the net such as the weighting or signal strength of each neurone or transfer function, which acts as a switch to allow the neurone to be activated. It does this according to predescribed input conditions that can be linear or a more useful sigmoidal exponential. After the input layer of neurones, there is a so-called hidden layer, which processes the results and minimizes the variance on incoming data from different nodes before allowing the data to be sent to the dependent-variable output layers. For our particular example, it could be blood pressure together with certain kinetic parameters. |
|
|
|
|
|
|
|
|
As the network is trained, the characteristics of the net change according to the success of the output predictions. This is achieved by changing individual neuronal weights. As these adjustments become reinforced, prior old values are forgotten and the new memory is updated. This teaching process is undertaken by a supervised backward and forward propagation system improving on the estimates of the predictive output each time. Thus, the system is left to learn the dynamics of the net until certain preset conditions are met; e.g., all predictions will be within 15% of the observed value. Such an analysis is less than precise with a certain amount of error being inherent in the method, but with the introduction of fuzzy logic, which uses the imprecision of the neural net output, it can provide an optimized answer based on the ranges of predictions given [77]. Not surprisingly, these methods have found great use in predictive mathematical modeling for forecasting the weather, stock market, and horse racing, or for voice pattern recognition, writing, and images. The few studies that have so far been published do show that neural nets could have a useful application to kinetics [78,79] and dynamics [80,81] (Fig. 10), but they |
|
|
|
|
|