Techniques for optimizing neural networks for medical applications

The overwhelming amount and diversity of clinical data that may be collected from patients have necessitated the development and use of sophisticated processing methods in order to make the data useful in medical applications. Artificial neural networks (ANNs) are one such method that has found growing success in the medical field. In the past, techniques for optimizing neural networks have focused almost completely on ANN training algorithms and architectures in order to improve performance on specific applications. However, a more data-centric view of optimization may prove more valuable in making ANNs useful to a wider community of medical researchers. Four specific characteristics of medical data processing are addressed: (1) large numbers of classes into which we wish to categorize data; (2) nonuniform distribution of classes; (3) features from input data with non-Gaussian distributions and widely variable ranges; (4) large, complex data sets from which it is difficult to choose appropriate data for model development. These characteristics present potential problems for ANN development. It is proposed that these problems may be overcome by using new data normalization techniques, implementing hierarchical networks, and using distance metrics to choose the most useful training patterns. These proposals should be generalizable to many medical applications because they focus on solutions to data problems and not specific applications. The specific proposals are applied to two practical medical applications: blood pressure determination and classification of breathing circuit faults. The proposals are compared with standard ANN training techniques. For each of the proposals, there was a significant increase in training performance, or network classification capability, or both. It is concluded that the proposals should be generalizable to other medical applications and that a data-centric view of optimization is valuable to ANN development in medical applications.