The value of accurate neural networks for future medical decision making

Heidelberg 19 September 1999In the Department of Experimental Medicine in Galway, Ireland, the effectiveness of neural networks for medical diagnosis is being tested. A diagnostic system for jaundice related disease detection has to be as accurate when it comes to medical decision making as the physician in the emergency room. At MedNet'99, Mr. Ben Kanagaratnam explained how neural networks learn to solve new problems in very much the same way as humans use their past experience to solve current issues. If properly prepared and used, neural networks can be of great practical use in the diagnostic practice.


The software for the neural network, as it has been implemented in Galway, is comprised of three different elements, which are the neuroshell or neural network shell programme, the basic language interpreter, and the FOXPRO relational database applications. The built-in diagnostic tool based on a questionnaire, sorts out whether the 92 symptoms which refer to jaundice related disease diagnosis, respond to a positive or a negative indicator. In the case of gallstones, 114 symptoms have been defined, whereas for cancer only 68 are required.

The team has designed four different models in which an increasingly higher degree of efficiency has been achieved. The binary model is able to obtain an accuracy of 67%. By changing the number of hidden nodes, the analogue model can reach a percentage of 79,5% since the network factors have been influenced. As a result, the learning rate as well as the momentum are being optimized. In the third model, the database has been checked to track down missing data, and to discover possibly conflicting information, data which is out of range, or logical errors. This has amounted in an accuracy of 81%.

In model four, only 8 characteristic features have been used which has led to an accuracy of 93%. This has brought Mr. Kanagaratnam to the conclusion that it is of capital importance to collect excellent quality data with regard to format and range of values, as well as to limit the amount of missing data if one wants neural networks to deliver reliable diagnostic outcomes. On top of this, the recognition rate can be increased by checking the configuration file and applying an error control. Provided that all network factors are being refined and that simultaneously more cases get introduced, neural networks are bound to play a major role in the future of the daily hospital practice.

Leslie Versweyveld

[Medical IT News][Calendar][Virtual Medical Worlds Community][News on Advanced IT]