Designing an expert system for differential diagnosis of β-Thalassemia minor and Iron-Deficiency anemia using neural network

Artificial neural networks are a type of systems that use very complex technologies and non-algorithmic solutions for problem solving. These characteristics make them suitable for various medical applications. This study set out to investigate the application of artificial neural networks for differential diagnosis of thalassemia minor and iron-deficiency anemia.
It is a developmental study with a cross-sectional-descriptive design. The statistical population included CBC results of 395 individuals visiting for premarital tests from 21 March to 21 June, 2013. For development of the neural network, MATLAB 2011 was used. Different training algorithms were compared after error propagation in the neural network. Finally, the best network structure (concerning diagnostic sensitivity, specificity, and accuracy) was selected, using the confusion matrix and the receiver operating characteristic (ROC).
The proposed system was based on a multi-layer perceptron algorithm with 4 inputs, 100 neurons, and 1 hidden layer. It was used as the most powerful differential diagnosis instrument with specificity, sensitivity and accuracy of 92%, 94%, and 93.9%, respectively.
The artificial neural networks have powerful structures for categorizing data and learning the patterns. Among different training methods, the Levenberg-Marquardt backpropagation algorithm produced the best results due to faster convergence in network training. It also showed considerable accuracy in differentiating patients from healthy individuals. The proposed method allows accurate, correct, timely, and cost-effective diagnoses. In line with the application of intelligent expert systems, development of this system is presented as a new outlook for medical systems.