Machine learning algorithms have great potential for exploring the hidden patterns in the data sets of the medical domain. These patterns can be utilized for clinical diagnosis. Since medical diagnostic decision support systems have become more popular in clinical routine, it is critical to decide the best model to provide reliable machine learning techniques in diagnostic problems. The purpose of my research thesis is to employ existing machine learning approaches and enhance those using better pre-processing techniques with improved classification accuracy for disease diagnosis. The emphasis of this research thesis is particularly on diabetes disease.
In this research thesis, the performance of efficient pre-processing and the classification techniques have been evaluated. We proposed the application of automatic multilayer perceptron (AutoMLP) combined with outlier detection to predict diabetes. Parameter Optimization (adjusting parameters to get best accuracy of a classifier) is a long run problem of neural networks while defining the network topology of the neural network architecture. Human intervention is required in the training process to choose the best suitable parameters for the network. Researchers and practitioners found problem choosing learning rates and hidden units that work well with their dataset. This problem is being solved by AutoMLP algorithm, which to our knowledge has not been applied for diabetes prediction before. The experiments are conducted on two publicly available datasets: Pima Indians obtained from UCI repository and Biostat Diabetes Dataset. The proposed model helped achieve an accuracy of 88.70% and 95.00% on Pima Indians dataset and Biostat diabetes dataset respectively.
Published: