THROMBO - EMBOLIC STROKE PREDICTION AND DIAGNOSIS USING ARTIFICIAL NEURAL NETWORK AND GENETIC ALGORITHM

The Back Propagation Neural Network algorithm was validated using hypothetical data from fifty patients with symptoms of Stroke. The data set was divided into training set and test set while the validation data set were chosen randomly from the testing data. Forty-two (42) data set were used for the training set while eight data set were used for the test. Four data were chosen from the test set and used for the validation. A MATLAB program was written for training, testing and validation of the neural network. Three different architectures with 5, 10 and 20 hidden neurons in the network architecture were tested to avoid overfitting and inaccuracy after which neural network with 10 hidden neurons was chosen as the best architecture. The training error converged to 0 after 50 iterations with architecture of 10 hidden neurons while convergence was almost achieved after 100 and 1000 iteration steps with 5 and 10 hidden neurons respectively. The ANN was trained and tested after optimizing the input parameters using Genetic Algorithm, the overall predictive accuracy obtained for the thrombo-embolic stroke was 90%.