Maximum Likelihood Estimation improved EEG-based classification accuracy for major depressive disorder detection to 86.66% and prediction accuracy to 100%, outperforming PCA and baseline methods.
Does Maximum Likelihood Estimation for EEG channel selection improve the accuracy of detecting Major Depressive Disorder compared to PCA or baseline neural networks?
Maximum Likelihood Estimation for EEG channel selection enhances the performance of machine learning classifiers in detecting Major Depressive Disorder.
Effect estimate: Accuracy increased from 77.77% (baseline) to 86.66% (MLE); prediction accuracy improved to 100% with MLE vs 98.02% with PCA; sensitivity increased from 70% baseline to 74.77% with MLE
Absolute Event Rate: 86.66% vs 77.77%
Abstract Major depressive disorder is a serious and debilitating mental health condition that affects millions worldwide. Diagnosis is important mainly for the successful treatment of the disorder, while traditional clinical judgment can be subjective and sometimes less than accurate. In this work, we take the opposite tack by exploring electroencephalogram signals. We applied MLE to select only those EEG channels that have the highest ranking weight. This filtering sharpens the data and helps in improving the performance of machine learning classifiers. We compared EEG data from people with MDD to that with controls, making sure the comparisons were fair. The complete process was tested with an MLPNN classifier. These results show that for classification accuracy, the use of MLE for channel selection improves upon relying on the neural network alone and points toward a strong tool for the improvement of MDD detection. EEG patterns clearly separate the subjects with MDD from the controls. The difference is striking. Though the combination of MLE with the MLPNN holds great promise for diagnosis, we will go even further: new classification algorithms and new strategies for channel selection in a hunt for maximum precision. This work brings us closer to automated mental health assessment effectively. Speaking for themselves are the metrics: baseline accuracy sat at 77.77%, rose to 83.33% with PCA, and climbed to 86.66% with MLE. Error rates fell commensurately-from 22.23% to 13.34% with MLE. Prediction accuracy was especially dramatic: 100% with MLE, versus 98.02% with PCA. Sensitivity followed the same upward path, with true-positive rates rising from 70% to 72.22% with PCA and to 74.77% with MLE. These figures show that both techniques are soundly effective.
Choudhary et al. (Mon,) conducted a other in Adults with major depressive disorder and healthy controls undergoing EEG for MDD detection. Maximum Likelihood Estimation (MLE) channel selection combined with Multi-layer Perceptron Neural Network (MLPNN) classifier vs. Principal Component Analysis (PCA) channel selection and baseline neural network without feature selection was evaluated on Classification accuracy, error rate, prediction accuracy, sensitivity for detection of major depressive disorder using EEG data (Accuracy increased from 77.77% (baseline) to 86.66% (MLE); prediction accuracy improved to 100% with MLE vs 98.02% with PCA; sensitivity increased from 70% baseline to 74.77% with MLE). Maximum Likelihood Estimation improved EEG-based classification accuracy for major depressive disorder detection to 86.66% and prediction accuracy to 100%, outperforming PCA and baseline methods.