Heuristic Optimization of Deep and Shallow Classifiers: An Application for Electroencephalogram Cyclic Alternating Pattern Detection

Abstract

Methodologies for automatic non-rapid eye movement and cyclic alternating pattern analysis were proposed to examine the signal from one electroencephalogram monopolar derivation for the A phase, cyclic alternating pattern cycles, and cyclic alternating pattern rate assessments. A population composed of subjects free of neurological disorders and subjects diagnosed with sleep-disordered breathing was studied. Parallel classifications were performed for non-rapid eye movement and A phase estimations, examining a one-dimension convolutional neural network (fed with the electroencephalogram signal), a long short-term memory (fed with the electroencephalogram signal or with proposed features), and a feed-forward neural network (fed with proposed features), along with a finite state machine for the cyclic alternating pattern cycle scoring. Two hyper-parameter tuning algorithms were developed to optimize the classifiers. The model with long short-term memory fed with proposed features was found to be the best, with accuracy and area under the receiver operating characteristic curve of 83% and 0.88, respectively, for the A phase classification, while for the non-rapid eye movement estimation, the results were 88% and 0.95, respectively. The cyclic alternating pattern cycle classification accuracy was 79% for the same model, while the cyclic alternating pattern rate percentage error was 22%.

Towards automatic EEG cyclic alternating pattern analysis: a systematic review

“…The results indicated that using features improved performance, likely because the limited data did not allow the deep learning model to learn all relevant characteristics. These features were later used by Mendonça et al [ 65 ] that conducted a similar analysis but proposed the Heuristic Oriented Search Algorithm (HOSA) for optimizing the structure of deep learning models. The authors examined the performance of LSTM fead with features agains the LSTM fead with the preprocessed EEG signal, and concluded again that the use of the feature-based model was superior for the same reason as previously stated.…”

Prediction of IOL decentration, tilt and axial position using anterior segment OCT data

“…For the prediction of the axial IOL equator plane position (as a scalar parameter), a simple univariate model is sufficient. We implemented 2 different models for each prediction: First, we set up a shallow feedforward neural network [ 20 , 30 ], and second, we implemented a multilinear regression model (bivariate for decentration and tilt and univariate for axial IOL equator position) as a reference [ 23 ].…”

“…the optimisation was performed in terms of minimising the mean squared prediction error, which refers to a metric for the performance of the prediction [ 20 ]: …”

“…The Levenberg-Marquardt algorithm [18,19] was used as the training function, as this algorithm is known to have a good performance in terms of convergence and stability. Based on the squared prediction error (derived from the 2 component vectors for A and B or the scalar for C) (difference of predicted and observed/measured target value) the optimisation was performed in terms of minimising the mean squared prediction error, which refers to a metric for the performance of the prediction [20]:…”