António Dourado

Interpretability and learning in neuro-fuzzy systems

Abstract A methodology for the development of linguistically interpretable fuzzy models from data is presented. The implementation of the model is conducted through the training of a neuro-fuzzy network, i.e., a neural net architecture capable of representing a fuzzy system. In the first phase, the structure of the model is obtained by means of subtractive clustering, which allows the extraction of a set of relevant rules based on a set of representative input–output data samples. In the second phase, the parameters of the model are tuned via the training of a neural network through backpropagation. In order to attain interpretability goals, the method proposed imposes some constraints on the tuning of the parameters and performs membership function merging. In this way, it will be easy to assign linguistic labels to each of the membership functions obtained, after training. Therefore, the model obtained for the system under analysis will be described by a set of linguistic rules, easily interpretable.

Epileptic seizure prediction using relative spectral power features

OBJECTIVE Prediction of epileptic seizures can improve the living conditions for refractory epilepsy patients. We aimed to improve sensitivity and specificity of prediction methods, and to reduce the number of false alarms. METHODS Relative combinations of sub-band spectral powers of electroencephalogram (EEG) recordings across all possible channel pairs were utilized for tracking gradual changes preceding seizures. By using a specifically developed feature selection method, a set of best candidate features were fed to support vector machines in order to discriminate cerebral state as preictal or non-preictal. RESULTS Proposed algorithm was evaluated on continuous long-term multichannel scalp and invasive recordings (183 seizures, 3565 h). The best results demonstrated a sensitivity of 75.8% (66 out of 87 seizures) and a false prediction rate of 0.1h(-1). Performance was validated statistically, and was superior to that of analytical random predictor. CONCLUSION Applying machine learning methods on a reduced subset of proposed features could predict seizure onsets with high performance. SIGNIFICANCE Our method was evaluated on long-term continuous recordings of overall about 5 months, contrary to majority of previous studies using short-term fragmented data. It is of very low computational cost, while providing acceptable levels of alarm sensitivity and specificity.

The EPILEPSIAE database: an extensive electroencephalography database of epilepsy patients.

From the very beginning the seizure prediction community faced problems concerning evaluation, standardization, and reproducibility of its studies. One of the main reasons for these shortcomings was the lack of access to high‐quality long‐term electroencephalography (EEG) data. In this article we present the EPILEPSIAE database, which was made publicly available in 2012. We illustrate its content and scope. The EPILEPSIAE database provides long‐term EEG recordings of 275 patients as well as extensive metadata and standardized annotation of the data sets. It will adhere to the current standards in the field of prediction and facilitate reproducibility and comparison of those studies. Beyond seizure prediction, it may also be of considerable benefit for studies focusing on seizure detection, basic neurophysiology, and other fields.

EPILAB: A software package for studies on the prediction of epileptic seizures

A Matlab®-based software package, EPILAB, was developed for supporting researchers in performing studies on the prediction of epileptic seizures. It provides an intuitive and convenient graphical user interface. Fundamental concepts that are crucial for epileptic seizure prediction studies were implemented. This includes, for example, the development and statistical validation of prediction methodologies in long-term continuous recordings. Seizure prediction is usually based on electroencephalography (EEG) and electrocardiography (ECG) signals. EPILAB is able to process both EEG and ECG data stored in different formats. More than 35 time and frequency domain measures (features) can be extracted based on univariate and multivariate data analysis. These features can be post-processed and used for prediction purposes. The predictions may be conducted based on optimized thresholds or by applying classifications methods such as artificial neural networks, cellular neuronal networks, and support vector machines. EPILAB proved to be an efficient tool for seizure prediction, and aims to be a way to communicate, evaluate, and compare results and data among the seizure prediction community.

Preprocessing effects of 22 linear univariate features on the performance of seizure prediction methods

Combining multiple linear univariate features in one feature space and classifying the feature space using machine learning methods could predict epileptic seizures in patients suffering from refractory epilepsy. For each patient, a set of twenty-two linear univariate features were extracted from 6 electroencephalogram (EEG) signals to make a 132 dimensional feature space. Preprocessing and normalization methods of the features, which affect the output of the seizure prediction algorithm, were studied in terms of alarm sensitivity and false prediction rate (FPR). The problem of choosing an optimal preictal time was tackled using 4 distinct values of 10, 20, 30, and 40 min. The seizure prediction problem has traditionally been considered a two-class classification problem, which is also exercised here. These studies have been conducted on the features obtained from 10 patients. For each patient, 48 different combinations of methods are compared to find the best configuration. Normalization by dividing by the maximum and smoothing are found to be the best configuration in most of the patients. The results also indicate that applying machine learning methods on a multidimensional feature space of 22 univariate features predicted seizure onsets with high performance. On average, the seizures were predicted in 73.9% of the cases (34 out of 46 in 737.9h of test data), with a FPR of 0.15 h(-1).

Adaptive RBFNN versus conventional self-tuning: comparison of two parametric model approaches for non-linear control

Abstract In this work a practical study evaluates two parametric modelling approaches — linear and non-linear (neural) — for automatic adaptive control. The neural adaptive control is based on a developed hybrid learning technique using an adaptive (on-line) learning rate for a Gaussian radial basis function neural network. The linear approach is used for a self-tuning pole-placement controller. A selective forgetting factor method is applied to both control schemes: in the neural case to estimate on-line the second-layer weights and in the linear case to estimate the parameters of the linear process model. These two techniques are applied to a laboratory-scaled bench plant with the possibility of dynamic changes and different types of disturbances. Experimental results show the superior performance of the neural approach particularly when there are dynamic changes in the process.

Calibration of ActiGraph GT3X, Actical and RT3 accelerometers in adolescents

Abstract The objective of this study was to develop count cut-points for three different accelerometer models: ActiGraph GT3X, RT3 and Actical to accurately classify physical activity intensity levels in adolescents. Seventy-nine adolescents (10–15 years) participated in this study. Accelerometers and oxygen consumption () data were collected at rest and during 11 physical activities of different intensities. Accelerometers were worn on the waist and was measured by a portable metabolic system: Cosmed K4b2. Receiver operating characteristic (ROC) curves were used to determine cut-points. Cut-points for sedentary (SED), moderate-to-vigorous (MVPA) and vigorous-intensity physical activity (VPA) were 46, 607 and 818 counts·15s−1 to the vertical axis of ActiGraph; 180, 757 and 1112 counts·15s−1 to the vector magnitude of ActiGraph; 17, 441 and 873 counts·15s−1 for Actical; and 5.6, 20.4 and 32.2 counts·s−1 for RT3, respectively. For all three accelerometer models, there was an almost perfect discrimination of SED and MVPA (ROC >0.97) and an excellent discrimination of VPA (ROC>0.90) observed. Areas under the ROC curves indicated better discrimination of MVPA by ActiGraph (AUC=0.994) and Actical (AUC=0.993) when compared to RT3 (AUC=0.983). The cut-points developed in this study for the ActiGraph (vector magnitude), RT3 and Actical accelerometer models can be used to monitor physical activity level of adolescents.

Supervision and c-Means clustering of PID controllers for a solar power plant

Abstract A hierarchical control strategy consisting on a supervisory switching of PID controllers, simplified using the c-Means clustering technique, is developed and applied to the distributed collector field of a solar power plant. The main characteristic of this solar plant is that the primary energy source, the solar radiation, cannot be manipulated. It varies throughout the day, causing changes in plant dynamics conducting to distinct several operating points. To guarantee good performances in all operating points, a local PID controller is tuned to each operating point and a supervisory strategy is proposed and applied to switch among these controllers accordingly to the actual measured conditions. Each PID controller has been tuned off-line, by the combination of a dynamic recurrent non-linear neural network model with a pole placement control design. To reduce the number of local controllers, to be selected by the supervisor, a c-Means clustering technique was used. Simulation and experimental results, obtained at Plataforma Solar de Almeria, Spain, are presented showing the effectiveness of the proposed approach.

Global optimization of energy and production in process industries: a genetic algorithm application

Abstract The process industries exhibit an increasing need for efficient management of all the factors that can reduce their operating costs, leading to the necessity for a global multi-objective optimization methodology that will enable the generation of optimum strategies, fulfilling the required restrictions. In this paper, a genetic algorithm is developed and applied for the optimal assignment of all the production sections in a particular mill in the kraft pulp and paper industry, in order to optimize energy the costs and production rate changes. This system is intended to implement all programmed or forced maintenance shutdowns, as well as all the reductions imposed in production rates.

Epileptic Seizure Classification Using Neural Networks with 14 Features

Epilepsy is one of the most frequent neurological disorders. The main method used in epilepsy diagnosis is electroencephalogram (EEG) signal analysis. However this method requires a time-consuming analysis when made manually by an expert due to the length of EEG recordings. This paper proposes an automatic classification system for epilepsy based on neural networks and EEG signals. The neural networks use 14 features (extracted from EEG) in order to classify the brain state into one of four possible epileptic behaviors: inter-ictal, pre-ictal, ictal and pos-ictal. Experiments were made in a (i) single patient (ii) different patients and (ii) multiple patients, using two datasets. The classification accuracies of 6 types of neural networks architectures are compared. We concluded that with the 14 features and using the data of a single patient results in a classification accuracy of 99%, while using a network trained for multiple patients an accuracy of 98% is achieved.