Pavel Dohnálek

Comparison of Classification Algorithms for Physical Activity Recognition

The main aim of this work is to compare different algorithms for human physical activity recognition from accelerometric and gyroscopic data which are recorded by a smartphone. Three classification algorithms were compared: the Linear Discriminant Analysis, the Random Forest, and the K-Nearest Neighbours. For better classification performance, two feature extraction methods were tested: the Correlation Subset Evaluation Method and the Principal Component Analysis. The results of experiment were expressed by confusion matrixes.

Classification of cardiotocography records by random forest

The cardiotocography (CTG) is a diagnostic method which is widely used in prenatal care. The CTG is indicated since 27 weeks of pregnancy and it measures heart activity, uterine contraction and fetal movement. Results of the CTG allow recognizing of three basic different fetal states (physiological, suspect and pathological) and an obstetrician can determine a diagnosis and evaluate situation which can lead to the fetus death. The main aim of this work is to suggest and to test algorithm for automatic recognition of above mentioned states. This task is especially used in prenatal care as a support decision system.

Human activity recognition on raw sensor data via sparse approximation

Human physical activity monitoring is a relatively new problem drawing much attention over the last years due to its wide application in medicine, homecare systems, prisoner monitoring etc. This paper presents Orthogonal Matching Pursuit based classifier as a method for activity recognition and proposes a modification to the classifier that significantly increases recognition accuracy. Both methods show promising results in both total recognition and differentiation between certain activities achieving up to 99.60% recognition accuracy even without any prior data processing. A comparison with other methods is also provided.

Pattern Recognition in EEG Cognitive Signals Accelerated by GPU

Analysing of Electroencephalography (EEG) cognitive signals becomes more popular today due to availability of essential hardware (EEG headsets) and sufficient computation power of common computers. Fast and precise pattern matching of acquired signals represents one of the most important challenges. In this article, a method for signal pattern matching based on Non-negative Matrix Factorization is proposed. We also utilize short-time Fourier transform to preprocess EEG data and Cosine Similarity Measure to perform query-based classification. The recognition algorithm shows promising results in execution speed and is suitable for implementation on graphics processors to achieve real-time processing, making the proposed method suitable for real-world, real-time applications. In terms of recognition accuracy, our experiments show that accuracy greatly depends on the choice of input parameters.

Human Fetus Health Classification on Cardiotocographic Data Using Random Forests

Pregnancy and fetus development is an extremely complex biological process that, while generally successful and without complications, can go wrong. One of the methods to determine if the fetus is developing according to expectations is cardiotocography. This diagnostic technique’s purpose is to measure the heartbeat of the fetus and uterine contractions of its mother, usually during the third trimester of pregnancy when the fetus’ heart is fully functional. Outputs of a cardiotocogram are usually interpreted as belonging to one of three states: physiological, suspicious and pathological. Automatic classification of these states based on cardiotocographic data is the goal of this paper. In this research, the Random Forest method is show to perform very well, capable of classifying the data with 94.69% accuracy. A comparison with the Classification and Regression Tree and Self-organizing Map methods is also provided.

Performance evaluation of Random Forest regression model in tracking Parkinson's disease progress

In this paper, capabilities of the Random Forest algorithm are tested with application to the Parkinsons disease progression that can be determined from speech. Results are compared with the linear regression model and the Classification and Regression Tree method. Mean Squared Error and Mean Absolute Error values were calculated and compared for each of the approaches. The Random Forest algorithm belongs to the group model category and usually improves the results achieved by regression trees, making it more suitable for fighting the disease.

A parallel Fruchterman–Reingold algorithm optimized for fast visualization of large graphs and swarms of data

Abstract Graphs in computer science are widely used in social network analysis, computer networks, transportation networks, and many other areas. In general, they can visualize relationships between objects. However, fast drawing of graphs and other structures containing large numbers of data points with readable layouts is still a challenge. This paper describes a novel variant of the Fruchterman–Reingold graph layout algorithm which is adapted to GPU parallel architecture. A new approach based on space-filling curves and a new way of repulsive forces computation on GPU are described. The paper contains both performance and quality tests of the new algorithm.

Recognition of pathological beats in ECG signals based on Singular Value Decomposition of wavelet coefficients and support vector machine

The main goal of this work is to describe possibilities of the Singular Value Decomposition in the task of arrhythmia recognition. Many approaches try to recognize pathological beats in the time domain, our approach transforms an ECG signal from time to frequency domain, where it is reduced by Singular Value Decomposition. The new feature subspace was classified by three basic algorithms: Support Vector Machine, Linear Discriminant Analysis and Classification tree. The results were compared. Our approach increases the quality of classification and the obtained results are comparable with the results available in literature. The main aim of the proposed solution is to differentiate between physiological and pathological beats such as Premature Ventricular Contraction, Right Bundle Branch Block and Left Bundle Branch Block beats.

Comparison of Feature Reduction Methods in the Task of Arrhythmia Classification

The main goal of the work is to test two well-known algorithms for feature transform such as Singular Value Decomposition and Principal Component Analysis in the task of arrhythmia recognition in ECG records. The original signal were transformed by these two techniques and a neural network was used for classification. Values of sensitivity and accuracy were observed and consequently compared for each transformation. Unlike in other similar works, our experiments were performed on a high number of beats and the tested database included over 47 000 experimental heart beats with different diseases.

Common Tensor Discriminant Analysis for human brainwave recognition accelerated by massive parallelism

In this paper, a massively parallel implementation of Common Tensor Discriminant Analysis is presented with applications to human brainwave pattern recognition. The implementation, accelerated by the NVIDIA Compute Unified Device Architecture technology, is shown to be 11.49x faster than the original MATLAB version. Before processing by the discriminant analysis, the data is segmented by a sliding window and converted into the time-frequency domain by the continuous wavelet transform.