Rui Pedro Paiva

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.

Melody Detection in Polyphonic Musical Signals: Exploiting Perceptual Rules, Note Salience, and Melodic Smoothness

80 Computer Music Journal Melody extraction from polyphonic audio is a research area of increasing interest. It has a wide range of applications in various fields, including music information retrieval (MIR, particularly in query-by-humming, where the user hums a tune to search a database of musical audio), automatic melody transcription, performance and expressiveness analysis, extraction of melodic descriptors for music content metadata, and plagiarism detection, to name but a few. This area has become increasingly relevant in recent years, as digital music archives are continuously expanding. The current state of affairs presents new challenges to music librarians and service providers regarding the organization of large-scale music databases and the development of meaningful methods of interaction and retrieval. In this article, we address the problem of melody detection in polyphonic audio following a multistage approach, inspired by principles from perceptual theory and musical practice. Our system comprises three main modules: pitch detection, determination of musical notes (with precise temporal boundaries, pitches, and intensity levels), and identification of melodic notes. The main contribution of this article is in the last module, in which a number of rule-based systems are proposed that attempt to extract the notes that convey the main melodic line among the whole set of detected notes. The system performs satisfactorily in a small database collected by us and in the database created for the ISMIR 2004 melody extraction contest. However, the performance of the algorithm decreased in the MIREX 2005 database. Related Work

Noise detection during heart sound recording using periodicity signatures.

Heart sound is a valuable biosignal for diagnosis of a large set of cardiac diseases. Ambient and physiological noise interference is one of the most usual and highly probable incidents during heart sound acquisition. It tends to change the morphological characteristics of heart sound that may carry important information for heart disease diagnosis. In this paper, we propose a new method applicable in real time to detect ambient and internal body noises manifested in heart sound during acquisition. The algorithm is developed on the basis of the periodic nature of heart sounds and physiologically inspired criteria. A small segment of uncontaminated heart sound exhibiting periodicity in time as well as in the time-frequency domain is first detected and applied as a reference signal in discriminating noise from the sound. The proposed technique has been tested with a database of heart sounds collected from 71 subjects with several types of heart disease inducing several noises during recording. The achieved average sensitivity and specificity are 95.88% and 97.56%, respectively.

Comparison of systolic time interval measurement modalities for portable devices

Systolic time intervals (STI) have shown significant diagnostic and prognostic value to assess the global cardiac function. Their value has been largely established in hospital settings. Currently, STI are considered a promising tool for long-term patient follow-up with chronic cardiovascular diseases. Several technologies exist that enable beat-by-beat assessment of STI in personal health application scenarios. A comparative study is presented using the echocardiographic gold standard synchronized with impedance cardiography (ICG), phonocardiography (PCG) and photoplethysmography (PPG). The ability of these competing technologies in assessing the pre ejection period (PEP) and the left ventricle ejection time (LVET) is given a general overview with comparative results.

Detection of motion artifact patterns in photoplethysmographic signals based on time and period domain analysis.

The presence of motion artifacts in photoplethysmographic (PPG) signals is one of the major obstacles in the extraction of reliable cardiovascular parameters in continuous monitoring applications. In the current paper we present an algorithm for motion artifact detection based on the analysis of the variations in the time and the period domain characteristics of the PPG signal. The extracted features are ranked using a normalized mutual information feature selection algorithm and the best features are used in a support vector machine classification model to distinguish between clean and corrupted sections of the PPG signal. The proposed method has been tested in healthy and cardiovascular diseased volunteers, considering 11 different motion artifact sources. The results achieved by the current algorithm (sensitivity--SE: 84.3%, specificity--SP: 91.5% and accuracy--ACC: 88.5%) show that the current methodology is able to identify both corrupted and clean PPG sections with high accuracy in both healthy (ACC: 87.5%) and cardiovascular diseases (ACC: 89.5%) context.

Heart murmur classification with feature selection

Heart sounds entail crucial heart function information. In conditions of heart abnormalities, such as valve dysfunctions and rapid blood flow, additional sounds are heard in regular heart sounds, which can be employed in pathology diagnosis. These additional sounds, or so-called murmurs, show different characteristics with respect to cardiovascular heart diseases, namely heart valve disorders. In this paper, we present a method of heart murmur classification composed by three basic steps: feature extraction, feature selection, and classification using a nonlinear classifier. A new set of 17 features extracted in the time, frequency and in the state space domain is suggested. The features applied for murmur classification are selected using the floating sequential forward method (SFFS). Using this approach, the original set of 17 features is reduced to 10 features. The classification results achieved using the proposed method are compared on a common database with the classification results obtained using the feature sets proposed in two well-known state of the art methods for murmur classification. The achieved results suggest that the proposed method achieves slightly better results using a smaller feature set.

Beat-to-beat systolic time-interval measurement from heart sounds and ECG

Systolic time intervals are highly correlated to fundamental cardiac functions. Several studies have shown that these measurements have significant diagnostic and prognostic value in heart failure condition and are adequate for long-term patient follow-up and disease management. In this paper, we investigate the feasibility of using heart sound (HS) to accurately measure the opening and closing moments of the aortic heart valve. These moments are crucial to define the main systolic timings of the heart cycle, i.e. pre-ejection period (PEP) and left ventricular ejection time (LVET). We introduce an algorithm for automatic extraction of PEP and LVET using HS and electrocardiogram. PEP is estimated with a Bayesian approach using the signals instantaneous amplitude and patient-specific time intervals between atrio-ventricular valve closure and aortic valve opening. As for LVET, since the aortic valve closure corresponds to the start of the S2 HS component, we base LVET estimation on the detection of the S2 onset. A comparative assessment of the main systolic time intervals is performed using synchronous signal acquisitions of the current gold standard in cardiac time-interval measurement, i.e. echocardiography, and HS. The algorithms were evaluated on a healthy population, as well as on a group of subjects with different cardiovascular diseases (CVD). In the healthy group, from a set of 942 heartbeats, the proposed algorithm achieved 7.66 ± 5.92 ms absolute PEP estimation error. For LVET, the absolute estimation error was 11.39 ± 8.98 ms. For the CVD population, 404 beats were used, leading to 11.86 ± 8.30 and 17.51 ± 17.21 ms absolute PEP and LVET errors, respectively. The results achieved in this study suggest that HS can be used to accurately estimate LVET and PEP.

Detection of motion artifacts in photoplethysmographic signals based on time and period domain analysis

The presence of motion artifacts in the photoplethysmographic (PPG) signals is one of the major obstacles in the extraction of reliable cardiovascular parameters in real time and continuous monitoring applications. In the current paper we present an algorithm for motion artifact detection, which is based on the analysis of the variations in the time and period domain characteristics of the PPG signal. The extracted features are ranked using a feature selection algorithm (NMIFS) and the best features are used in a Support Vector Machine classification model to distinguish between clean and corrupted sections of the PPG signal. The results achieved by the current algorithm (SE: 0.827 and SP: 0.927) show that both time and especially period domain features play an important role in the discrimination of motion artifacts from clean PPG pulses.

Assessing Systolic Time-Intervals from Heart Sound: A Feasibility Study

Systolic time intervals are highly correlated to fundamental cardiac functions. In this paper we investigate the feasibility of using heart sound (HS) to accurately measure the opening and closing moments of the aortic valve, since these are crucial moments to define the main systolic timings of the heart cycle, i.e. the pre-ejection period (PEP) and the left ventricular ejection time (LVET). We introduce a HS model, which is applied to define several features that provide clear markers to identify these moments in the HS. Using these features and a comparative analysis with registered echocardiographies from 17 subjects, the results achieved in this study suggest that HS can be used to accurately estimate LVET and PEP.

Assessing PEP and LVET from heart sounds: Algorithms and evaluation

This paper addresses the estimation of systolic time intervals, namely the pre-ejection period (PEP) and the left ventricular ejection time (LVET), using heart sound. PEP is estimated with a Bayesian approach resorting to the signals instantaneous amplitude and typical time intervals between atrio-ventricular valve closure and aortic valve opening. As for LVET, aortic valve closure is determined through the analysis of a high-frequency signature of S2. Additionally, LVET has also been estimated from a PPG signal at a peripheral site, for the sake of comparison over a subset of data. We evaluated our algorithms on a set of 658 heartbeats and achieved 10.32 msec average absolute PEP estimation error with 7.3 msec standard deviation and for LVET, 15.8 msec average estimation error with 13.6 msec standard deviation. Current results support our assumption that heart sounds can be applied to detect the onset of the aortic valve movement processes.