Publications

Gas chromatography—ion mobility spectrometry (GC-IMS) allows the fast, reliable, and inexpensive chemical composition analysis of volatile mixtures. This sensing technology has been successfully employed in food science to determine food origin, freshness and preventing alimentary fraud. However, GC-IMS data is highly dimensional, complex, and suffers from strong non-linearities, baseline problems, misalignments, peak overlaps, long peak tails, etc., all of which must be corrected to properly extract the relevant features from samples. In this work, a pipeline for signal pre-processing, followed by four different approaches for feature extraction in GC-IMS data, is presented. More precisely, these approaches consist of extracting data features from: (1) the total area of the reactant ion peak chromatogram (RIC); (2) the full RIC response; (3) the unfolded sample matrix; and (4) the ion peak volumes. The resulting pipelines for data processing were applied to a dataset consisting of two different quality class Iberian ham samples, based on their feeding regime. The ability to infer chemical information from samples was tested by comparing the classification results obtained from partial least-squares discriminant analysis (PLS-DA) and the samples’ variable importance for projection (VIP) scores. The choice of a feature extraction strategy is a trade-off between the amount of chemical information that is preserved, and the computational effort required to generate the data models.

JTD Keywords: authenticity, classification, electronic-nose, feature extraction, food analysis, gc-ims, headspace, least-squares, models, pld-da, pre-processing, quality, sensory analysis, wine, Feature extraction, Food analysis, Gc-ims, Hs-gc-ims, Pld-da, Pre-processing

This paper describes the design of a linear phase low-pass differentiator filter with a finite impulse response (FIR) for extracting transient features of gas sensor signals (the so-called “bouts”). The detection of these bouts is relevant for estimating the distance of a gas source in a turbulent plume. Our current proposal addresses the shortcomings of previous ‘bout’ estimation methods, namely: (i) they were based in non-causal digital filters precluding real time operation, (ii) they used non-linear phase filters leading to waveform distortions and (iii) the smoothing action was achieved by two filters in cascade, precluding an easy tuning of filter performance. The presented method is based on a low-pass FIR differentiator, plus proper post-processing, allowing easy algorithmic implementation for real-time robotic exploration. Linear phase filters preserve signal waveform in the bandpass region for maximum reliability concerning both ‘bout’ detection and amplitude estimation. As a case study, we apply the proposed filter to predict the source distance from recordings obtained with metal oxide (MOX) gas sensors in a wind tunnel. We first perform a joint optimization of the cut-off frequency of the filter and the bout amplitude threshold, for different wind speeds, uncovering interesting relationships between these two parameters. We demonstrate that certain combinations of parameters can reduce the prediction error to 8 cm (in a distance range of 1.45 m) improving previously reported performances in the same dataset by a factor of 2.5. These results are benchmarked against traditional source distance estimators such as the mean, variance and maximum of the response. We also study how the length of the measurement window affects the performance of different signal features, and how to select the filter parameters to make the predictive models more robust to changes in wind speed. Finally, we provide a MATLAB implementation of the bout detection algorithm and all analysis code used in this study.

JTD Keywords: Gas sensors, Differentiator, Low pass filter, Metal oxide semiconductor, MOX sensors, Signal processing, Feature extraction, Gas source localization, Robotics

This paper explores which signal features of a gas sensor are optimum for assessing the proximity to a gas source in an open environment. Specifically, we compare three statistical descriptors of the signal (mean, variance and maximum response) against the 'bout' frequency, a feature computed in the derivative of the response. The experimental setup includes a generator of turbulent plumes and a sensing board composed of three metal oxide (MOX) sensors of different types. The main conclusion is that the maximum response is the most robust feature across the three sensors. The 'bout' frequency can be very sensitive to an additional parameter (the noise threshold).

JTD Keywords: Feature extraction, Gas plume, Gas sensors, Gas source localization, MOX, Signal processing

Estimating the affective state of a user during a computer task traditionally relies on either subjective reports or analysis of physiological signals, facial expressions, and other measures. These methods have known limitations, can be intrusive and may require specialized equipment. An alternative would be employing a ubiquitous device of everyday use such as a standard keyboard. Here we investigate if we can infer the emotional state of a user by analyzing their typing patterns. To test this hypothesis, we asked 400 participants to caption a set of emotionally charged images taken from a standard database with known ratings of arousal and valence. We computed different keystroke pattern dynamics, including keystroke duration (dwell time) and latency (flight time). By computing the mean value of all of these features for each image, we found a statistically significant negative correlation between dwell times and valence, and between flight times and arousal. These results highlight the potential of using keystroke dynamics to estimate the affective state of a user in a non-obtrusive way and without the need for specialized devices.

JTD Keywords: Feature extraction, Correlation, Keyboards, Task analysis, Statistical analysis, Affective computing, Standards, Keystroke, Keyboard, Typing, Arousal, Valence, Affect

Differentiating normal from adventitious respiratory sounds (RS) is a major challenge in the diagnosis of pulmonary diseases. Particularly, continuous adventitious sounds (CAS) are of clinical interest because they reflect the severity of certain diseases. This study presents a new classifier that automatically distinguishes normal sounds from CAS. It is based on the multi-scale analysis of instantaneous frequency (IF) and envelope (IE) calculated after ensemble empirical mode decomposition (EEMD). These techniques have two major advantages over previous techniques: high temporal resolution is achieved by calculating IF-IE and a priori knowledge of signal characteristics is not required for EEMD. The classifier is based on the fact that the IF dispersion of RS signals markedly decreases when CAS appear in respiratory cycles. Therefore, CAS were detected by using a moving window to calculate the dispersion of IF sequences. The study dataset contained 1494 RS segments extracted from 870 inspiratory cycles recorded from 30 patients with asthma. All cycles and their RS segments were previously classified as containing normal sounds or CAS by a highly experienced physician to obtain a gold standard classification. A support vector machine classifier was trained and tested using an iterative procedure in which the dataset was randomly divided into training (65%) and testing (35%) sets inside a loop. The SVM classifier was also tested on 4592 simulated CAS cycles. High total accuracy was obtained with both recorded (94.6% ± 0.3%) and simulated (92.8% ± 3.6%) signals. We conclude that the proposed method is promising for RS analysis and classification.

JTD Keywords: Diseases, Dispersion, Empirical mode decomposition, Feature extraction, Informatics, Support vector machines

Changes in the left ventricle function produce alternans in the hemodynamic and electric behavior of the cardiovascular system. A total of 49 cardiomyopathy patients have been studied based on the blood pressure signal (BP), and were classified according to the left ventricular ejection fraction (LVEF) in low risk (LR: LVEF>35%, 17 patients) and high risk (HR: LVEF≤35, 32 patients) groups. We propose to characterize these patients using a linear and a nonlinear methods, based on the spectral estimation and the recurrence plot, respectively. From BP signal, we extracted each systolic time interval (STI), upward systolic slope (BPsl), and the difference between systolic and diastolic BP, defined as pulse pressure (PP). After, the best subset of parameters were obtained through the sequential feature selection (SFS) method. According to the results, the best classification was obtained using a combination of linear and nonlinear features from STI and PP parameters. For STI, the best combination was obtained considering the frequency peak and the diagonal structures of RP, with an area under the curve (AUC) of 79%. The same results were obtained when comparing PP values. Consequently, the use of combined linear and nonlinear parameters could improve the risk stratification of cardiomyopathy patients.

JTD Keywords: Feature extraction, Blood pressure, Heart rate, Estimation, Data mining, Covariance matrices, Hospitals

Aging population is a major concern that is reflected in the increase of chronic diseases. Heart Failure (HF) is one of the most common chronic diseases of elderly people that is punctuated with acute episodes, which result in hospitalization. The periodic modulation of the amplitude of the breathing pattern is proved to be one of the multiple symptoms of an acute episode, and thus, the features extracted from its characterization contribute in the improvement of the first diagnosis of the clinical practice. The main objective of this study is to evaluate if the features extracted from the breathing pattern along with common clinical variables are reliable enough to detect Periodic Breathing (PB). A dataset of 44 elderly patients containing clinical information and a short record of electrocardiogram and respiratory flow signal was used to train two machine learning classification methods: Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA). All the available clinical parameters within the dataset along with the parameters characterizing the respiratory pattern were used to classify the observations into two groups. SVM classification was optimized and performed using a = -8 and C = 10.04 giving an accuracy of 88.2 % sensitivity of 90 % and specificity of 85.7 % Similar results were achieved with LDA classifying with an accuracy of 82.4 %, a sensitivity of 81.8% and specificity of 83.3 % PB has been accurately detected using both classifiers.

JTD Keywords: Support vector machines, Feature extraction, Training, Senior citizens, Standards, Training data, Hospitals

One of the most challenging problems in intensive care is still the process of discontinuing mechanical ventilation, called weaning process. Both an unnecessary delay in the discontinuation process and a weaning trial that is undertaken too early are undesirable. In this study, we analyzed respiratory pattern variability using the respiratory volume signal of patients submitted to two different levels of pressure support ventilation (PSV), prior to withdrawal of the mechanical ventilation. In order to characterize the respiratory pattern, we analyzed the following time series: inspiratory time, expiratory time, breath duration, tidal volume, fractional inspiratory time, mean inspiratory flow and rapid shallow breathing. Several autoregressive modeling techniques were considered: autoregressive models (AR), autoregressive moving average models (ARMA), and autoregressive models with exogenous input (ARX). The following classification methods were used: logistic regression (LR), linear discriminant analysis (LDA) and support vector machines (SVM). 20 patients on weaning trials from mechanical ventilation were analyzed. The patients, submitted to two different levels of PSV, were classified as low PSV and high PSV. The variability of the respiratory patterns of these patients were analyzed. The most relevant parameters were extracted using the classifiers methods. The best results were obtained with the interquartile range and the final prediction errors of AR, ARMA and ARX models. An accuracy of 95% (93% sensitivity and 90% specificity) was obtained when the interquartile range of the expiratory time and the breath duration time series were used a LDA model. All classifiers showed a good compromise between sensitivity and specificity.

JTD Keywords: autoregressive moving average processes, feature extraction, medical signal processing, patient care, pneumodynamics, signal classification, support vector machines, time series, ARX, autoregressive modeling techniques, autoregressive models with exogenous input, autoregressive moving average model, breath duration time series, classification method, classifier method, discontinuing mechanical ventilation, expiratory time, feature extraction, final prediction errors, fractional inspiratory time, intensive care, interquartile range, linear discriminant analysis, logistic regression analysis, mean inspiratory flow, patient respiratory volume signal, pressure support level, pressure support ventilation, rapid shallow breathing, respiratory pattern variability characterization, support vector machines, tidal volume, weaning trial, Analytical models, Autoregressive processes, Biological system modeling, Estimation, Support vector machines, Time series analysis, Ventilation

This work investigates the very low frequency (VLF) modulation of QRS slopes and heart rate variability (HRV). Electrocardiogram (ECG) and respiratory flow signal were acquired from patients with dilated cardiomyopathy and ischemic cardiomyopathy. HRV as well as the upward QRS slope (I_US) and downward QRS slope (I_DS) were extracted from the ECG. The relation between HRV and QRS slopes in the VLF band was measured using ordinary coherence in 5-minute segments. Partial coherence was then used to remove the influence that respiration simultaneously exerts on HRV and QRS slopes. A statistical threshold was determined, below which coherence values were considered not to represent a linear relation. 7 out of 276 segments belonging to 5 out of 29 patients for I_US and 10 segments belonging to 5 patients for I_DS presented a VLF modulation in QRS slopes, HRV and respiration. In these segments spectral coherence was statistically significant, while partial coherence decreased, indicating that the coupling HRV and QRS slopes was related to respiration. 4 segments had a partial coherence value below the threshold for I_US, 3 segments for I_DS. The rest of the segments also presented a notable decrease in partial coherence, but still above the threshold, which means that other non-linearly effects may also affect this modulation.

JTD Keywords: diseases, electrocardiography, feature extraction, medical signal processing, pneumodynamics, statistical analysis, ECG, QRS slopes, cardiomyopathy patients, dilated cardiomyopathy, electrocardiogram, feature extraction, heart rate variability, ischemic cardiomyopathy, ordinary coherence, partial coherence value, respiration, respiratory flow signal acquisition, spectral coherence, statistical threshold, time 5 min, very low frequency modulation, Coherence, Educational institutions, Electrocardiography, Frequency modulation, Heart rate variability

Some of the most common clinical problems in elderly patients are related to diseases of the cardiac and respiratory systems. Elderly patients often have altered breathing patterns, such as periodic breathing (PB) and Cheyne-Stokes respiration (CSR), which may coincide with chronic heart failure. In this study, we used the envelope of the respiratory flow signal to characterize respiratory patterns in elderly patients. To study different breathing patterns in the same patient, the signals were segmented into windows of 5 min. In oscillatory breathing patterns, frequency and time-frequency parameters that characterize the discriminant band were evaluated to identify periodic and non-periodic breathing (PB and nPB). In order to evaluate the accuracy of this characterization, we used a feature selection process, followed by linear discriminant analysis. 22 elderly patients (7 patients with PB and 15 with nPB pattern) were studied. The following classification problems were analyzed: patients with either PB (with and without apnea) or nPB patterns, and patients with CSR versus PB, CSR versus nPB and PB versus nPB patterns. The results showed 81.8% accuracy in the comparisons of nPB and PB patients, using the power of the modulation peak. For the segmented signal, the power of the modulation peak, the frequency variability and the interquartile ranges provided the best results with 84.8% accuracy, for classifying nPB and PB patients.

JTD Keywords: cardiovascular system, diseases, feature extraction, geriatrics, medical signal processing, oscillations, pneumodynamics, signal classification, time-frequency analysis, Cheyne-Stokes respiration, apnea, cardiac systems, chronic heart failure, classification problems, discriminant band, diseases, elderly patients, feature selection process, frequency variability, interquartile ranges, linear discriminant analysis, nonperiodic breathing, oscillatory breathing pattern, periodic breathing, respiratory How signal, respiratory systems, signal segmentation, time 5 min, time-frequency parameters, Accuracy, Aging, Frequency modulation, Heart, Senior citizens, Time-frequency analysis

The differentiation of obstructive and central respiratory events is a major challenge in the diagnosis of sleep disordered breathing. Esophageal pressure (Pes) measurement is the gold-standard method to identify these events but its invasiveness deters its usage in clinical routine. Flattening patterns appear in the airflow signal during episodes of inspiratory flow limitation (IFL) and have been shown with invasive techniques to be useful to differentiate between central and obstructive hypopneas. In this study we present a new method for the automatic non-invasive differentiation of obstructive and central hypopneas solely with nasal airflow. An overall of 36 patients underwent full night polysomnography with systematic Pes recording and a total of 1069 hypopneas were manually scored by human experts to create a gold-standard annotation set. Features were automatically extracted from the nasal airflow signal to train and test our automatic classifier (Discriminant Analysis). Flattening patterns were non-invasively assessed in the airflow signal using spectral and time analysis. The automatic non-invasive classifier obtained a sensitivity of 0.71 and an accuracy of 0.69, similar to the results obtained with a manual non-invasive classification algorithm. Hence, flattening airflow patterns seem promising for the non-invasive differentiation of obstructive and central hypopneas.

JTD Keywords: Practical, Experimental/ biomedical measurement, Feature extraction, Flow measurement, Medical disorders, Medical signal processing, Patient diagnosis, Pneumodynamics, Pressure measurement, Signal classification, Sleep, Spectral analysis/ automatic noninvasive differentiation, Obstructive hypopnea, Central hypopnea, Inspiratory flow limitation, Nasal airflow, Esophageal pressure, Polysomnography, Feature extraction, Discriminant analysis, Spectral analysis

Sleep Apnea-Hypopnea Syndrome (SAHS) diagnosis is still done with an overnight multi-channel polysomnography. Several efforts are being made to study profoundly the snore mechanism and discover how it can provide an opportunity to diagnose the disease. This work introduces the concept of regular snores, defined as the ones produced in consecutive respiratory cycles, since they are produced in a regular way, without interruptions. We applied 2 thresholds (TH/sub adaptive/ and TH/sub median/) to the time interval between successive snores of 34 subjects in order to select regular snores from the whole all-night snore sequence. Afterwards, we studied the effectiveness that parameters, such as time interval between successive snores and the mean intensity of snores, have on distinguishing between different levels of SAHS severity (AHI (Apnea-Hypopnea Index)<5h/sup -1/, AHI<10 h/sup -1/, AHI<15h/sup -1/, AHI<30h/sup -1/). Results showed that TH/sub adaptive/ outperformed TH/sub median/ on selecting regular snores. Moreover, the outcome achieved with non-regular snores intensity features suggests that these carry key information on SAHS severity.

JTD Keywords: Practical, Experimental/ acoustic signal processing, Bioacoustics, Biomedical measurement, Diseases, Feature extraction, Medical signal processing, Patient diagnosis, Pneumodynamics, Sleep/ nonregular snore features, SAHS markers, Sleep apnea hypopnea syndrome, Overnight multichannel polysomnography, Snore mechanism