Voiceprint and machine learning models for early detection of bulbar dysfunction in ALS.

BACKGROUND AND OBJECTIVE: Bulbar dysfunction is a term used in amyotrophic lateral sclerosis (ALS). It refers to motor neuron disability in the corticobulbar area of the brainstem which leads to a dysfunction of speech and swallowing. One of the earliest symptoms of bulbar dysfunction is voice deterioration characterized by grossly defective articulation, extremely slow laborious speech, marked hypernasality and severe harshness. Recently, research efforts have focused on voice analysis to capture this dysfunction. The main aim of this paper is to provide a new methodology to diagnose this dysfunction automatically at early stages of the disease, earlier than clinicians can do. METHODS: The study focused on the creation of a voiceprint consisting of a pattern generated from the quasi-periodic components of a steady portion of the five Spanish vowels and the computation of the five principal and independent components of this pattern. Then, a set of statistically significant features was obtained using multivariate analysis of variance and the outcomes of the most common supervised classification models were obtained. RESULTS: The best model (random forest) obtained an accuracy, sensitivity and specificity of 88.3%, 85.0% and 95.0% respectively when classifying bulbar vs. control participants but the results worsened when classifying bulbar vs. no-bulbar patients (accuracy, sensitivity and specificity of 78.7%, 80.0% and 77.5% respectively for support vector machines). Due to the great uncertainty found in the annotated corpus of the ALS patients without bulbar involvement, we used a safe semi-supervised support vector machine to relabel the ALS participants diagnosed without bulbar involvement as bulbar and no-bulbar. The performance of the results obtained increased, especially when classifying bulbar and no-bulbar patients obtaining an accuracy, sensitivity and specificity of 91.0%, 83.3% and 100.0% respectively for support vector machines. This demonstrates that our model can improve the diagnosis of bulbar dysfunction compared not only with clinicians, but also the methods published to date. CONCLUSIONS: The results obtained demonstrate the efficiency and applicability of the methodology presented in this paper. It may lead to the development of a cheap and easy-to-use tool to identify this dysfunction in early stages of the disease and monitor progress.

Detecting Bulbar Involvement in Patients with Amyotrophic Lateral Sclerosis Based on Phonatory and Time-Frequency Features.

The term "bulbar involvement" is employed in ALS to refer to deterioration of motor neurons within the corticobulbar area of the brainstem, which results in speech and swallowing dysfunctions. One of the primary symptoms is a deterioration of the voice. Early detection is crucial for improving the quality of life and lifespan of ALS patients suffering from bulbar involvement. The main objective, and the principal contribution, of this research, was to design a new methodology, based on the phonatory-subsystem and time-frequency characteristics for detecting bulbar involvement automatically. This study focused on providing a set of 50 phonatory-subsystem and time-frequency features to detect this deficiency in males and females through the utterance of the five Spanish vowels. Multivariant Analysis of Variance was then used to select the statistically significant features, and the most common supervised classifications models were analyzed. A set of statistically significant features was obtained for males and females to capture this dysfunction. To date, the accuracy obtained (98.01% for females and 96.10% for males employing a random forest) outperformed the models in the literature. Adding time-frequency features to more classical phonatory-subsystem features increases the prediction capabilities of the machine-learning models for detecting bulbar involvement. Studying men and women separately gives greater success. The proposed method can be deployed in any kind of recording device (i.e., smartphone).

Automated detection of COVID-19 cough.

Easy detection of COVID-19 is a challenge. Quick biological tests do not give enough accuracy. Success in the fight against new outbreaks depends not only on the efficiency of the tests used, but also on the cost, time elapsed and the number of tests that can be done massively. Our proposal provides a solution to this challenge. The main objective is to design a freely available, quick and efficient methodology for the automatic detection of COVID-19 in raw audio files. Our proposal is based on automated extraction of time-frequency cough features and selection of the more significant ones to be used to diagnose COVID-19 using a supervised machine-learning algorithm. Random Forest has performed better than the other models analysed in this study. An accuracy close to 90% was obtained. This study demonstrates the feasibility of the automatic diagnose of COVID-19 from coughs, and its applicability to detecting new outbreaks.

Measuring heart rate variability by means of information entropies based on Choi-Williams distribution.

The Shannon entropy theory was applied to the Choi-Williams time-frequency distribution (CWD) of cardiac time series (RR series) in order to extract entropy information in both time and frequency domains. From this distribution, four indexes were defined: (1) instantaneous partial entropy; (2) spectral partial entropy; (3) instantaneous complete entropy; (4) spectral complete entropy. These indexes were used for analyzing the heart rate variability of ischemic cardiomyopathy patients (ICM) with different sudden cardiac death risk. The results have shown that the values of these indexes tend to decrease, with different proportion, when the severity of pathological condition increases. Statistical differences (p-value < 0.0005) of these indexes were found comparing low risk and high risk of cardiac death during night and between daytime and nighttime periods of ICM patients. Finally, these indexes have demonstrated to be useful tools to quantify the different complex components of the cardiac time series.

Filtering and thresholding the analytic signal envelope in order to improve peak and spike noise reduction in EEG signals.

To remove peak and spike artifacts in biological time series has represented a hard challenge in the last decades. Several methods have been implemented mainly based on adaptive filtering in order to solve this problem. This work presents an algorithm for removing peak and spike artifacts based on a threshold built on the analytic signal envelope. The algorithm was tested on simulated and real EEG signals that contain peak and spike artifacts with random amplitude and frequency occurrence. The performance of the filter was compared with commonly used adaptive filters. Three indexes were used for testing the performance of the filters: Correlation coefficient (ρ), mean of coherence function (C), and rate of absolute error (RAE). All these indexes were calculated between filtered signal and original signal without noise. It was found that the new proposed filter was able to reduce the amplitude of peak and spike artifacts with ρ>0.85, C>0.8, and RAE<0.5. These values were significantly better than the performance of LMS adaptive filter (ρ<0.85, C<0.6, and RAE>1).

Prediction of nociceptive responses during sedation by time-frequency representation.

The level of sedation in patients undergoing medical procedures evolves continuously, such as the effect of the anesthetic and analgesic agents is counteracted by pain stimuli. The monitors of depth of anesthesia, based on the analysis of the electroencephalogram (EEG), have been progressively introduced into the daily practice to provide additional information about the state of the patient. However, the quantification of analgesia still remains an open problem. The purpose of this work is to analyze the capability of prediction of nociceptive responses based on the time-frequency representation (TFR) of EEG signal. Functions of spectral entropy, instantaneous power and instantaneous frequency were calculated in order to predict the presence or absence of the nociceptive responses to different stimuli during sedation in endoscopy procedure. Values of prediction probability of Pk above 0.75 and percentages of sensitivity and specificity above 70% and 65% respectively were achieved combining TFR functions with bispectral index (BIS) and with concentrations of propofol (CeProp) and remifentanil (CeRemi).

Removal of peak and spike noise in EEG signals based on the analytic signal magnitude.

Peak and spike artifacts in time series represent a serious problem for signal analysis especially in biomedical field. From the last decades, different techniques have been used for their removal mainly based on adaptive filters. This work presents a new approach for removing peak and spike artifacts based on the analytic signal envelope, filtered with a low-pass filter. The proposed algorithm was tested on electroencephalogram signals containing peak and spike artifacts. Results showed that this method permitted to remove the peak and spike artifacts preserving both high correlation (ρ>0.9) and spectral coherence equation with the original signal.

Characterization of the cerebral activity by time-frequency representation of evoked EEG potentials.

Event-related brain potentials (ERPs) are the electrical response of the brain while performing a particular task. Methods traditionally used to study ERPs measure the amplitude and duration of the waveform in order to quantify the changes, being signal morphology dependent. However, the frequency characteristics of those events remain uncovered. The aim of this work was the study of new measures to characterize, by means of time-frequency representation (TFR) techniques, the ERPs recorded while subjects conducted a choice reaction time task (Ericksen flanker task) following the administration of different alprazolam doses. Several measures defined from energy, instantaneous frequency and group delay functions were obtained by means of TFR techniques applied to the Choi-Williams distribution (CWD) of EEG signals. These measures, which are signal morphology independent, were studied in four frequency bands, δ (0-4 Hz), θ (4-8 Hz), α (8-15 Hz), ß (15-30 Hz), and for certain time periods. Based on these measures, differences between ERPs were analyzed by comparing the different response types (successes or successfully corrected failures) of the subject performing the task, and comparing the applied drug doses. For each subject, the CWD of EEG signals was applied in two different ways: (a) all ERPs were averaged per channel, and then the CWD was applied; (b) the CWD was applied to each one of the ERPs. When the CWD was applied to each ERP, the energy measures in the δ, θ and ß bands, the instantaneous frequency measures in the α and ß bands, and the group delay measures in the δ, θ and α bands showed a statistically significant level p < 0.0005 in the analysis of the response type. Also, the energy measures in the θ and ß bands and the instantaneous frequency measures in the α band showed statistically significant differences (p < 0.0005) between placebo and low and high drug doses. In contrast, poor results were obtained when all epochs of each subject were averaged per channel. Finally, it was concluded that these results showed that the new proposed measures based on the energy offered a new and more robust way to characterize ERP signals.

Time-frequency characterization of mismatch negativity in nociceptive responses.

An efficient way to investigate the neural basis of nociceptive responses is the event-related brain potentials (ERPs). One component belonging to this family of ERPs is the mismatch negativity (MMN). It reflects pre-attentive detection of changes in the incoming stimulus by comparing the new stimulus with sensory memory traces. In this work, single trials of ERP taken from EEG signal recorded under thermal and electric stimulation were analyzed with time-frequency representation (TFR). The main objective of this work was to characterize responses to frequent and infrequent stimuli with TFR functions. Variables defined on instantaneous frequency and instantaneous power presented a statistical significance (p-value<0.0001) differentiating these two kind of responses. Furthermore, differences between the averaged instantaneous power and instantaneous frequency were also analyzed. It was found that instantaneous power and instantaneous frequency were able to better isolate the MMN components from EEG noise in certain frequency bands.

Exons and introns characterization in nucleic acid sequences by time-frequency analysis.

A current problem in deoxyribonucleic acid (DNA) sequence analysis is to determine the exact locations of the genes and also in eukaryotes, the protein-coding regions in the mRNA primary transcript (pre-mRNA).The conversion into discrete numerical values of the symbols associated to the nucleotides of these sequences allows for a signal to address the problems related to localization and annotation of genes. In this work, thermodynamic data of free energy changes (ΔG°) on the formation of a duplex structure of DNA or RNA are used to convert the symbols into numerical values associated with the nucleotide sequence pre-mRNA. This study presents an analysis, based on techniques of time-frequency representation of a large number of gene sequences, in order to find variables related to pre-mRNA that could best characterize and discriminate coding regions from non-coding regions. It has been found that instantaneous frequency variables and instantaneous spectral energy variables in different frequency bands, allowed exons and introns to be correctly classified with more than 85%.

Transcription factor binding site detection through position cross-mutual information variability analysis.

Regulatory sequence detection is a fundamental challenge in computational biology. One key process in protein synthesis starts with the binding of the transcription factor to its binding site. Different sites can show binding to the same factor. This variability found in binding sequences increases the difficulty of their detection using computational algorithms. In this manuscript, a method for the detection of binding sites is proposed, based on the correlation between binding sequence positions through information theoretical measures. Efficiency values of the method are reported in the form of Receiver Operating Characteristic curves on the detection of different transcription factors of the Saccharomyces cerevisiae organism. We compare our results with other known motif detection Motif Discovery scan (MDscan).

Refined multiscale entropy: application to 24-h Holter recordings of heart period variability in healthy and aortic stenosis subjects.

Multiscale entropy (MSE) was proposed to characterize complexity as a function of the time-scale factor tau. Despite its broad use, this technique suffers from two limitations: 1) the artificial MSE reduction due to the coarse graining procedure and 2) the introduction of spurious MSE oscillations due to the suboptimal procedure for the elimination of the fast temporal scales. We propose a refined MSE (RMSE), and we apply it to simulations and to 24-h Holter recordings of heart rate variability (HRV) obtained from healthy and aortic stenosis (AS) groups. The study showed that the refinement relevant to the elimination of the fast temporal scales was more helpful at short scales (spanning the range of short-term HRV oscillations), while that relevant to the procedure of coarse graining was more useful at large scales. In healthy subjects, during daytime, RMSE was smaller at short scales (i.e., tau = 1-2) and larger at longer scales (i.e., tau = 4-20) than during nighttime. In AS population, RMSE was smaller during daytime both at short and long time scales (i.e., tau = 1 -11) than during nighttime. RMSE was larger in healthy group than in AS population during both daytime (i.e., tau = 2 -9) and nighttime (i.e., tau = 2). RMSE overcomes two limitations of MSE and confirms the complementary information that can be derived by observing complexity as a function of the temporal scale.

Multiscale sample entropy in heart rate variability of aortic stenosis patients.

In the present document the multiscale entropy (MSE) methodology has been applied to analyze the complex behavior of the heart rate variability (HRV), in patients with aortic stenosis (AS). A set of healthy voluntaries have been used as a control group. MSE analysis calculates an entropy rate over different time scales to assess the complexity of time series, evaluating short-term and long-term correlations. Daytime and nighttime have been considered to study variations of the complexity inside the same group of population. A statistical analysis showed that entropy was significantly higher in healthy subjects than in AS subjects in all the scales during daytime, with exception at scale 1. During nighttime, entropy in healthy subjects was significantly higher than in AS subjects only in scales from 1 to 7. Multiscale entropy is helpful to characterize AS patients and distinguish them from healthy subjects.

DNA binding sites characterization by means of Rényi entropy measures on nucleotide transitions.

In this work, parametric information-theory measures for the characterization of binding sites in DNA are extended with the use of transitional probabilities on the sequence. We propose the use of parametric uncertainty measure such as Renyi entropies obtained from the transition probabilities for the study of the binding sites, in addition to nucleotide frequency based Renyi measures. Results are reported in this manuscript comparing transition frequencies (i.e. dinucelotides) and base frequencies for Shannon and parametric Renyi for a number of binding sites found in E. Coli, lambda and T7 organisms. We observe that, for the evaluated datasets, the information provided by both approaches is not redundant, as they evolve differently under increasing Renyi orders.