RU2313280C1 - Method for studying vocal cords functional state - Google Patents

Abstract

FIELD: medicine.

SUBSTANCE: method involves transforming voice signal of vocalized test phoneme into electric signal by means of acoustic transducer and recording it as discrete readings sequence on magnetic carrier. The so produced discrete sequence is reproduced as wavelet-plane segmented in a way that each segment represents spectral structure of the test phoneme. Characteristic attribute correlating to vocal cord state is selected in each interval. Its value is estimated by expert. Machalanobis distance from a point corresponding to attribute vector to class centroids is determined on the so built attribute space. The centroids have been defined a priory for given learning samples in the same attribute space. Vocal cords state is considered to belong to the class on which the Machalanobis distance reaches minimum.

EFFECT: higher specificity of functional state analysis.

4 dwg, 1 tbl

Description

The invention relates to medicine, namely to methods for qualitatively-quantitative analysis of the functional state of the vocal folds.

There is a method for assessing phonator vibrations of the vocal folds, in which, based on an expert ball score of five qualitative characteristics of the vibrations of the vocal folds, the index of vibratory disorders is determined, the magnitude of which makes a conclusion about the presence of voice disorders (see Using video laryngostroboscopy in phoniatric practice - a manual for a doctor. Moscow Scientific - Research Institute of the Ear, Throat and Nose. Moscow. 1997.22 p.).

However, this diagnostic method is not effective enough due to the integral nature of the index of vibratory disturbances, which does not allow to differentiate voice disturbances; the limited dynamic range of the expert assessment (from one to three points), which reduces both the sensitivity and the specificity of the method; the frequent lack of the ability to determine the index of vibratory disturbances on the basis of all five characteristics and the lack of complete identical conditions when setting up the experiment, which complicates the comparability of the type of voice disturbances and the numerical value of the indices of vibratory disturbances; and the lack of retrospective analysis.

Closest to the invention is a method of voice research, which determines the characteristics of the ability of the voice-reproducing apparatus for speech formation (see A.A. Skoromets, T.A. Skoromets "Topical Diagnosis of Nervous System Diseases. Polytechnic, St. Petersburg, 2000). p. 141). At the same time, devices are used to record the studied speech sounds, in which, through acoustic sensors, speech sounds are converted into electrical vibrations, which are then amplified and can be represented as recordings, for example, on magnetic tape or in the form of a curve on paper (see N.M Liventsev. "The course of physics." M.: Higher School, 1974, pp. 114-115).

The disadvantages of this method, as well as the previous one, are the low specificity and the requirement for a highly qualified expert to decide on the presence or absence of pathology of the vocal apparatus.

The objective of the invention is to increase the specificity of the analysis of the functional state of the vocal folds, which will reduce the treatment time and increase the time of remission.

To do this, in the known method of voice research, which includes converting a vocalized test phoneme signal into an electrical signal using an acoustic sensor and recording it onto a magnetic medium as a sequence of discrete samples, the resulting discrete sequence is represented as a wavelet plane that is segmented into segments, each of which reflects the spectral structure of the test phoneme, a characteristic feature is selected in each segment, which is correlated with the state of voice folds, its quantitative value is estimated by an expert, and in the attribute space thus obtained, the distances of the Mahalanobis from the point corresponding to the attribute vector to the centroid classes determined a priori for specific training samples in the same attribute space are determined; and the state of the vocal folds is assigned to the class for which the Mahalanobis distance is the smallest.

Figure 1 shows the structural diagram of an automated system designed to implement the proposed method. It includes a stereo headset 3, designed to connect to patient 2; analog interface 4, the inputs of which are connected to the output of the stereo headset 3; A computer 5 connected to the output of the analog interface 4; monitor 6 connected to the computer 5 and designed to implement the interaction between the doctor 1, patient 2 and the module for digitizing the signal of the computer 5.

Figure 2a shows a segmented Fourier spectrum of the voice signal (the phoneme And is selected as an illustration), and figure 2b shows a segmented wavelet transform of the same signal.

Figure 3 shows three segments A of the wavelet planes of the phoneme And: with a high degree of frequency regularity (X1 = 4,5) at the top; with an average degree (X1 = 2.5) in the center; with low frequency regularity (X1 = 1.5) below.

Figure 4a shows a fragment of the experimental data table for training samples of three classes, and figure 4b shows a fragment of a table with Mahalanobis distances obtained in STATISTICA 6 for training and control samples of the same classes.

The method is as follows.

A stereo headset 3 is connected to patient 2. After this, doctor 1 offers him to phonate one of the vowels, for example I. The commands for starting the phoning, phonation time, and setting the sampling frequency of the voice signal are set by doctor 1 through the computer interface module 5 and are the input parameters of the signal digitizing module, which included in computer software 5. The phonation time is 5 s. The sampling frequency of the voice signal is 11000 Hz. Feedback on the moment of the start of phonation and on the time of phonation is carried out between the doctor 1, patient 2 and the module for digitizing the signal of the computer 5 through the monitor 6 (see figure 1). As a result of this interaction, a voice signal with a length of 5 s is recorded in the memory of the computer 5.

Then, by visual analysis of the recorded signal using the interface module and the computer signal digitization module 5, from the obtained array of samples of the voice signal, a length of 0.15 s is selected that is least susceptible to artifacts and interference, and its wavelet transform is obtained. The indicated procedures can be performed in the Matlab 6.1 package (see Dyakonov VP Wavelets. From theory to practice. Solon. 2004. 448 pp.).

Figure 2, b shows the wavelet plane of the voice signal recorded with the above parameters during sound phonation AND by a woman with a healthy voice apparatus. To diagnose the state of the vocal apparatus along the wavelet plane, it is necessary to synthesize the feature space, for which the wavelet plane of FIG. 2, b is divided into five segments: A, B, C, D, and E, the dislocations of which are shown by the corresponding arrows. The principle of this division is determined by the spectral structure of the test phoneme.

Figure 2, a shows a typical Fourier spectrum of the voice signal of the phoneme And and its segments corresponding to the segments of the wavelet plane of figure 2, b. Segment A includes the high-frequency region of the wavelet plane — the overtone zone, segment B corresponds to the region of the second formant, segment C is the boundary region between the second and first formant, segment D is the region of the first formant, segment E is the low-frequency region, to which all wavelet plane frequencies below the fundamental frequency. The attribute space is synthesized using this segmentation as follows.

The first informative feature X1 characterizes the frequency regularity in time in segment A and is estimated using a five-point system based on expert judgment. Figure 3 shows three segments A for three patients with different regularity of wavelet coefficients: from 1.5 to 4.5.

The second attribute X2 characterizes the degree of variability of the values of the wavelet coefficients along the time and frequency axes in segments B and D corresponding to the dislocation zones of the second and first phoneme formants I. It is also evaluated using a five-point system based on expert judgment.

The third informative feature X3 is characterized by the ratio of energies in segments B and D and can approximately be calculated as the ratio of the areas of these segments.

The fourth and fifth informative features X4 and X5 characterize the degree of localization of formants and are determined by the distance (along the frequency axis) between the upper and lower boundaries of segments B and D, respectively. So that these signs do not depend on the parameters of the wavelet transform, it is advisable to normalize them over the frequency range that the wavelet plane occupies.

To obtain ranges of changes in these signs, we conducted statistical studies of the wavelet planes of patients in normal conditions and with characteristic pathologies. The table presents the results of a study for functional and organic dysphonias.

Table Wavelet descriptions of informative signs of the phoneme AND for some classes of diseases of the vocal tract Indicator Norm Functional dysphonia Organic dysphonia Inflammatory diseases of the larynx X1 4 ... 5 3 ... 4 1 ... 2 0 ... 1 X2 0 ... 1 1 ... 2 2 ... 3 3 ... 5 X3 0.35 ... 0.44 0,1 ... 0,15 0.01 ... 0.1 0 ... 0.1 X4 0.05 ... 0.1 0.04 ... 0.08 0.02 ... 0.04 0 ... 0.02 X5 0,1 ... 0,15 0,1 ... 0,12 0.12 ... 0.15 0,1 ... 0,15

Received wavelet descriptions of informative features of the class "norm" illustrates figure 2, b.

In figure 2, b we see that in the region of overtones there is a high regularity, which can be estimated 4 ... 5 points. The lowest X2 variability in the first formant (segment D), which can be rated 0 ... 1 points. The average ratio of the energies of segments B and D lies in the range 0.35 ... 0.44. The localization indicators of formants are 0.05 ... 0.1 and 0.1 ... 0.15.

To classify violations of the vocal apparatus in the selected feature space, we use discriminant analysis in the STATISTICA 6 package (see Borovikov V. STATISTICA. The Art of Computer Data Analysis: For Professionals. 2nd ed. (+ CD). - St. Petersburg: Peter, 2003 .688 p.).

For each set of objects of a certain class (sample), you can define a point whose coordinates represent the averages for all variables in the multidimensional space defined in the model under consideration. These points are called class centroids. For each observation, you can then calculate its Mahalanobis distance to each centroid (to each class). An observation is recognized as belonging to the class to which it is closer, i.e. for which the distance of Mahalanobis is minimal (see Borovikov V. STISTICA. The Art of Computer Data Analysis: For Professionals. 2nd ed. (+ CD). - St. Petersburg: Peter, 2003. 688 pp. CD, section “Discriminant Analysis” )

A fragment of a table with a training set in the STATISTICA 6 package is shown in Fig. 4, a. Figure 4, b shows a table with the distances of Mahalanobis for training and control samples.

In general, a patient is considered to belong to the class for which the Mahalanobis distance is minimal. However, the doctor, analyzing the numerical values of the distances of the Mahalanobis, can introduce his own criteria and assign the patient to another class with a close, but not minimum value of the distance of the Mahalanobis.

The proposed method has been tested on more than 300 patients with diseases of the vocal apparatus of varying severity and different ages.

Case studies

Example 1. Patient N., 12 years old, complains of rapid voice fatigue, a sensation of "coma" in the throat. The voice is quiet, weak, dull, with a long voice load hoarse.

Wavelet diagnosis: X1 = 2.5; X2 = 2; X3 = 0.1; X4 = 0.05; X5 = 0.12.

Computer diagnosis: functional dysphonia.

Indirect laryngoscopy: vocal folds of normal color. With phonation, their incomplete closure with the formation of a glottis of the oval shape is noted.

Microlaryngoscopy: a slight increase in the vascular pattern of the mucous membrane of the vocal folds.

Laryngostroboscopy: the amplitude of the oscillatory movements of the vocal folds is reduced, the phonation time is shortened.

Diagnosis: Hypotonic dysphonia.

Example 2. Patient C., 14 years old, worried about a feeling of discomfort in the larynx. The voice is hoarse, often breaking from the chest sound to falsetto.

Wavelet diagnosis: X1 = 3; X2 = 2; X3 = 0.12; X4 = 0.05; X5 = 0.11.

Computer diagnosis: functional dysphonia. Indirect laryngoscopy: hyperemia and swelling of the mucous membrane of the vocal folds. During phonation, their incomplete closure is noted with the formation of a mutation triangle in the interchaloid space.

Microlaryngoscopy: injection of small blood vessels of the laryngeal mucosa.

Laryngostroboscopy: asynchronous phonatory vibrations of the vocal folds.

Diagnosis: Mutational dysphonia.

Example 3. Patient E., 10 years old, complains of obsessive coughing and unpleasant, pain in the throat, larynx, neck, rapid voice fatigue. The voice is hoarse, it sounds intense.

Wavelet diagnosis: X1 = 3.5; X2 = 1.5; X3 = 0.12; X4 = 0.05; X5 = 0.11.

Computer diagnosis: functional dysphonia.

Indirect laryngoscopy: tight contact of the vocal folds during phonation, hyperemia and slight swelling of their free edge.

Laryngostroboscopy: the vocal folds look motionless, are in a tense closed state in combination with short-term periods of rapidly decaying oscillations of small amplitude.

Diagnosis: hypertonic dysphonia.

Example 4. Patient A., 9 years old, complains of a feeling of persistent perspiration, discomfort in the throat, cough. The voice is quiet, weak, more hoarse in the morning. In the afternoon, the voice becomes cleaner and coarser again in the evening. Wavelet diagnosis: X1 = 2.4; X2 = 2.4; X3 = 0.12; X4 = 0.05; X5 = 0.15.

Computer diagnosis: organic dysphonia is a stage of incomplete remission.

Indirect laryngoscopy: vocal folds are hyperemic, thickened, with phonation they do not close completely. The vestibular folds are compacted, hyperemic, swollen.

Microlaryngoscopy: injection of blood vessels of the vocal folds and accumulation of mucus on their surface.

Laryngostroboscopy: phonator oscillations of the vocal folds are asynchronous, weakened.

Diagnosis: Chronic laryngitis - stage of incomplete remission.

Example 5. Patient Yu., 7 years old, weak voice, hoarse. Sick for 1.5 months after suffering a severe form of an infectious disease (flu).

Wavelet diagnosis: X1 = 2.5; X2 = 1.5; X3 = 0.15; X4 = 0.1; X5 = 0.11.

Computer diagnosis: functional dysphonia.

Indirect laryngoscopy, microlaryngoscopy: the right vocal fold is in the paramedial position, the movement of the left half of the larynx is in full.

Laryngostroboscopy: vibrations of the vocal fold on the affected side, weakened, lethargic with small amplitude.

Diagnosis: Infectious paresis of the larynx.

Thus, choosing the appropriate test phoneme and determining the appropriate training samples, we obtain a decisive rule that allows us to divide patients according to the functional state of the vocal folds into classes corresponding to the training samples. Moreover, the proposed method allows both objective testing and an interactive mode in which the expert (doctor) analyzes the location of the research object (patient) in the space of signs and, when making a decision, is guided by the data of both computer diagnostics and the results of traditional studies.

Claims (1)

A method for studying the functional state of voice folds, including converting a vocalized test phoneme voice signal into an electrical signal using an acoustic sensor and recording it onto a magnetic medium in the form of a sequence of discrete samples, characterized in that the obtained discrete sequence is presented in the form of a wavelet plane that is segmented into segments, each of which reflects the spectral structure of the test phoneme, in each segment choose a characteristic feature, which correlated with the state of the vocal folds, the expert evaluates its value, and in the attribute space obtained in this way, the distances of the Mahalanobis from the point corresponding to the attribute vector to the centroid classes a priori determined for specific training samples in the same attribute space are determined, and the state of the vocal folds is referred to that class for which the distance of the Mahalanobis is the smallest.

Images