WO2016000061A1

WO2016000061A1 - Acoustic neck fluid volume assessment system and method

Info

Publication number: WO2016000061A1
Application number: PCT/CA2014/050627
Authority: WO
Inventors: Azadeh Yadollahi; Frank RUDZICZ
Original assignee: University Health Network
Priority date: 2014-07-02
Filing date: 2014-07-02
Publication date: 2016-01-07

Abstract

Described herein are various embodiments of an acoustic neck fluid volume assessment system and method. In one such embodiment, a neck fluid volume assessment device comprises a microphone to be positioned in an area of the subject so to acquire acoustic breath sounds emanating from the subject while breathing and generate a signal representative thereof, a digital storage device having stored thereon a neck fluid volume assessment engine having associated therewith one or more designated acoustic features previously identified to provide a measure of neck fluid volume, and a data processor operatively coupled to the digital storage device to implement the neck fluid volume assessment engine to act on the signal to automatically extract the one or more designated acoustic features therefrom and output an indication of the subject's neck fluid volume as a function of the extracted one or more features.

Description

ACOUSTIC NECK FLUID VOLUME ASSESSMENT SYSTEM AND METHOD

FIELD OF THE DISCLOSURE

[0001] The present disclosure relates to pharyngeal and/or airway assessment methods and systems, and, in particular, to an acoustic neck fluid volume assessment system and method. BACKGROUND

[0002] Obstructive sleep apnea (OSA) is a common disorder that increases cardiovascular morbidity and mortality. Although OSA occurs due to a partial or complete collapse of the upper airway during sleep, the underlying mechanisms of this collapse are not fully understood. Fluid accumulation in the neck could cause distension of the neck veins and/or edema of the peripharyngeal soft tissue, narrow the upper airway, and facilitate its obstruction. Especially in fluid-retaining patients, such as renal failure and heart failure patients, severity of OSA is strongly correlated with the amount of edema in the pharyngeal tissue. In this regard, while reducing total body water using compression stockings, diuretics and ultrafiltration could reduce OSA severity; fluid overloading could induce or worsen OSA severity. Therefore, developing convenient and non-invasive techniques to measure fluid accumulation in the neck could contribute to monitoring the effects of neck edema on the severity of sleep apnea and to modify and evaluate various treatments for reducing neck edema to prevent their adverse effects.

[0003] A sedentary lifestyle causes fluid retention in the legs during the day, which would be redistributed to the thorax and neck when lying down at night. Fluid redistribution into the neck could contribute to the OSA severity, as assessed by apnea-hypopnea index (AHI), by narrowing the upper airway, increasing its resistance, and collapsibility. In patients with end- stage renal disease, the only predictors of AHI were the mucosal water content in the pharynx and internal jugular vein volume as assessed by magnetic resonance imaging. Ultrafiltration reduces total body water, neck fluid volume, and OSA severity in patients with end stage renal disease. In otherwise healthy men with OSA and in patients with chronic venous insufficiency, wearing compression stockings for 1 day or 1 week respectively reduces overnight decreases in leg fluid volume and increases in neck circumference, in association with a 30% decrease in

1

1004P-ANF-WO01 AHI. Similarly, intensive diuretic therapy in patients with heart failure and severe OSA decreases AHI along with a decrease in total body water and overnight increase in neck circumference. On the other hand, in older men with mild to moderate OSA, fluid overloading during sleep with intravenous injection of normal saline causes a significant three-fold increase in AHI. These results provide strong evidence that fluid accumulation in the neck changes the properties of the pharynx and increases the risk of developing OSA.

[0004] Available methods for measuring neck fluid volume (NFV) are based on magnetic resonance imaging or detecting changes in neck circumference. However, these methods are either inconvenient, expensive, subjective, sensitive to movement, or are technically difficult to perform.

[0005] This background information is provided to reveal information believed by the applicant to be of possible relevance. No admission is necessarily intended, nor should be construed, that any of the preceding information constitutes prior art.

SUMMARY [0006] The following presents a simplified summary of the general inventive concept(s) described herein to provide a basic understanding of some aspects of the invention. This summary is not an extensive overview of the invention. It is not intended to restrict key or critical elements of the invention or to delineate the scope of the invention beyond that which is explicitly or implicitly described by the following description and claims. [0007] A need exists for a new neck fluid volume assessment system and method that overcome some of the drawbacks of known techniques, or at least, provides a useful alternative thereto.

[0008] In accordance with one broad aspect of the present disclosure, there is provided such a system and method involving the performance of acoustic assessments, for instance, to assess the amount of fluid in the neck.

[0009] In accordance with one aspect, there is provided a neck fluid volume assessment device for use with a subject while breathing, the device comprising: a microphone to be

2

1004P-ANF-WO01 positioned in an area of the subject so to acquire acoustic breath sounds emanating from the subject and generate a signal representative thereof; a digital storage device having stored thereon a neck fluid volume assessment engine having associated therewith one or more designated acoustic features previously identified to provide a measure of neck fluid volume; and a data processor operatively coupled to said digital storage device to implement said neck fluid volume assessment engine to act on said signal to automatically extract said one or more designated acoustic features therefrom and output an indication of the subject's neck fluid volume as a function of said extracted one or more features.

[0010] In accordance with another embodiment, there is provided a non-invasive neck fluid volume assessment method to be performed on a subject, the method comprising: receiving as input a signal representative of breath sounds generated by the subject; extracting one or more designated acoustic features from said input signal, wherein said one or more designated acoustic features define one or more preset neck fluid volume assessment metrics; comparing said one or more extracted features with said one or more preset neck fluid volume assessment metrics; and outputting, based on said comparing, characterization of the subject's neck fluid volume as a function of said one or more metrics.

[0011] In accordance with another embodiment, there is provided a computer-readable medium having statements and instructions stored thereon for implementation by a processor to act on a signal representative of acoustic breath sounds emanated by a subject in outputting an assessment of the subject's neck fluid volume by performing the steps of the above method.

[0012] In accordance with another embodiment, there is provided a non-invasive method for assessing a neck fluid volume in a subject, the method comprising: acquiring acoustic breath sounds emanating from the subject over a time period; generating a data signal representative of said acquired acoustic breath sounds; extracting one or more designated acoustic features from said data signal, wherein said one or more designated acoustic features define, alone or in combination, a neck fluid volume metric; comparing said one or more extracted features with said neck fluid volume metric; and outputting, based on said comparing, an estimated neck fluid volume in the subject during said time period.

3

1004P-ANF-WO01 [0013] In accordance with another embodiment, there is provided a method of manufacturing a neck fluid volume assessment device, the method comprising: acquiring acoustic breath sounds from multiple subjects to generate respective data signals representative thereof; concurrently measuring respective neck fluid volumes in said subjects during said acquiring; extracting one or more designated acoustic features from each of said signals; classifying said one or more extracted features, alone or in combination, for each of said signals, to correspond with an estimated neck fluid volume based on said concurrently measured neck fluid volumes; defining a neck fluid volume metric based on said classifying to be applied to said one or more designated features once extracted from input breath sound signals to output a corresponding neck fluid volume assessment; and programming a computing device with said defined metric so to act on new input breath sound signals to: extract said one or more designated acoustic features therefrom; compare said one or more extracted features with said neck fluid volume metric; and output, based on said comparing, an estimated neck fluid volume indication. [0014] Other aspects, features and/or advantages will become more apparent upon reading of the following non-restrictive description of specific embodiments, given by way of example only with reference to the accompanying drawings.

BRIEF DESCRIPTION OF THE FIGURES

[0015] Several embodiments of the present disclosure will be provided, by way of examples only, with reference to the appended drawings, wherein:

[0016] Figure 1 is a schematic diagram of a neck fluid volume (NFV) assessment system, in accordance with one embodiment of the invention;

[0017] Figure 2 is a schematic diagram of a NFV assessment device, and components thereof, in accordance with one embodiment of the invention; [0018] Figure 3 is a diagram of an experimental setup for recording neck bioelectrical impedance and measuring NFV, in accordance with one embodiment;

4

1004P-ANF-WO01 [0019] Figure 4 is a chart illustrating an average and standard error of FV among all subjects in different time periods, in accordance with one example (**: P<0.001);

[0020] Figures 5A and 5B are charts of average and standard error of changes in select acoustic features, namely Mel-Frequency Cepstral Coefficients (MFCC) and Mel-Power frequencies acoustic features, respectively, after 90 minutes from baseline {%.p < 0.1, *: ><0.05, and % / 0.01), in accordance with one embodiment;

[0021] Figure 6 is a chart of changes in 1^st and 2^nd formants after 90 minutes in different individuals, in accordance with one embodiment;

[0022] Figure 7 is a chart of recorded and estimated NFV measures based on a regression method implemented in accordance with one embodiment;

[0023] Figure 8 is a flow chart of an exemplary acoustic neck fluid assessment method, in accordance with one embodiment; and

[0024] Figure 9 is a flow chart of a more detailed acoustic neck fluid assessment method, in accordance with one embodiment.

DETAILED DESCRIPTION

[0025] In accordance with some aspects of the herein-described embodiments, an acoustic neck fluid volume assessment system and method will now be described. In particular, the systems and methods considered herein rely on acoustic variations observed in relation to the amount of fluid in the neck, for example. For example, the methods and systems described herein can be used to accurately and non-invasively assess an increase in a candidate's neck fluid, which increase constricts the airway and can be correlated with OSA in some instances, by identifying acoustic changes in breathing and/or snoring sounds resulting therefrom.

[0026] For instance, tracheal sound analysis, in the context of the below-described embodiments, can provide an effective and non-invasive way to investigate variations in the pathophysiology of the airways and monitor upper airway obstruction during both wakefulness and sleep. Different mechanisms including turbulence of respiratory airflow and pressure fluctuations in the pharynx can contribute to the generation of tracheal sounds. In some

5

1004P-ANF-WO01 embodiments, the vibrations so generated are transmitted to the skin through the tracheal wall and tissue beneath the skin, and can be picked up by a microphone placed over the trachea, for example, but also for example via a microphone mounted to or in the ear, the cheek, a face mask disposed above a nose and mouth area of the subject's face, or again, but subject to greater ambient noise, freestanding, mounted or positioned in a room near the subject. For example, ambient and other noise may be reduced upon positioning the microphone in skin- contact with the subject, for example in a throat, cheek or ear area. As demonstrated in the below examples, normal tracheal sound have been shown to exhibit a broad-band spectrum, its shape and spectral peaks changing with the geometry and pathology of the upper airway. Accordingly, in accordance with some aspects, the system and methods described herein allow for the estimation of fluid volume in a biophysical airway based on acoustics.

[0027] With reference now to Figure 1, and in accordance with one embodiment, a system for assessing neck fluid volume, generally referred to using the numeral 100, will now be described. In this example, the system 100 generally comprises a microphone 102 or the like to be attached on the surface of a throat area of a candidate for acquiring acoustic sounds and/or signals over time. The microphone 102 is operatively coupled to a data processing device 104 having stored and implemented thereon one or more neck fluid volume assessment tools/engines to automatically process the acquired data according to one or more designated assessment protocols for output. While the data processing device 104 is illustrated in Figure 1 as distinct from the microphone/recording device 102, in some embodiments, the microphone 102 and data processing device 104 may be integral to or combined in a common data recording device to be worn on the subject's neck area, for example. While the term "data processing device" is used genetically herein to refer not only to a device for performing automated or semi-automated acoustic neck fluid volume assessments, it may also refer to similar devices also configured for the detection or assessment of other more or less related conditions, symptoms, and/or biological processes.

[0028] The processing device 104 is depicted herein as a distinctly implemented device operatively coupled to microphone 102 for communication of data thereto, for example, via one or more data communication media such as wires, cables, optical fibres, and the like, and/or one or more wireless data transfer protocols, as would be readily appreciated by one of

1004P-ANF-WO01 ordinary skill in the art. The processing device may, however, in accordance with other embodiments, be implemented integrally with a recording device embodying the microphone (e.g. within the context of a self-contained assessment tool or device that can be secured to or on the subject's body during data acquisition and processing), for example, depending on the intended practicality of the system 100, and/or context within which it is to be implemented. As will be appreciated by the skilled artisan, the processing device 104 may further or alternatively be coupled to, or operated in conjunction with, an external processing and/or interfacing device, such as a local or remote computing device or platform provided for the further processing and/or display of raw and/or processed data, or again for the interactive display of system implementation data, protocols and/or screening/assessment tools.

[0029] With reference to Figure 2, the processing device, depicted herein generically as a self-contained device 200, generally comprises a power supply 202, such as a battery or other known power source, and various input/output port(s) 204 for the transfer of data, commands, instructions and the like with interactive and/or peripheral devices and/or components (not shown), such as for example, a distinctly operated microphone and/or acoustic data recorder, external data processing device, display or the like. The device 200 further comprises one or more computer-readable media 208 having stored thereon statements and instructions for implementation by one or more processors 206 in automatically implementing various computational tasks with respect to, for example, acoustic data acquisition and processing 210, operation of the device 212 (e.g. one or more clinically accepted operation protocols, testing and/or validation sequences, etc.), or again in the implementation of one or more acoustic assessment tools/engines (e.g. NFV engine 214) implemented on or in conjunction with the device 200. The device 200 may further comprise a user interface 216, either integral thereto, or distinctly and/or remotely operated therefrom for the input of data and/or commands (e.g. keyboard, mouse, scroll pad, touch screen, push-buttons, switches, etc.) by an operator thereof, and/or for the presentation of raw, processed and/or assessment data outputs (e.g. graphical user interface such as CRT, LCD, LED screen, touchscreen, or the like, visual and/or audible signals/alerts/warnings/cues, numerical displays, etc.)

[0030] As will be appreciated by those of ordinary skill in the art, additional and/or alternative components operable in conjunction and/or in parallel with the above-described

7

1004P-ANF-WO01 illustrative embodiment of device 200 may be considered herein without departing from the general scope and nature of the present disclosure. It will further be appreciated that device 200 may equally be implemented as a distinct and dedicated device, such as a dedicated home, clinical or bedside assessment device, or again implemented by a multi-purpose device, such as a multi-purpose clinical or bedside device, or again as an application operating on a conventional computing device, such as a laptop or PC, or other personal computing devices such as a PDA, smartphone, tablet or the like.

[0031] In the illustrative example of Figure 2, the stored statements and instructions of computer-readable medium 208 encompass one or more acoustic FV assessment tools/engines 214 that, when launched via processor 206, act on acquired acoustic data to output one or more assessments useful in characterizing an amount of fluid in the subject's neck, for example.

[0032] In accordance with some embodiments, the assessment tool/engine 214 may be configured to receive as input (e.g. via input port 204) acoustic data of interest acquired, for example, via a recording device and/or microphone, such as microphone 102 of Figure 1. In some embodiments, the engine will include an optional preprocessing utility, a feature extraction utility, an estimation utility, and one or more optional post-processing utilities, the later generating a global or respective outputs to be rendered or otherwise provided via the system's input/output port 204 and/or user interface 216. [0033] For example, and with reference to Figure 8, a neck fluid volume assessment method 800 may initiate from recorded breath sounds, indicated herein as breath sound signal 802, be they prerecorded, stored and/or recorded in real or quasi real-time fashion. The breath sound signal 802 is processed via a predefined feature extraction process 804 which takes as input both the (optionally preprocessed) sound signal 802 and a set of designated acoustic features 806 previously identified to accurately characterize (i.e. distinctly identify or quantify) neck fluid volumes, for example. The extracted features output from step 806 are then processed through a predefined feature characterization process 808 that takes as input a designated neck fluid volume metric 810 predefined as a function of the designated features 806 so to output a neck fluid volume characterization 812.

8

1004P-ANF-WO01 [0034] With reference to Figure 9, a more detailed neck fluid volume assessment method 900 may again initiate from a breath sound signal 902, in this embodiment, that is optionally preprocessed via a noise reduction algorithm at step 904. The (preprocessed) signal is then spliced into distinct time segments at step 906 so to allow for an assessment of neck fluid volume over time. Each time-segmented signal is then optionally transformed via time and/or frequency transformation utilities at step 908, to have time, frequency and/or time-frequency domain features respectively extracted therefrom at step 910 based on input designated acoustic features 912. Once the designated features extracted, they are input for parallel or joint characterization at step 914 based on one or more input neck fluid volume metrics 916 predefined as a function of the designated features 912. Output characterization(s) are then optionally post-processed (e.g. normalized) at step 918 to output a time-evolving (i.e. time- dependent) neck fluid characterization at step 920. It will be appreciated by the skilled artisan that the above presents illustrative process steps and sequences in the output of neck fluid characterizations from acquired breath sounds, and that other steps and/or sequences may also or alternatively be considered to provide similar results without departing from the general scope and nature of the present disclosure.

[0035] The input device is operable to obtain an acoustic signal that is to be used for estimation, and may comprise a microphone as noted above, or another sound source, or again may include an input communicatively linked to a microphone or other sound source, for example. For example, a sound source could be a sound file stored on a memory or an output of a sound producing device, and used as an alternative to direct acoustic sound recording of pharyngeal/airway sounds. In yet another example, the sound may be a pre-recorded sound that is synthesized to resemble a natural sound. For example, a simulation computer may be programmed to output a particular sound that resembles bodily sounds. The fluid estimation engine may be applied to the outputted sound for the purposes of model simulation.

[0036] In some embodiments, the optional preprocessing utility may apply noise reduction on the acoustic signal by applying a noise reduction algorithm, such as spectral subtraction, for example. The one or more feature extraction utilities may then measure specific attributes of the acoustic signal, designated in accordance with the methods described herein, to produce quantifiable results reliably indicative of fluid volume. From these extracted features, the

1004P-ANF-WO01 estimation utility may then estimate the amount of fluid in the airway that can be post- processed or normalized for output, such as in the form of a stored output on a computer- readable memory or device, a readout such as on screen or display, and the like.

[0037] As noted above, the engine(s) may be implemented by a computerized device, such as a desktop computer, laptop computer, tablet, mobile device, or other device having one or more computer processors and a memory having stored thereon statements and instructions which, when executed by the one or more computer processors, provide the functionality described herein.

[0038] The engine(s) may be embodied in a single-use device or in respective single use devices. The device could, for example, be a handheld computerized device comprising a microphone as the input device, a screen or speaker as the output device, and one or more processors, controllers and/or electric circuitry implementing, for example, one or more of a signal splicing utility, a time transformation utility and a frequency transformation utility, for example, or may otherwise be implemented within a more general device, such as depicted in Figure 2, or again within the context of a general purpose computer.

[0039] One particular example of such a device is a mobile device whose input device is pressed against the neck or airway under consideration. Another example of such a device is an implantable or wearable device (for example, worn around the neck). Another example of such a device is a microphone connected to a stationary computational device in which the estimation occurs.

[0040] The estimation engine(s) may be applied to different sounds represented by an acoustic signal. In one example, the sound may be the breathing or snoring of an individual. For example, the engine(s) may be applied jointly or independently to the breathing of an individual with OSA in order to estimate the amount of fluid in their neck, for example. [0041] As will be demonstrated by the below example, the processing of acoustic sound data in accordance with the herein-described embodiments allows for a characterization of a subject's neck fluid volume, and this, despite significant variations in neck tissue

10

1004P-ANF-WO01 compositions (fat, muscle, etc.), variations between subjects, and the general dynamics of the neck as compared to other more static or uniform portions of the body.

EXAMPLES:

[0042] The following provides different examples, in accordance with some aspects of the above-described embodiments, of an acoustic neck fluid volume assessment method and system. It will be appreciated by the skilled artisan that this example is not intended to limit the general scope and nature of the present disclosure, but rather provide further evidence as to the utility, applicability and/or accuracy of the methods and system described herein in accordance with different embodiments of the invention. Example 1

[0043] The following example, provided in accordance with one embodiment, focuses on the estimation of fluid in the neck, particularly during breathing.

[0044] In order to estimate neck fluid, the estimation engine makes use of a model based on relevant acoustic measurements and estimates of fluid. [0045] Bioelectrical impedance is a non-invasive technique to estimate fluid volume of tissues. Accordingly, and in one example, the estimation engine is modeled using bioelectrical impedance measures indicative of quantifiable fluid volumes, taken in parallel with acoustic measures to be modeled; bioelectrical impedance measures may include, but are not limited to, single frequency methods to measure extracellular fluid, multi -frequency methods that sweep across a range of frequencies, and bio-impedance tomography in which several electrodes are placed around a relevant body part and activated in succession.

[0046] An accurate model of neck fluid volume (NFV) which considers the length, L, and circumference, C, of the segment in the model, given resistivity p and neck resistance R_e may be defined as

1 1

1004P-ANF-WO01 [0047] In another example, an administrator or user of the fluid estimation engine could populate a sample database with a set of sound recordings and bioelectrical impedance measurements.

[0048] In yet another example, an administrator or user of the fluid estimation engine could populate a sample database with a set of sound recordings and fluid measurements based on different imaging modalities such as MRI of the neck, for example.

[0049] Tracheal respiratory sounds can be recorded by a microphone (for example, the Sony ECM-44B omni-directional microphone embedded in a chamber) and can be attached to the suprasternal notch of the subject. Tracheal sounds can be low-pass filtered with a cut-off frequency of 5 kHz using an, for example, the Biopac DA100C amplifier). Both FV and tracheal sounds can be digitized and recorded simultaneously with a given sampling rate (for example, 12.5 kHz).

[0050] Tracheal sounds can be band-pass filtered, for example in the frequency range of [30-3000] Hz to remove low- and high-frequency noise, including motion artefacts and measurement noi se respectively .

[0051] For each designated time period, for example each inspiratory breath cycle, several features in the temporal and spectral domains can be extracted from the sound signal. These may include, but are not limited to features such as total duration, average energy, skewness, kurtosis, the ratio between vocalized and unvocalized segments of breath sound lengths, recurrence features such as recurrence period density entropy, and zero crossing rate.

[0052] In the spectral domain, pitch frequency can be extracted, for example using the robust-adaptive pitch tracking algorithm, and one or more of the first four formants can be estimated in overlapping or non-overlapping windows, for example Hamming windows of 10 ms. Pitch and formant frequencies can be calculated using analysis of linear prediction coefficients, for example, with frequencies above 90 Hz and bandwidths below 400 Hz. Average power of breath sound may also be calculated, including in specific frequency bands, including Mel-frequency bands. Further processing of the spectrum may be performed using cepstral analysis, for example.

12

1004P-ANF-WO01 [0053] The features extracted from the acoustic signal can be reduced or transformed in a number of ways. For example, principal component analysis or independent components analysis may be performed to transform the available data into a smaller dimensionality. Specific features may also be isolated from others, for example, by forward selection, minimum-redundancy-maximum-relevance or another statistics-based systems.

[0054] Given a designated set of features (whether original, transformed, or selected), the estimation engine outputs an estimate of fluid, for example, fluid in the neck, in an established measurement scheme, such as millilitres. The output may be derived by a neural network, Bayes network inference, or regression, for example.

[0055] This output may be further post-processed, for example, by z-score normalization oorr ffiilltteerriinngg.. FFoorr eexxaammppllee,, tthhee ffiilltteering utility may apply a 10 -order low-pass Butterworth filter whose magnitude response is

ISO; 10)|² = |H(z; 10) |² =

where z is the complex frequency in polar coordinates and z^* _Norm is the cutoff frequency in that domain. This provides the transfer function

1

B(z; 10) = H(z; 10) =

1 + z¹⁰ +∑^ _Ciz whose poles occur at known symmetric intervals around the unit complex-domain circle. These poles may then be transformed by a function that produces the state-space coefficients a, and βί that describe the output signal resulting from applying the low-pass Butterworth filter to the discrete signal xfnj. These coefficients may further be converted by

→ * ^→—1 b =—z_Norm( ¹ ?)

13

1004P-ANF-WO01 giving the high-pass Butterworth filter with the same cutoff frequency of z _Norm. This continuous system may be converted to a discrete equivalent thereof using an impulse- invariant discretization method, which may be provided by the difference equation

[0056] The estimated fluid value may be output using the output device, saved onto a storage device, or transmitted over a transmission line.

[0057] The results outlined below describe one particular example in which acoustic data was demonstrated to accurately and reliably estimate neck fluid volume in candidates and thus provide a useful tool, for instance, within the context of diagnosing and/or characterizing breathing conditions such as OSA. It will be appreciated that this study is provided as an example only and is not intended to limit the scope of the application, but rather, serve to demonstrate the utility and practicality of the methods and systems described herein.

Subjects

[0058] In this study, inclusion criteria admitted healthy men between 18 and 70 years of age and healthy women more than 18 years of age who were premenopausal and did not have their menstrual cycle at the time of experiments, with a body mass index (BMI) < 30 kg/m², and a blood pressure of < 140/90 mmHg. The exclusion criteria were a history of hysterectomy, having metal implants, cardiovascular, renal, neurological or respiratory diseases, taking any medication for them, or taking any over the counter medication that might influence fluid retention.

Data recording

[0059] Subjects lay down in the supine position for 90 minutes without a pillow while awake. FV and tracheal respiratory sounds were recorded simultaneously. FV was recorded according to bioelectrical impedance, which is a non-invasive, well validated, and highly reproducible technique to measure fluid volume within tissues. In this method, two electrodes inject high frequency (50 kHz), low amplitude (400 μΑ) current into the tissues and

14

1004P-ANF-WO01 two sensing electrodes measure bioelectrical impedance, which is inversely related to the amount of fluid in the tissue. For FV, and with reference to Figure 3, sensing electrodes (V+, V-) were placed on the right side of the neck, one below the right ear and one at the base of the neck, whereas injecting electrodes (I+, I-) were placed one inch away from the sensing electrodes. Bioelectrical impedance is inversely related to the fluid content of each segment and can be estimated as defined above by equation (1). At the beginning of the study, neck length and circumference were measured with a measuring tape.

[0060] Tracheal respiratory sounds were recorded by a Sony ECM-44B omni-directional microphone embedded in a chamber (diameter of 6 mm) and attached to the suprasternal notch of the subject with double-sided tape. Tracheal sounds were low-pass filtered with a cut-off frequency of 5 kHz. Both NFV and tracheal sounds were digitized and recorded simultaneously with a sampling rate of 12.5 kHz (MP150, Biopac Systems).

Feature Extraction

[0061] Tracheal sounds were band-pass filtered in the frequency range of [30-3000] Hz to remove low- and high-frequency noise, including motion artifacts. For every subject, 4 periods of data between 0-10 minutes (Period 1), 20-30 minutes (Period 2), 50-60 minutes (Period 3), and 80-90 minutes (Period 4) were selected by an expert annotator and the inspiratory breath cycles without noise artifacts were marked manually. For each inspiratory breath cycle, several features in the temporal and spectral domains were extracted. [0062] Temporal features included duration, average energy, skewness and kurtosis of amplitudes over time, the ratio between voiced and unvoiced segments of breath sound lengths, recurrence period density entropy (RPDE), and zero crossing rate (normalized by the duration) of the inspiratory breath cycle.

[0063] In the spectral domain, pitch frequency (extracted using the RAPT algorithm) and the first three formants were estimated in non-overlapping Hamming windows of 10 ms. Pitch and formant frequencies were calculated using analysis of linear prediction coefficients with frequencies above 90 Hz and bandwidths below 400 Hz; their average and standard deviation were calculated for 'voiced' segments within each inspiratory cycle. Average power of breath

15

1004P-ANF-WO01 sound was calculated in the following bands: [30 - 100], [100 - 450], [450 - 600], [600 - 800], [800 - 1200], [1200 - 2000], and [2000 - 3000]. For each inspiratory cycle, the power of breath sounds was also calculated over 19 Mel -frequency bands, and the first 12 Mel -frequency cepstral coefficients (MFCCs) were also extracted. Estimation of NFV

[0064] The changes in NFV from the beginning to the end of the study and also between the four time periods (Period 1, Period 2, Period 3, and Period 4) were compared with paired Student's t-test. Repeated two-way analysis of variance (ANOVA) was used to investigate which acoustic features changed significantly over time. This was considered as the first level of screening to reduce the number of features.

[0065] To estimate NFV, k-fo\d (k=\0) cross-validation was used to divide the data of every individual into training and test sets. Two methods were developed to select features and derive fluid estimation algorithms from the training data of every individual. In the first method (regression method), stepwise regression selects features and derives the model parameters for fluid estimation. Here, features are iteratively added and removed from the input set based on a combination of the t-test and root mean squared error fitting. This method assumes no features are part of an initial set, features are added only if their associated p-va\ue is below 0.05, and removed only if their associated p-va\ue is above 0.10. In the regression- based method, fluid estimation parameters are set along with the feature selection step to achieve the minimum root mean square error between the estimated fluid and measured fluid volume.

[0066] In the second method (neural network method), a novel algorithm was developed for feature selection based on minimum-redundancy-maximum-relevance (mRMR). This algorithm produces a feature set which simultaneously maximizes the strength of the statistical relationship between the estimated variable (i.e., fluid volume) and the predictor variables while minimizing the strengths of the statistical relationships among those predictor variables. The mRMR algorithm chooses a feature set S^* consisting of N features f_; from among F features in total, given class c, using Pearson's correlation coefficient r(-,-) such that ||S^* II = JV and

16

1004P-ANF-WO01

[0067] One drawback of mRMR is that it does not consider multi-feature interaction - each feature is compared with the class individually and to reduce redundancy only pairwise comparisons between features are made. In this example, an additional empirical parameter, λ G R^{[a i]}, is added that balances class-feature relevance against feature-feature redundancy. This approach is used to minimize the difference between vectors of correlation coefficients for the set of features and the class while maximizing the average difference of those vectors among the features. Given the correlation table #₍F+I₎X₍F+ I₎ where

a matrix E of Euclidean dist nces is calculated between rows in /?, i.e..

[0068] The matrix D provides a similarity measure between two variables in terms of their overall similarity with all other variables in the system. Optimization then becomes a matter of finding the N values of i that minimize

[0069] In addition to optimizing over vectors of similarities, rather than individual similarities as in mRMR, this method can be computed in 0(F²) time, without the need for iterative 'hill-climbing'. Both mRMR and this method can replace Pearson's correlation coefficient with other measures of statistical similarity, including mutual information.

17

1004P-ANF-WO01 [0070] After selecting feature sets, fluid volume was estimated based on a mixture density neural network consisting of a single output Gaussian, an input vector of the N selected features for the given frame and δ optional frames of context before and after the current frame. The network uses one hidden layer with |ίίίί≤≤±!2±22| _unit_s and up to 400 training epochs. Given the input (with optional context), the output Gaussian represents a distribution over the estimated NFV, with the centroid of that Gaussian taken as the estimate.

[0071] Both methods of feature selection and fluid estimation were evaluated on each test set and the same routine was repeated for all k folds. Finally, based on the results of all folds, a score was assigned to each feature based on its significance for each selection, number of selections in all folds, and the root mean square error of NFV estimation when the feature was selected. For both methods and every individual, the top 10 features were selected as the optimum feature set and the error of NFV estimation was recalculated. Since the baseline NFV is highly variable among subjects, the relative error is calculated as:

E[NFV_meas - NFV_est]²

Relative Error = ^ r * , , > x l00

E[NFV_meas\² where E[.] is expected values, NFVmeas is the measured NFV, and NFVest is the estimated NFV. The performance of two NFV estimation methods among all subjects was also compared by Student's paired t-test.

Results

[0072] The study included 28 subjects (13 men, 15 women) aged 36.6±10.8 years, with a mean BMI of 23.5±3.6. Figure 4 shows the average and standard deviation of absolute values of NFV among all subjects. NFV increased progressively and significantly in all subjects from baseline to 90 minutes (p < 0.001). It has been shown that relative to baseline (just after lying supine), the changes in NFV follow an exponential model over time. Since the change in NFV over time is smaller than the baseline amplitude of NFV, such exponential changes in NFV are not visible in Figure 4. However, since the main objective was to estimate absolute values of NFV and not the changes in NFV, absolute values of NFV were demonstrated at various time segments.

18

1004P-ANF-WO01 [0073] Among all acoustic features, the average power of tracheal sounds in the frequency range of [2000-3000] Hz, Mel-power frequencies at seven bands above 1000HZ (Figure 5), the 2^nd and 6^th MFCC frequencies (Figure 5), and RPDE changed significantly for all subjects after 90 minutes relative to baseline (p < 0.01). The change in RPDE is expected to some extent, since we expect the increase in NFV to coincide with increased turbulence, hence higher measures of aperiodicity. The significance of 2^nd and 6^th MFC coefficients is more surprising. The lower-order cepstral coefficients tend to refer to the upper vocal tract or 'filter' of the anatomical source-filter model. A pronounced decrease in the amplitude here implies a decreased filter function nearer the microphone, which in this case corresponds to a greater attenuation of the sound in the neck.

[0074] It was found in this study that after 90 minutes, the total energy of tracheal sounds did not change significantly. Thus the decrease in higher frequency components of tracheal sounds may show a tendency in the tracheal sound power to move towards lower frequencies. In this regard, Figure 6 shows the variations in the first and second formant frequencies of every individual after 90 minutes. In some subjects there was a decrease in the first or the second formants, which complies with the shift of the power spectrum to the lower frequencies as observed in Figure 5. However as presented in Figure 6, this pattern was not consistent in all subjects. This may be due to the differences among subjects in the upper airway (UA) dilator muscle activities or reflexes. [0075] Among those features selected for each subject, the 10 most frequent features across all subjects were identified. These features (Table 1) can be considered as a globally optimum set of features based on each method for estimating NFV. For both methods, MFCC was selected.

19

1004P-ANF-WO01 Regression Neural Networks

Duration Formant 2

MFCC2 Formant 3

MFCC3 MFCC1

MFCC5 MFCC2 MFCC3

Selected

Features MFCC 10 MFCC4

MFCC 11 MFCC10

MFCC 12 Average Power at [100-450]Hz

Mel-Power atl 380Hz Average Power at [450-600]Hz

Mel-Power atl550Hz Average Power at [600-800]Hz

Table 1: Ten most frequently selected features among all subjects. (MFCC: Mel- frequency cepstral coefficient)

[0076] Figure 7 shows an example of recorded and estimated NFV based on a regression model for a typical subject. The results show that while the absolute values of recorded NFV changed from 199 millilitres (ml) to 208ml, the error of estimating NFV based on regression model was less than ±lml. The average and standard deviation of absolute and relative errors over all subjects for each method is shown in Table 2. Both methods achieved high accuracy in estimating NFV from the selected acoustic features for each subject. However, compared to the neural network method, the absolute and relative errors were significantly smaller when

20

1004P-ANF-WO01 using the regression method (t(27) = -11.13, p < 0.0001, and t(27) = -10.02, p < 0.0001, respectively). The absolute and relative errors in estimating NFV based on the regression method were 3.20±1.71ml and 1.25±0.70%, respectively. The absolute and relative errors for estimating NFV with neural network method were 8.14±2.38ml and 3.23±1.00%, respectively.

Method Absolute Error, ml Relative Error, %

Regression 3.20 ± 1.71 1.25 ± 0.70

Neural Networks 8.14 ± 2.38 3.23 ± 1.00

Table 2: Average and standard deviation of errors for each method for estimating NFV.

Discussion

[0077] As demonstrated by the above results, the herein-described methods and systems provide a reliable approach to estimating absolute values of NFV and their variations over time from the analysis of tracheal sounds. It was demonstrated that tracheal sound features change over time in association with fluid accumulation in the neck. Two models were developed for NFV estimation, one based on linear regression and the other on feed-forward neural networks. Results show that tracheal sound features can be used to estimate absolute values of NFV in healthy awake subjects with relative accuracies of more than 98%. These results indicate that tracheal sound analysis can be used to develop reliable and non-invasive techniques to estimate NFV, which estimates can be used in the pathophysiology of OSA, for example.

[0078] Accordingly, the herein-described methods and system allowed for the investigation of physiological factors that may contribute to the changes in tracheal sound features, and the analysis of the relative utility of different acoustic features in the estimation of NFV. Acoustic features were compared using two methods: 1) an approach using feature selection with stepwise regression; and 2) an approach that weighs the relative similarity of a

21

1004P-ANF-WO01 set of features and the predicted variable against the relative similarity among those features. This latter approach allows for a weighting between the relevance of a set of features to a predicted variable, and their internal redundancy. One use of this approach is to avoid issues of over-fitting or over-specification. Additionally, this latter approach to feature selection avoids the need to perform iterative 'hill-climbing' optimization and instead finds a global optimum quickly.

[0079] Fluid accumulation in the neck may increase pharyngeal tissue pressure around the UA and consequently narrow the UA. Based on a tubular model of the UA, narrowing in the oro-pharyngeal part of the UA could change the spectral shape of the generated tracheal sound. In vowel articulation in speech, narrowing in the top or back of the oral cavity due to tongue movement can decrease the first and second formants, respectively. Posterior movements of the tongue could reflect narrowing in the UA. Therefore, it may be expected that UA narrowing could cause a shift in the power components of the tracheal sounds to the lower frequencies and decrease the second formant of tracheal sound. These changes occurred in some of our subjects, independent of their age or gender. However, as presented in Figure 4, this pattern was not consistent in all subjects. This may be due to the differences among subjects in the UA dilator muscle activities or reflexes. The variability in the formant frequencies may explain why the increases in lower power frequency components of tracheal sounds were not significant among all subjects. [0080] Across both methods of acoustic feature selection, the majority of selected features included Mel-frequency cepstral coefficients, which are widely used to develop source-filter models of speech and separate the effects of the glottal source from those of the vocal tract. This aligns with the observation that the 2^nd and 6^th MFC coefficients changed significantly over time. Other time-frequency domain features that may be expected to produce similarly accurate results may include, but are not limited to wavelet, short time Fourier, Bilinear time- frequency, bispectrum and Wigner coefficients, for example.

[0081] As demonstrated, tracheal sound analysis can be used, as described above, to develop non-invasive, convenient and reliable methods to estimate fluid accumulation in the neck, which can provide a useful, non-invasive tool in the study of the pathophysiology of

22

1004P-ANF-WO01 sleep apnea, for example, and in providing useful diagnostic and/or monitoring information for the treatment of OSA, of example.

[0082] While the present disclosure describes various exemplary embodiments, the disclosure is not so limited. To the contrary, the disclosure is intended to cover various modifications and equivalent arrangements included within the general scope of the present disclosure.

23

1004P-ANF-WO01

Claims

What is claimed is: 1. A neck fluid volume assessment device for use with a subject while breathing, the device comprising:

a microphone to be positioned in an area of the subject so to acquire acoustic breath sounds emanating from the subject and generate a signal representative thereof;

a digital storage device having stored thereon a neck fluid volume assessment engine having associated therewith one or more designated acoustic features previously identified to provide a measure of neck fluid volume; and

a data processor operatively coupled to said digital storage device to implement said neck fluid volume assessment engine to act on said signal to automatically extract said one or more designated acoustic features therefrom and output an indication of the subject's neck fluid volume as a function of said extracted one or more features.

2. The device of claim 1, wherein said microphone is to be positioned in skin-contact with one of a throat, cheek or ear area of the subject.

3. The device of claim 2, wherein said microphone is to be positioned in skin-contact with a throat area of the subject.

4. The device of claim 2, wherein said microphone is to be positioned inside an ear canal of the subject.

5. The device of claim 1, wherein said area is selected from in front of the subject's face, in a same room as where the subject sleeps and above the subject's head while sleeping on a bed.

6. The device of any one of claims 1 to 5, wherein said output indication comprises an estimated fluid volume in the subject's neck.

24

1004P-ANF-WO01

7. The device of any one of claims 1 to 6, wherein said one or more designated acoustic features comprise at least one time domain feature, frequency domain feature or time- frequency domain feature.

8. The device of any one of claims 1 to 7, wherein said one or more designated acoustic features comprise at least one predominantly speech-related feature.

9. The device of claim 8, wherein said speech-related feature comprises at least one of a 2^nd and a 6^th mel -frequency cepstrum coefficient.

10. The device of any one of claims 1 to 9, wherein said one or more designated acoustic features comprise a designated set of acoustic features.

11. The device of any one of claims 1 to 10, wherein said neck fluid volume engine comprises a signal classifier previously trained against a known data set of acoustic breath sound signals, wherein each of said one or more features were extracted from each of said acoustic breath sound signals and classified, alone or in combination, to correspond with a defined output indication, thereby defining characterization criteria to which said one or more designated features extracted from newly acquired breath sound signals are compared in automatically generating each said output indication.

12. The device of claim 11, wherein said signal classifier comprises one of a regression- type classifier and a neural network classifier.

13. The device of any one of claims 1 to 12, wherein said signal is time-segmented to define multiple data segments, wherein said one or more designated acoustic features are extracted from each of said data segments so distinctly characterize the subject's neck fluid volume over time.

25

1004P-ANF-WO01

14. A non-invasive neck fluid volume assessment method to be performed on a subject, the method comprising:

receiving as input a signal representative of breath sounds generated by the subject; extracting one or more designated acoustic features from said input signal, wherein said one or more designated acoustic features define one or more preset neck fluid volume assessment metrics;

comparing said one or more extracted features with said one or more preset neck fluid volume assessment metrics; and

outputting, based on said comparing, characterization of the subject's neck fluid volume as a function of said one or more metrics.

15. The method of claim 14, wherein said one or more metrics comprise a quantifiable neck fluid volume.

16. The method of claim 14 or claim 15, wherein the method further comprises acquiring said breath sounds via a microphone positioned in skin-contact with one of a throat, cheek or ear area of the subject.

17. The device of claim 16, wherein said microphone is to be positioned inside an ear canal of the subject.

18. The device of claim 14, wherein said area is selected from in front of the subject's face, in a same room as where the subject sleeps and above the subject's head while sleeping on a bed.

19. The method of any one of claims 14 to 18, wherein said one or more designated acoustic features comprise at least one time domain feature, frequency domain feature or time- frequency domain feature.

20. The method of any one of claims 14 to 19, wherein said one or more designated acoustic features comprise at least one predominantly speech-related feature.

26

1004P-ANF-WO01

21. The method of any one of claims 14 to 20, wherein said one or more designated acoustic features comprise a designated set of acoustic features.

22. The method of any one of claim 14 to 21, wherein the method further comprises segmenting said signal into respective time segments, and wherein said extracting and said comparing is implemented for each of said time segments to characterize the subject's neck fluid volume over time.

23. A computer-readable medium having statements and instructions stored thereon for implementation by a processor to act on a signal representative of acoustic breath sounds emanated by a subject in outputting an assessment of the subject's neck fluid volume by performing the steps of any one of claims 14 to 22.

24. A non-invasive method for assessing a neck fluid volume in a subject, the method comprising:

acquiring acoustic breath sounds emanating from the subject over a time period;

generating a data signal representative of said acquired acoustic breath sounds;

extracting one or more designated acoustic features from said data signal, wherein said one or more designated acoustic features define, alone or in combination, a neck fluid volume metric;

comparing said one or more extracted features with said neck fluid volume metric; and outputting, based on said comparing, an estimated neck fluid volume in the subject during said time period.

25. The method of claim 24, wherein said signal is time-segmented to define multiple time periods, and wherein said extracting, comparing and outputting is implemented for each of said multiple time periods to monitor a variation of said estimated neck fluid volume over time.

27

1004P-ANF-WO01

26. The method of claim 24 or claim 25, further comprising, prior to said acquiring, having a microphone positioned in skin-contact with one of a throat, cheek and ear area of the subject, wherein said acquiring comprises acquiring said acoustic breath sounds via said microphone.

27. A method of manufacturing a neck fluid volume assessment device, the method comprising:

acquiring acoustic breath sounds from multiple subjects to generate respective data signals representative thereof;

concurrently measuring respective neck fluid volumes in said subjects during said acquiring;

extracting one or more designated acoustic features from each of said signals;

classifying said one or more extracted features, alone or in combination, for each of said signals, to correspond with an estimated neck fluid volume based on said concurrently measured neck fluid volumes;

defining a neck fluid volume metric based on said classifying to be applied to said one or more designated features once extracted from input breath sound signals to output a corresponding neck fluid volume assessment; and

programming a computing device with said defined metric so to act on new input breath sound signals to:

extract said one or more designated acoustic features therefrom; compare said one or more extracted features with said neck fluid volume metric; and

output, based on said comparing, an estimated neck fluid volume indication.

28. The method of claim 27, wherein said measuring comprises measuring a bioelectrical impedance of the neck.

29. The method of claim 27, wherein said measuring comprises measuring said respective neck fluid volumes from magnetic resonance imaging of the neck.

28

1004P-ANF-WO01

30. The method of any one of claims 27 to 29, wherein said classifying comprises implementing a regression-type or a neural network classifier on said extracted features.

29

1004P-ANF-WO01