WO2013086615A1

WO2013086615A1 - Device and method for detecting congenital dysphagia

Info

Publication number: WO2013086615A1
Application number: PCT/CA2012/001145
Authority: WO
Inventors: Celeste MEREY; Thomas T. K. CHAU
Original assignee: Holland Bloorview Kids Rehabilitation Hospital
Priority date: 2011-12-16
Filing date: 2012-12-14
Publication date: 2013-06-20

Abstract

Disclosed herein are methods and devices for identifying congenital dysphagia, for in pediatric candidates. In general, methods and devices considered herein make use of cervical dual-axis accelerometer in distinguishing safe from unsafe swallowing events.

Description

DEVICE AND METHOD FOR DETECTING CONGENITAL DYSPHAGIA

FIELD OF THE DISCLOSURE

[0001] The present disclosure relates to dysphagia, and in particular, to a device and method for detecting congenital dysphagia. BACKGROUND

[0002] Feeding disorders encompass a broad range of problems associated with eating solid and liquid foods. Difficulty with the process of swallowing is known as dysphagia, and can occur in both adult and pediatric populations. Epidemiologic data on the prevalence of dysphagia in children is not readily available. However, feeding disorders as a whole are estimated to be present in a significant and increasing portion of the pediatric population: in 25% to 45% of typically developing children and in 33% to 80% of children with developmental disorders. Dysphagia impacts the health and well- being of a child as the disorder may lead to malnutrition, dehydration, impairment of physical growth and developmental delay. Dysphagia can also induce feeding-related stress and challenges, affecting the psychosocial well-being of the child, family and other caregivers. A particularly dangerous condition, aspiration pneumonia, is frequently associated with dysphagia.

[0003] Early evaluation by a clinical team may greatly reduce health issues that can result from dysphagia. In particular, instrumental evaluations of swallowing facilitate the visualization of the bolus trajectory and motion of anatomical structures throughout the different phases of swallowing. The current standard is the videoflouroscopic swallow study (VFSS), where the patient swallows barium coated substances of various consistencies while lateral X-ray images of the oral cavity, pharynx, larynx and upper esophagus are displayed in real-time for live viewing and recorded for subsequent review. As one part of the assessment, the clinical team discerns whether or not the bolus passes into the airway, either into the laryngeal vestibulum above the vocal chords or past the vocal chords and into the inferior airways.

[0004] Clinical evaluations of the health of the swallow have yielded varying levels of agreement amongst clinicians and trained experts. Although perfect agreement was not achieved, expert diagnosis matched on most occasions, indicating the effectiveness of VFSS as a means of detecting unsafe swallowing. Earlier studies resulted in agreement rates of 93-95% in clinical diagnosis of an unhealthy swallow, whereas more recent studies found a higher variability and lower accuracy of detecting unhealthy swallows: between 77% and 88% agreement amongst participating clinicians. [0005 j VFSS is not without its shortcomings. Proper interpretation of the swallow usually requires an experienced practitioner or a team of assessors. The method itself also exposes children to ionizing radiation and therefore should be used minimally. As well, the process is expensive both in terms of equipment and human resources. Furthermore, VFSS can only provide a snapshot of a patient's swallowing function, despite the fact that this function can vary from day to day. Finally, many children find VFSS frightening and uncomfortable.

[0006] Recently, the use of accelerometry has been under investigation as a noninvasive, low cost technique for characterizing swallowing. Initial research involving the placement of a single-axis accelerometer to measure throat vibrations during swallowing has garnered positive results in the identification of dysphagic activity within the adult population. For example, single axis accelerometry was proposed as a viable option for detecting aspiration in U.S. Patent No. 7,749,177 to Chau et al. Nonetheless, it is generally observed that the patient's compliance to measurement protocol can affect the accuracy of classifying swallows. In the pediatric case, one can reasonably expect that a child may be inclined to exhibit more spontaneous vocalizations and bodily movements during the feeding protocol, thus in turn generating more contaminant vibration signals. Accordingly, the classification of swallowing accelerometry in the pediatric population presents different analytical challenges than those encountered in the adult case. While, the classification of uniaxial vibrations associated with swallowing activity in children was proposed using a radial basis classifier, the 80% adjusted accuracy achieved using this classifier only detected a very specific event, i.e., airway entry with inspiratory airflow, and was thus of limited utility.

[0007] Therefore, there remains a need for a method and device for detecting congenital dysphagia, for example in pediatrics, that overcome at least some of the drawbacks of known techniques, or at least, provides a useful alternative. [0008) This background information is provided to reveal information believed by the applicant to be of possible relevance to the invention. No admission is necessarily intended, nor should be construed, that any of the preceding information constitutes prior art against the invention.

SUMMARY [0009] Some aspects of this disclosure provide a method and device for detecting congenital dysphagia, for example in pediatrics, that overcome at least some of the drawbacks of known techniques, or at least, provides a useful alternative. In accordance with one embodiment, there is provided device for use in identifying congenital dysphagia from execution of a swallowing event, the device comprising: a dual axis accelerometer to be positioned in a region of the candidate's throat and configured to acquire axis-specific vibrational data representative of the swallowing event; and a processing module operatively coupled to said accelerometer and configured to process said axis-specific data to extract therefrom, for each axis, a set of designated features representative of the swallowing event, and classify said vibrational data, based on at least some of said extracted features, as indicative of one of normal swallowing and aspiration due to congenital dysphagia. [0010] In one embodiment of the device, the device is for use with pediatric candidates.

[0011) In one embodiment of the device, the processing module comprises a classifier configured to receive as input said extracted features and output a classification specific to the swallowing event indicative of a normal event or an aspiration event.

[0012] In one embodiment of the device, said classifier comprises a non-linear classifier.

[0013] In one embodiment of the device, said classifier comprises a Support Vector Machines (SVM) classifier with a Radial Basis Function (RBF) kernel. [0014] In one embodiment of the device, said plurality of features comprises at least one feature representative of a main lobe of vibrational data.

[0015] In one embodiment of the device, said at least one feature representative of said main lobe of vibrational data comprises a peak vibrational amplitude.

[0016] In one embodiment of the device, said accelerometer is configured for alignment along an anterior-posterior axis (A-P) and a superior-inferior axis (S-I) of the candidate's throat.

[0017] In one embodiment of the device, said plurality of features comprises at least one time-frequency feature.

[0018] In one embodiment of the device, said at least one time-frequency feature comprises a feature related to a short-time Fourier transform of said vibrational data.

[0019] In one embodiment of the device, said processing module having stored therein a designated input feature vector identified from a known vibrational data set to exploit correlations between said set of designated features, said processing module further configured to transform said extracted features as a function of said input feature vector and classify said vibrational data as a function of said transformed features.

[0020] In one embodiment of the device, a dimension of said designated input feature vector is less than that of said designated set of features. [0021] in one embodiment of the device, said designated input feature vector is defined by a selected set of principle components identified from a principle component analysis of said known vibrational data set.

[0022] In accordance with another embodiment of the invention, there is provided a method for identifying congenital dysphagia in a candidate via a designated candidate swallowing event, the method comprising: recording dual-axis vibrational data representative of the swallowing event; extracting, for each axis, a set of designated features from said dual axis vibrational data; and classifying said extracted features as indicative of one of normal swallowing and aspiration due to congenital dysphagia.

[0023] In one embodiment of the method, the candidate consists of a pediatric candidate.

[0024] In one embodiment of the method, said classifying step comprises comparing said extracted features with a pre-classified set of such features, said pre-classified set extracted from a known vibrational data set and classified as representative of candidates having known congenital dysphagia conditions. [0025] In one embodiment of the method, said comparing step comprises transforming said extracted features in accordance with a designated feature vector identified from said known vibrational data set to exploit correlations between said set of designated features; and classifying said transformed features as a function of one or more designated classification criteria previously identified from said pre-classified set to distinguish between such feature vectors representative of congenital dysphagia and those representative of normal swallowing.

[0026] In accordance with another embodiment of the invention, there is provided a method for classifying cervical dual-axis acceierometry data acquired in respect of a candidate swallowing event to detect congenital dysphagia, comprising: receiving as input axis-specific vibrational data representative of the swallowing event; extracting a plurality of designated features representative of the swallowing event for each axis of said axis-specific vibrational data; comparing said extracted features with one or more designated classification criteria previously identified from such features when extracted from a known vibration data set to distinguish between normal and impaired swallowing events; and outputting, based on said comparing, classification of said vibrational data as indicative of one of normal swallowing and aspiration due to congenital dysphagia.

(0027] In one embodiment of the method, said comparing comprises transforming said extracted features in accordance with a designated feature vector identified from said known vibrational data set to exploit correlations between said set of designated features, said one or more designated classification criteria previously identified to distinguish between such feature vectors representative of congenital dysphagia and those representative of normal swallowing.

[0028] In one embodiment of the method, said designated input feature vector is defined by a selected set of principle components identified from a principle component analysis of said known vibrational data set.

[0029] In accordance with another embodiment of the invention, there is provided a computer readable-medium having statements and instructions stored thereon for implementation by a processor to automatically implement the above methods. [0030] In accordance with another embodiment of the invention, there is provided a device for use in identifying congenital dysphagia from execution of a swallowing event, the device comprising: an input operable to receive axis-specific vibrational data representative of the swallowing event, said vibrational data acquired from a dual-axis accelerometer positioned in a region of the candidate's throat; the above computer readable medium; a processor operable to implement the statements and instructions stored on said computer readable medium; and an output operable to render classification of said vibrational data as indicative of one of normal swallowing and aspiration due to congenital dysphagia. [0031] In one embodiment of the device, the device further comprises said dual-axis accelerometer.

[0032] Other aims, objects, advantages and features of the invention will become more apparent upon reading of the following non-restrictive description of specific embodiments thereof, given by way of example only with reference to the accompanying drawings.

BRIEF DESCRIPTION OF THE FIGURES

[0033] Several embodiments of the present disclosure will be provided, by way of examples only, with reference to the appended drawings, wherein:

[0034] Figure 1 is a schematic diagram of an aspiration detection device in operation, in accordance with an embodiment of the invention;

[0035] Figure 2 is a schematic diagram of an aspirometer, and components thereof, in accordance with an embodiment of the invention; [0036] Figure 3 is a high level dual axis accelerometry data processing flow diagram for implementation by an aspiration detection device, in accordance with an embodiment of the invention;

[0037] Figure 4 is an illustrative dual axis accelerometry data processing flow diagram, for implementation by an aspiration detection device, in accordance with an embodiment of the invention;

[0038] Figure 5 are illustrative plots of dual-axis swallowing accelerometry signals, wherein (a) and (b) show signals for the A-P and S-I axes, respectively, having been cleansed of vocalizations and head movements, and subsequently segmented, wherein the same signals are shown after deno i s i ng and trimming in the A-P (c) and S-I (d) axes; and

[0039] Figure 6 is an exemplary detailed dual axis accelerometry data processing flow diagram, in accordance with one embodiment of the invention.

DETAILED DESCRIPTION [0040] As introduced above, dysphagia or swallowing disorder can negatively impact a child's health and development, as can it adversely affect the health of older subjects when not properly diagnosed, treated and/or managed. The gold standard of dysphagia detection is videofluoroscopy which exposes the child or older subject to ionizing radiation, and requires specialized clinical expertise and expensive institutionally-based equipment, precluding day-to-day and repeated assessment of fluctuating swallowing function.

[0041] As an alternative to videofluoroscopy, the methods, systems and devices described herein, in accordance with different embodiments of the invention, provide for the identification of unsafe swallows via swallowing acceierometry, which is the non- invasive measurement of cervical vibrations during swallowing, and which can provide a portable and cost-effective bedside alternative to the current standards. In one embodiment, dual-axis accelerometry is used to accurately identify unsafe swallowing events in pediatrics, or again within the context of older candidates manifesting symptoms of, or suspected of suffering from congenital dysphagia, for example, via implementation of a classifier, classification system and method, as described in greater detail below. Namely, in the Example described below, both male and female participants were examined and each had an identifiable neurological condition such as Cerebral Palsy, seizure disorder, developmental delay, brain injury and Downe Syndrome. Meanwhile, the age range of the participants was wide: from 6 months to 18 years. Accordingly, while each candidate was effectively qualified for pediatric care, the overall range in physical characteristics such as anatomical size and movements varied widely between participants, without adversely affecting the performance of the exemplary method implemented in classifying vibrational data acquired from these candidates. Accordingly, the embodiments of the invention herein considered, did not show dependence on age or physical size, but rather provided, in one example, for the recognition of neurological dysphagia.

[0042] In particular, accurate classification of dual-axis cervical accelerometry signals was accomplished despite a general reduction in patient compliance with measurement protocols expected from pediatric candidates, which can generally have an impact on classification accuracy. For instance, a child undergoing testing may be inclined to exhibit more spontaneous vocalizations and bodily movements during the feeding protocol, in turn, generating more contaminant vibration signals that can impede or reduce accurate classification. Nonetheless, the classifier, classification system and method described below, in accordance with different embodiments of the invention, have proven sufficiently robust to accommodate the increased analytical challenges brought forth by the pediatric population, and thus equally applicable to older candidates suspected of suffering from congenital dysphagia. [0043] To validate this approach, and as will be described in greater detail below, dual-axis accelerometric signals were collected simultaneous to videofluoroscopic records from 29 pediatric participants (age 6.8 ± 4.8 years; 20 males) previously diagnosed with neurogenic dysphagia. Participants swallowed 3-5 sips of barium-coated boluses of different consistencies (normally, from thick puree to thin liquid) by spoon or bottle. Videofluoroscopic records were reviewed retrospectively by a clinical expert to extract swallow timings and ratings. The dual-axis acceleration signals corresponding to each identified swallow were pre-processed, segmented and trimmed prior to feature extraction from time, frequency, time-frequency and information theoretic domains. Feature space dimensionality was reduced via principal component analysis.

[0044] Using 8-fold cross-validation, 16-17 dimensions and a support vector machine classifier with an RBF kernel, an adjusted accuracy of 89.6% ± 0.9 was achieved for the discrimination between swallows with and without airway entry. Results thus validate the use of dual-axis acceierometry in the non-invasive detection of unsafe swallows in children, and candidates suspected of suffering from a congenital swallowing disorder in general.

[0045] Referring now to Figure 1, a system for use in aspiration detection, generally referred to using the numeral 100, and in accordance with an illustrative embodiment of the invention, will now be described. In this example, the system 100 generally comprises a dual axis accelerometer 102 to be attached in a throat area of a candidate (a child or older subject suspected of suffering from congenital dysphagia, in this example) for acquiring dual axis acceierometry data and/or signals during swallowing (e.g. see illustrative S-I Acceleration signal 104 shown in Figure 1). For example, acceierometry data may include, but is not limited to, throat vibration signals acquired along the anterior-posterior axis (A-P) and superior-inferior axis (S-I). [0046] The accelerometer is operatively coupled to an aspirometry data processing module or aspirometer 106 configured to process the acquired data in accordance with the classification system and method discussed below, in distinguishing healthy from unhealthy swallows. The processing module 106 is depicted herein as a distinctly implemented device, or aspirometer, operatively coupled to accelerometer 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 ordinary skill in the art. The processing module may, however, in accordance with another embodiment, be implemented integrally with the accelerometer, for example, depending on the intended practicality of the aspirometer, and/or context within which it is to be implemented. As will be appreciated by the skilled artisan, the processing module may further 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 diagnostics tools.

[0047] With reference to Figure 2, the processing module, depicted herein generically as a self-contained device or aspirometer 200, generally comprises a power supply 202, such as a batteiy 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 accelerometer (as shown in Figure 1), external data processing module, 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 associated with the acquisition and processing (i.e. classification) of accelerometry data for each swallowing event. The device 200 may further comprise a user interface 210, 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 and/or processed data with respect to aspiration detection, screening, etc. (e.g. graphical user interface such as CRT, LCD, LED screen or the like, visual and/or audible signals/alerts/warnings/cues, numerical displays, etc.)

[0048] 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 illustrative embodiment of aspirometer 200 may be considered herein without departing from the general scope and nature of the present disclosure. It will further be appreciated that aspirometer 200 may equally be implemented as a distinct and dedicated device, such as a dedicated home, clinical or bedside aspiration detection 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, or the like.

[0049] Referring now to Figure 3, an example of a data processing stream, in accordance with one embodiment of the invention, will now be described. In general terms, the processing of acquired or collected dual axis accelerometry data 302 representative of at least one swallowing event may be composed of two broad steps, namely a dual axis feature extraction step 304 applied for data representative of each swallowing event along each axis, and a swallowing event classification step 306 based on the extracted features of step 304. In applying this approach to dual axis cervical accelerometry data representative of respective swallowing events, such swallowing events may be effectively classified as normal swallowing events or aspiration events. [0050] Referring now to Figure 4, and in accordance with an embodiment of the invention, a further illustrative dual axis accelerometry data processing flow will be described. In this particular embodiment, accelerometry data 402 is acquired or provided in respect of multiple swallowing events. This data is then preprocessed at step 404 to condition the raw data and thus facilitate further processing thereof. For example, the raw data may be filtered, denoised and/or processed for signal artifact removal.

[0051] The preprocessed data is then manually or automatically segmented into distinct swallowing events (step 406). For example, an automated swallowing event segmentation process, such as described in co-pending United States Patent Application Publication No. 2010/0160833, the entire contents of which are incorporated herein by reference, may be applied to the data to segment this data by swallowing event. Alternatively, manual segmentation may be applied upon visual inspection of the data (e.g. identification of the start of each swallowing event, which may be readily and systematically recognized by an operator of the device). Alternatively, data may be recorded distinctly for each swallowing event, thus naturally providing event-specific data without need for further segmentation. In such embodiments, it will be appreciated that swallowing event-specific data may be preprocessed individually, thus effectively applying the manual signal segmentation step 406 of Figure 4 during acquisition of accelerometry data 402 and prior to preprocessing step 404. As will be appreciated by the skilled artisan, these and other such variations may be considered herein without departing from the general scope and nature of the present disclosure.

[0052] The event-specific data is then processed by a dual axis feature extraction module 408 configured to extract for each axis, a plurality of preset features previously identified as amenable to distinguishing signals respective to each output classification of interest (e.g. as identified via classifier training, discussed below), allowing for each swallowing event to be classified at step 412 based on these extracted features. As discussed above generally, such classification thus allows for the determination and output 414 of which swallowing event represented a normal swallowing event as compared to a likely aspiration event.

[0053] For example, in one embodiment, the device 200 of Figure 2 may be configured to implement the above classification as a function of a designated feature set previously identified to distinguish, when compared with designated classification criteria, between vibrational data representative of a normal swallowing event and that representative of congenital dysphagia.

[0054] For example, the device 208 may be configured to receive as input axis- specific vibrational data representative of the swallowing event (e.g. from an integrated, associated and/or remote or distinct dual-axis accelerometer) and extract therefrom (or from a pre-processed representation thereof) a plurality of designated features representative of the swallowing event for each axis of the axis-specific vibrational data. Once extracted the features may be compared, either directly or indirectly, with one or more designated classification criteria, these criteria previously identified from such features when extracted from a known vibrational data set to distinguish between normal and impaired swallowing events. As a result of this comparison, an output is provided classifying the input data as indicative of one of normal swallowing and aspiration due to congenital dysphagia. In one such example, comparison of input data with the one or more classification criteria is performed via a designated transformation of the input data, for example, in accordance with a designated feature vector identified from the known vibrational data set to exploit correlations between the set of designated features. Similarly, the classification criteria may be characterized as resulting from the application of a same transformation to the known data set, thus allowing for comparison of the feature vector associated with the acquired data, with those similarly associated with the blown data set and thus known to represent normal or impaired swallowing events. [00551 In the example provided below, the designated input feature vector is defined by a selected set of principle components identified from a principle component analysis of the known vibrational data set, and in particular, in which a dimensionality of the input vector is thus effectively reduced from an overall dimensionality of the extracted feature set. Designated classification criteria, in the below example output from the training of a selected data classifier operating on the known feature vectors, result when applied to a new input feature vector, in the classification of this new vector as indicative of a normal or impaired swallowing event due to congenital dysphagia.

[0056J Reference will now be made to the following non-limiting example, wherein an embodiment of the above-described classification method and system was applied in classifying accelerometry data acquired during distinct swallowing events in children for the puipose of distinguishing between healthy swallows irom those exhibiting signs of aspiration due to congenital dysphagia, and thus validating the above-described approach to aspirometry. [0057] In this example, cervical accelerometry swallowing samples were collected from 29 patients (aged 6.8 ± 4.8, 20 male) of the Holland Bloorview Kids Rehabilitation Hospital's feeding clinic. All participants were diagnosed with probable feeding disorders at a previous appointment by the attending pediatrician through medical history and feeding patterns. The data collection session occurred during the feeding clinic's VFSS assessment.

[0058] The accelerometer (Analog Devices, ADXL322.) was taped to the participant's neck at the level of the cricoid cartilage such that the sensitive axes of the accelerometer were aligned with the superior-inferior (S-I) and posterior-anterior (A-P) anatomical directions. Each child was fed barium-coated boluses of varying consistencies, ranging from thick puree to thin liquid. Children were generally fed by spoon and infants were fed by bottle. The starting consistency varied by child and was determined based on the recent case history.

[0059] The swallowing vibration signals from the accelerometer were filtered and amplified (Astro-Med Inc., Grass, P55 A.C. Preamplifier; filtered lHz to 3 kHz; amplification lOx) and sampled at 10 kHz via a data-acquisition card (National Instruments, USB NI-6210) prior to storage on a research computer. Bolus activity in the cervical region was captured by lateral fluoroscopic video recording onto a research computer via a PCI card (National Instruments, PCI 1405). The trigger to initiate and terminate the recording of these inputs was controlled through a custom LabVIEW application (National Instruments), which served to synchronize the VFSS video recording with the accelerometric data, enabling swallow time-stamp identification retrospectively.

[0060) All recordings of the VFSS video were reviewed off-line by the feeding clinic's pediatrician to ascertain the time-stamp of occurrence and nature of each swallow taking place. The VFSS video was played at one third speed and the point at which the bolus passed below the epiglottis was selected as the time of swallow. Each swallow was assigned one of three ratings, as outlined in Table 1, below.

Rating Clinical description

0 Material does not enter the airway

1 Material enters the airway, and contacts the true vocal folds but does not pass below

2 Material enters the airway, and passes below the true vocal folds

Table 1 : Swallow rating based on path of bolus; Each swallow was assigned a number using this system based on visual observation of the VFSS video playback [0061] Swallows rated as 1 were discarded from the study to avoid ambiguity between safe and unsafe swallowing. The resultant data set contained 94 swallows rated 2 (i.e. Unsafe swallows) and 544 swallows rated 0 (i.e. safe swallows).

[0062] The dual axis accelerometer generated time-stamped S-I and A-P vibration signals. Each signal was processed individually. Initially, the signal was down-sampled to 1 kHz. Next, segments deemed to contain vocalizations by the robust algorithm for pitch tracking (RAPT) were identified. The RAPT parameters were adjusted to minimize the number of discarded segments which contained swallows as follows: the maximum possible fundamental frequency was set to 1000 Hz, the duration of frame size was set to 0.01s, the duration of the correlation window was set to 0.01s and the maximum number of hypotheses at each frame was set to 50. Segments identified by the RAPT algorithm with the adjusted parameters were discarded. Any swallows within these removed segments were therefore eliminated from further consideration.

[0063] Low frequency components associated with head movement were de-trended using a least-squares spline approximation. Segmentation of each swallow from the accelerometric record proceeded as follows: the signal was marked at the time stamp identified via videofluoroscopy. Each segment was produced by expanding one second on either end of the time stamp identified via videofluoroscopy, ensuring swallow activity was captured. Each segment was trimmed through visual and audio inspection to eliminate extraneous data from the beginning or end of each segment.

[0064] Each segment was filtered using a denoising algorithm developed for swallowing accelerometry signals, such as that described in co-pending US Patent Application no. 12/819,216, the entire contents of which are incorporated herein by reference. This method was based on a ten-level discrete wavelet transform using the Meyer wavelet with soft thresholding. Figure 5 provides an example of a swallowing accelerometry signal that has been cleansed of vocalizations and head movements and subsequently segmented in the A-P (a) and S-I (b). The same signals are shown afte r de no i s i ng and trimming in the A-P (c) and S-I (d). Note that duration and visual characteristics depicted herein are meant to provide an example only, as such characteristics may vary between swallows. [0065] For each trimmed segment (i.e., each swallow) features were extracted from the time, frequency, time-frequency and information-theoretic domains to capture differences between safe and unsafe swallows. In this particular example, the following features were calculated for each of the A-P and S-I signal segments: mean, variance, skewness, kurtosis, entropy rate, memory, Lempel-Ziv complexity, measures of the relative energy and entropy in each wavelet decomposition level as extracted from a 10-level discrete wavelet decomposition of the signal using the discrete Meyer wavelet and as extracted from a 20-level Daubechies 5 wavelet decomposition, normality, stationarity, number of zero-crossings, dispersion ratio, interquartile range, peak Fast Fourier Transform (FFT) magnitude, frequency at spectral peak, and maximum hyolaryngeal excursion (estimated via double integration of accelerometry). Two features were calculated between the A-P and S-I signals, namely the cross-entropy rate and cross-correlation of the signals at zero lag.

[0066] The following additional features were also computed for each of the A-P and S-I signal segments. Namely, the following were computed from the time domain: median, the absolute difference between mean and median, p value of the chi-squared test for normality, total area under the main lobe (i.e. the portion of the signal corresponding to maximum hyolaryngeal excursion), and width of the main lobe. The following were calculated from the frequency domain: frequency corresponding to maximum spectral density over all time of the short-time Fourier transform spectrogram (max frequency); difference between frequency values corresponding to 75% and 25% of maximum spectral density at time corresponding to max frequency; and 20 features computed by the summation of power spectral density values within a 10Hz range, from 0 to 200Hz (calculated in unit of power per radian per sample using Welch's averaged modified periodogram method of spectral estimation).

[0067] Ultimately, 144 features were extracted for each of the 638 trimmed swallow samples. The number of healthy swallows largely outnumbered the alternate class, prompting the use of bootstrap re-sampling to create a balanced distribution class.

[0068] Classification was performed using 8 fold cross-validations. In each fold, training and test sets contained a balanced distribution of both classes. The dimensionality of the training subset was reduced using principal component analysis (PCA) with a variance threshold of 80%. The principle components served as the inputs to four different classifier models: linear discriminant analysis (LDA) using Euclidean and Mahalanobis distance measure variants, and support vector machines (SVM) trained with a linear kernel and with a Radial Basis Function (RBF) kernel. Adjusted accuracy was averaged over all folds and runs.

[0069] The classifier was further optimized through a method of removing redundant variables for a data set composed of more than 10 features. In particular, to determine which variables may be discarded, method B2 was deployed, as introduced by Joliffe in "Discarding variables in a principal component analysis. I: Artificial data. Journal of the Royal Statistical Society, Series C (Applied Statistics) 1972, 21(2): 160-173," the entire contents of which are hereby incorporated herein by reference. This method was selected given its ability to parsimoniously represent the data while achieving accuracies comparable to those attained with the full feature set. Using B2, the feature set was reduced by removing variables associated with the largest coefficient in each of the last 70 principal components when ordered by their contribution to variance, from largest to smallest. Principal component analysis was then performed on the new feature subset and dimensions were reduced to 16-17. The reduced subset of principal components was then input to the SVM classifier with the RBF kernel. All steps outlined in the preceding sections are outlined in Figure 6.

[0070] Table 2, below, displays averaged results of multiple runs of 8-fold cross- validation for each of the four different classification models using principal components comprising the full 144 feature set. The reduction in dimensionality as a result of PCA is also reported for each of the models.

Classifier Sensitivity Specificity Adjusted Dimensionality

.„(%} . (%) accuracy (%)

Using 144 features

LDA Euclidean 50.7±7.5 74.9±6.9 62.8±3.3 17

LDA Mahalanobis 69.8±2.1 51.4±4.7 60.6±1.9 16

SVM Linear 51.5±2.9 72.4±2.7 62.0±2.1 16

SVM RBF 80.0±7.0 81.2±5.3 80.6±3.9 18

Using 48 features

SVM RBF 86.6±0.9 92.2±1.6 89.9±0.6 16

Table 2: Averaged results for each classifier

[0071] Overall, SVM with the RBF kernel generated the highest average adjusted accuracy, 80.6%, a value which is significantly different from the adjusted accuracies of LDA with Euclidean, LDA with Mahalanobis and SVM with a linear kernel (Wilcoxon rank sum test, pO.001).

[0072] Using method B2, features calculated through wavelet decomposition (relative energy from each of the 10-levels of discrete Meyer wavelet, and the 20-level Daubechis 5 wavelet) as well as from the summation of power spectral densities were frequently associated with the largest coefficient of the last 70 components. These features were removed leaving a 48 feature subset. A listing of the resulting 48 features is provided in Table 3, below, wherein an identical set of features was used for each axis in this example, along with cross-correlation and cross-entropy rate between axes. It will be appreciated that distinct feature sets can alternatively be used, without departing from the general scope and nature of the present disclosure. Accordingly, PCA diminished the dimensionality from 144 to 16 or 17 dimensions, depending on the particular iteration and fold. The average input dimensionality to the classifier was 17 ± 1. Multiple runs of 8- fold cross-validation using this reduced feature set obtained an average adjusted accuracy of 89.6% ± 0.9, a sensitivity of 92.2% ± 1.6 and a specificity of 86.9% ± 0.9.

Table 3: Reduced Feature Set

[0073] New swallowing accelerometry features were introduced and used for classification in this study, with promising results given the attained accuracies.

[0074] For example, considering the features related to the main lobe, the peak accelerometry amplitude has been correlated to the movement of the hyoid bone and laiynx during a swallow. Studies evaluating dysphagic swallowing suggest penetration or aspiration may occur with delay or absence of hyolaryngeal movement or improper closure of the larynx. Therefore, differences in hyolaryngeal movement may result in a change in the duration and area of the main lobe, providing features which discriminate between healthy and unhealthy swallows. [0075] The selection of features relating to the short-time Fourier transform is also consistent with past studies that demonstrated the discriminative value of time-frequency features of swallowing accelerometry signals.

[0076J The dimensionality of the principal component sub-space balanced three criteria:

Most of the variability of the principal component space ought to be captured. To meet this criterion, principal components contributing to 80% of the variability were selected.

The "curse of dimensionality" ought to be mitigated. To fulfill this criterion, the ratio of sample size to the number of dimensions was 30, which exceeds the minimum of 5 prescribed in literature.

The number of principal components ought to minimize the error of the classifier. To meet this criterion, components were added until the "peaking" point was reached. These criteria were all met with the range of 16- 17 selected dimensions.

[0077] Removal of the wavelet decomposition and power spectrum summation components resulted in a significant improvement to the accuracy of the classifier. Metliod B2 may therefore be interpreted as a noise removal technique: the last 70 principal components out of 144 minimally contributed to the overall variability. Therefore the features that provide the most influence to these insignificant principal components can be considered to be disruptive to the overall accuracy of the classifier. Indeed, removal of these features resulted in a classifier with improved accuracy.

[0078] In reducing the number of features, variables associated with the largest coefficients of the last 70 components were evenly distributed between the AP and the SI axes. Likewise, the features retained were also sourced equally from each axis. No one axis was singled out as more useful than the other, reflecting the advantage of the dual axis combination.

[0079} The adjusted accuracy of the SVM classifier with the RBF kernel using the reduced feature set compares favorably with results attained in classifying adult dysphagic swallows, as well as studies of inter-rater reliability of videofluoroscopic examination.

[0080] The improvement in accuracy attained using the nonlinear RBF kernel indicates that the two classes were not linearly separable, and that the selection of a high dimensional input vector, i.e., 16 or 17 dimensions, may prove beneficial in achieving high accuracy. For instance, visual inspection of the accelerometer data did not reveal a regular signal pattern for either class, suggesting that no obvious low dimensional representation of the data would yield class separability. Further, the data were necessarily collected under a variety of conditions; feeding modality (spoon or cup), bolus viscosity (thin liquid to thick puree), participant age and etiology of dysphagia varied among participants. These variables may have introduced differential effects on the accelerometer signals, even within one class. Hence, a multi-dimensional, non-linear classifier may be selected, in accordance with one embodiment, to sort through such complex inputs, and thus provide a robust solution for widespread aspiration detection. [0081] In light of the above, dual-axis accelerometry for the discrimination of swallowing signals in the dysphagic pediatric population has been validated. In the above example, a nonlinear classifier was used to achieve an adjusted accuracy of 90% using an input vector of up to 17 dimensions. This study demonstrates that, regardless of the potential differences in input signal quality that may exist between adult and pediatric swallowing, dual-axis accelerometry holds significant promise for the screening for unsafe swallows in the pediatric domain. [0082] The above provides one example of a classification system for screening unsafe swallows in pediatrics, in accordance with one embodiment of the invention. Other embodiments may also be considered to provide similar effects. Namely, Table 4, below, outlines various alternative approaches to the above example.

Table 4: Alternative classification processes

[0083] For example, in one embodiment, Fisher's Ratio can be used instead of PCA to reduce dimensions prior to input to an LDA classifier. In other embodiments using PCA, the dimensions can be reduced according to different levels of percent contribution to variation (e.g. 75%, 80% or 85%), where in the above example an optimal percent of 80% was selected to achieve the above-reported accuracies. Similarly, using Fisher's ratio, a range of feature combinations can be selected for optimization. It was observed that the use of PCA to reduce dimensions (without use of B2 to remove features) generated generally higher adjusted accuracies than the use of the Fisher's Ratio to select features, but not significantly so. [0084] The number of folds utilized during cross-validation can also be adjusted, with general expectation that the accuracy will increase incrementally with increased folds.

[0085] Furthermore, different filters may be applied to the trimmed data, such as for example, a lowpass Butterworth filter with a passband of 0 to 190Hz with 3dB of ripple and a stopband of 220 to 500Hz with 60dB of attenuation, or a Daubechies 10 filter, to achieve similar results.

[0086] 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 spirit and scope of the appended claims. The scope of the following claims is to be accorded the broadest interpretation so as to encompass all such modifications and equivalent structures and functions.

Claims

CLAIMS:

1. A device for use in identifying congenital dysphagia from execution of a swallowing event, the device comprising:

a dual axis accelerometer to be positioned in a region of the candidate's throat and configured to acquire axis-specific vibrational data representative of the swallowing event; and

a processing module operatively coupled to said accelerometer and configured to process said axis-specific data to extract therefrom, for each axis, a set of designated features representative of the swallowing event, and classify said vibrational data, based on at least some of said extracted features, as indicative of one of normal swallowing and aspiration due to congenital dysphagia.

2. The device of claim 1, wherein the device is for use with pediatric candidates.

3. The device of claim 1, wherein said processing module comprises a classifier configured to receive as input said extracted features and output a classification specific to the swallowing event indicative of a normal event or an aspiration event.

4. The device of claim 3, wherein said classifier comprises a non-linear classifier.

5. The device of claim 4, wherein said classifier comprises a Support Vector Machines (SVM) classifier with a Radial Basis Function (RBF) kernel.

6. The device of any one of claims 1 to 5, wherein said plurality of features comprise at least one feature representative of a main lobe of vibrational data.

7. The device of claim 6, wherein said at least one feature representative of said main lobe of vibrational data comprises a peak vibrational amplitude.

8. The device of any one of claims 1 to 7, wherein said accelerometer is configured for alignment along an anterior-posterior axis (A-P) and a superior-inferior axis (S-I) of the candidate's throat.

9. The device of any one of claims 1 to 8, wherein said plurality of features comprise at least one time-frequency feature.

10. The device of claim 9, wherein said at least one time- frequency feature comprises a feature related to a short-time Fourier transform of said vibrational data.

1 1. The device of any one of claims 1 to 10, wherein said processing module has stored in association therewith a designated input feature vector identified from a known vibrational data set to exploit correlations between said set of designated features, said processing module further configured to transform said extracted features as a function of said input feature vector and classify said vibrational data as a function of said transformed features.

12. The device of claim 1 1 , wherein a dimension of said designated input feature vector is less than that of said designated set of features.

13. The device of claim 1 1 or 12, wherein said designated input feature vector is defined by a selected set of principle components identified from a principle component analysis of said known vibrational data set.

14. A method for identifying congenital dysphagia in a candidate via a designated candidate swallowing event, the method comprising:

recording dual-axis vibrational data representative of the swallowing event;

extracting, for each axis, a set of designated features from said dual axis vibrational data; and

classifying said extracted features as indicative of one of normal swallowing and aspiration due to congenital dysphagia.

15. The method of claim 14, wherein the candidate consists of a pediatric candidate,

16. The method of claim 14 or 15, wherein said classifying comprises comparing said extracted features with a pre-classified set of such features, said pre-classified set extracted from a known vibrational data set and classified as representative of candidates having known congenital dysphagia conditions.

17. The method of claim 16, wherein said comparing comprises:

transforming said extracted features in accordance with a designated feature vector identified from said known vibrational data set to exploit correlations between said set of designated features; and

classifying said transformed features as a function of one or more designated classification criteria previously identified from said pre-classified set to distinguish between such feature vectors representative of congenital dysphagia and those representative of normal swallowing.

18. A method for classifying cervical dual-axis accelerometry data acquired in respect of a candidate swallowing event to detect congenital dysphagia, comprising:

receiving as input axis-specific vibrational data representative of the swallowing event; extracting a plurality of designated features representative of the swallowing event for each axis of said axis-specific vibrational data;

comparing said extracted features with one or more designated classification criteria previously identified from such features when extracted from a known vibrational data set to distinguish between normal and impaired swallowing events; and

outputting, based on said comparing, classification of said vibrational data as indicative of one of normal swallowing and aspiration due to congenital dysphagia.

19. The method of claim 18, wherein said comparing comprises transforming said extracted features in accordance with a designated feature vector identified from said known vibrational data set to exploit correlations between said set of designated features, said one or more designated classification criteria previously identified to distinguish between such feature vectors representative of congenital dysphagia and those representative of normal swallowing.

20. The method of claim 19, wherein said designated input feature vector is defined by a selected set of principle components identified from a principle component analysis of said known vibrational data set.

21. A computer readable-medium having statements and instructions stored thereon for implementation by a processor to automatically implement the method of any one of claims 18 to 20.

22. A device for use in identifying congenital dysphagia from execution of a swallowing event, the device comprising:

an input operable to receive axis-specific vibrational data representative of the swallowing event, said vibrational data acquired from a dual-axis accelerometer positioned in a region of the candidate's throat; a computer readable medium as claimed in claim 23;

a processor operable to implement the statements and instructions stored on said computer readable medium; and

an output operable to render classification of said vibrational data as indicative of one of normal swallowing and aspiration due to congenital dysphagia.

23. The device of claim 22, wherein the device further comprises said dual-axis accelerometer.