If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Address correspondence and reprint requests to Julia Sidorova, Centro de Investigación Biomédica en Red de Enfermedades Hepáticas y Digestivas (CIBERehd), Hospital Clinic, Barcelona, Spain.
The possibility to estimate glucose value from voice would make a breakthrough in diabetes treatment: namely, remove the delay in the nonintrusive instantaneous blood glucose estimation, relieve medical budgets and significantly improve wellbeing of diabetics. In this review, different approaches have been described and systematized, in order to provide an objective snapshot of the state of the art. Since nonintrusive glucose estimation is notoriously difficult, we included a Transparence and Reproducibility Score aimed at revealing the biases in the primary research articles. The review is completed with the discussion on future research pathways.
A number of diseases and pathological conditions are reflected in the voice due to specific temporary or static changes that occur in the speaker's speech production organs or in the brain mechanisms controlling speech. Such voice changes can be heard and from a computational perspective can form learnable acoustic patterns generalizable across speakers and even across languages. The systems to detect such patterns have been built in the past decade to detect different neurological and psychological conditions and even in a disorder-independent way. The examples include stress recognition,
HALEF: an open-source standard-compliant telephony-based modular spoken dialog system: a review and an outlook.
in: Geunbae Lee G. Kook Kim H. Jeong M. Kim J.-H. Natural Language Dialog Systems and Intelligent Assistants. Springer International Publishing,
Switzerland2015: 53-61
Towards disorder-independent automatic assessment of emotional competence in neurological patients with a classical emotion recognition system: application in foreign accent syndrome.
for the cases, when condition-specific methods are not applicable: little data available for a new or rare disorder and/or different acoustic parameters affected from subject to subject. Formally speaking, a biomarker is “a characteristic that is objectively measured and evaluated as an indicator of normal biological processes, pathogenic processes, or pharmacologic responses to a therapeutic intervention.”
A vocal biomarker takes speech as an input and evaluates the subject's speech production (quality, competence or other aspects) at either a particular time moment or as a trend through a time interval (a recent review can be found in
). Vocal biomarkers possess an attractive combination of properties: noninvasive, instantaneous, and cost-effective. Currently, in diverse fields of healthcare, there is an urgent need for such solutions, to alleviate the ever growing medical budgets, better understand the disease, and improve therapies. Nonetheless, vocal biomarkers are only making it to become one of the core diagnostic criteria in neurology and are almost unheard of in endocrinology.
Diabetes mellitus is a chronic disease associated to one of the two mechanisms: inadequate production of insulin by pancreas (Type 1), or inadequate sensitivity of cells to the action of insulin (Type 2). Diligent and timely control of blood glucose (BG) levels is a crucial element in diabetes treatment. An instantaneous and reliable way to measure glucose value is obtaining a blood sample with a lancet and then estimating a glucose value with a blood glucometer. Many diabetics do not test their blood frequently enough because of inconvenience, lack of time, discomfort, and cost of the test strips for a glucometer. Consequently, they fail to discover the presence of the dangerous conditions of hypo- or hyperglycemia, which are too low or too high blood sugar, respectively. Among the diabetics, who do not monitor their glucose level diligently, the development of severe complications is frequent. Unobtrusive continuous blood glucose monitoring (CBGM) with implants was invented, but, unfortunately, it introduces a considerable delay in the measurements, which is long enough to trigger negative clinical consequences. Regarding truly noninvasive technologies, despite many research efforts, at the moment none of them is entirely reliable and convenient, and many noninvasive glucose estimation products were retracted. An interested reader can be referred to critical surveys of the topic, eg,
The problem of glucose value estimation is understood as the estimation of the numeric glucose value from a speech fragment or its mapping into hypoglycemia, normal BG, or hyperglycemia. In the adjacent field of research, glucose prediction refers to the prediction of future glucose values, eg,
therefore, to avoid ambiguity, the term BG estimation rather than prediction is more appropriate.
In this first literature survey on BG estimation from voice, we have answered the following questions.
1.
What makes one believe that the swings in BG trigger voice changes?
2.
What are the existing approaches to BG estimation from voice? Are there any research biases?
3.
What are the definite successes in the field? What remains to be an open problem?
The rest of the article is structured as follows. “The Search Protocol” Section describes the search strategy to find the publications on glucose estimation from voice. The sections “Why BG Affects Voice?” and “Existing Systems for Glucose Estimation From Voice” aim at creating a judgment-neutral snapshot of the state of the art. (A reader uninterested in computational details of the existing solutions can skim through the “Computational Approaches” Subsection.) The Sections “Missing Values” and “Biases in Data Collection” describe the research biases in the primary research articles and the pitfalls of data collection. The remaining sections cover Future Paths, Discussion and Conclusions.
THE SEARCH PROTOCOL
1
To search for the potentially applicable studies, we have searched in five databases of scientific literature covering the publications up to the article submission date in 2020: Cochrane library, PubMed, Scopus, Web of Science, and Google Scholar. The logical formula was: (speech OR voice) AND (diabetes OR glucose OR sugar OR hyperglycemia OR hypoglycemia). The queries are listed in Table 1, and the query was kept as large as possible, retrieving the articles containing the keywords anywhere. In Google Scholar the search was completed only over the titles, but the formula was extended with synonyms.
TABLE 1Queries for the Literature Search
Database
Query
# Hits (Not Listed, If Found in the Previous Database)
# Manually Selected as Relevant and Relevant Articles That Cite the Hit
Cochrane Library
#1 speech MeSH #2 glucose OR hypoglycemia OR sugar OR hypoglycemia OR diabetes OR SMBG OR self-monitoring of glucose #3 #1 AND #2
223
N/A
PubMed
(("speech"[MeSH Terms] OR "speech"[All Fields]) OR ("voice"[MeSH Terms] OR "voice"[All Fields])) AND (("glucose"[MeSH Terms] OR "glucose"[All Fields]) OR ("hyperglycemia"[All Fields] OR "hyperglycemia"[MeSH Terms] OR "hyperglycemia"[All Fields]) OR ("hypoglycemia"[All Fields] OR "hypoglycemia"[MeSH Terms] OR "hypoglycemia"[All Fields]) OR ("sugars"[MeSH Terms] OR "sugars"[All Fields] OR "sugar"[All Fields]) OR ("diabetes mellitus"[MeSH Terms] OR ("diabetes"[All Fields] AND "mellitus"[All Fields]) OR "diabetes mellitus"[All Fields] OR "diabetes"[All Fields] OR "diabetes insipidus"[MeSH Terms] OR ("diabetes"[All Fields] AND "insipidus"[All Fields]) OR "diabetes insipidus"[All Fields]) OR SMBG[All Fields] OR (("ego"[MeSH Terms] OR "ego"[All Fields] OR "self"[All Fields]) AND monitoring[All Fields] AND ("glucose"[MeSH Terms] OR "glucose"[All Fields])))
956 (more recent than 1999)
N/A
Web of Science
(speech OR voice) AND (diabetes OR glucose OR sugar OR hyperglycemia OR hypoglycemia)
allintitle: diabetes OR glucose OR hypoglycemia OR hyperglycemia OR sugar vocal OR acoustic OR perceptual OR speech OR voice; time span of 2009-2019 in articles and patents.
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Once the potentially applicable studies were retrieved, the irrelevant ones were discarded based on the manual scanning of the abstracts and full texts. A conservative approach was undertaken of including and summarizing every piece of research that describes either the construction of a vocal biomarker that detects glucose swings from voice or the voice changes in response to the change in glucose concentration.
3
The cited articles and the ones that cite the relevant articles were added to the pool of potentially applicable studies.
4
The research biases in the retrieved studies were analyzed: the description of the design alternatives is detailed with a Transparency and Reliability Matrix that was built to account for missing values, and the issues of data collection are addressed.
The search returned the works published on glucose swings reflected in voice: one journal article,
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Three more research articles and one literature survey describe the static changes in voice attributable to diabetes complications, and/or investigate the possibilities of diabetes detection from voice.
on noninvasive glucose monitoring provide a broader context for the vocal biomarker. Miscellaneous research articles were retrieved answering specific queries, when the primary sources lacked information or references. We give full consideration to primary research articles from gray literature, since conference proceedings have been the publication venues for many important technological breakthroughs in speech technology and computer science, as well as, for example, the fact that emotion recognition from voice is feasible first appeared in a patent in the late 1970s.
WHY BG AFFECTS VOICE?
Some studies simply claim that glucose-related changes in voice can be heard: the persons, who know the diabetic subject well can hear when they are hypoglycemic,
it is common knowledge that hypoglycemia is perceivable. Several studies provided scientific grounds to the not-obviously true statement that glucose swings are reflected in the voice. There are two explanations regarding the underlying mechanisms of voice changes as a response to BG shifts. One group of authors says that
•
the change in glucose level in the blood flowing in the larynx and the cords causes changes in the elastic properties of the biological tissue of these organs, which in turn results in the changes of spectral characteristics in compliance to the Hooke's law of physics.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
hypoglycemia is often accompanied by a feeling of anxiety, which causes people to speak faster and with greater urgency, whereas hyperglycemia on the contrary is often accompanied by feelings of lethargy thereby causing speech patterns to be slower or slurred.
Although no supporting references were provided, our investigation confirmed that both assumptions have scientific grounds. The subjective feelings by diabetics in the state of hypo- and hyperglycemia were reported in the 1990s, for example,
No related work regarding the other explanation about the change in elasticity of the larynx and cord was found by Google Scholar in English: (elasticity OR Hooke's law OR cord OR larynx) AND (hypoglycemia OR hyperglycemia OR glucose), but the search returned an article in Russian,
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
where it is stated that it was demonstrated experimentally that with the increase of BG the elasticity of those muscles decreases.
EXISTING SYSTEMS FOR GLUCOSE ESTIMATION FROM VOICE
The study designs differ according to the initial assumption why BG affects voice resulting in five approaches coming from different research communities, and generally, since the problem is new and the alternative system designs have not been compared yet, a plephora of approaches have been tried, similar to the early days of speech emotion recognition
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
hypoglycemia (dangerously low blood sugar) is defined as <70 mg/dL or <3.8 mmol/L, and hyperglycemia (dangerously high blood sugar) is defined as >180 mg/dL or >10 mmol/dL.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
BG readings are obtained from a drop of blood by means of a capillary finger-pricks with a lancet and then BG value is measured with a BG meter. In the case of CBGM, the BG value is obtained from a needle type sensor that is inserted into the subcutaneous system. The biases of ground truth values obtained with CBGM are covered in the “Biases of Data Collection” Section.
Patient groups
The studies were either clinical or involved volunteers (Table 3). The number of subjects (Table 6) was typically been below ten, except for a large-scale study with 7,000 subjects in.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
In contrast to the tradition of voice pathology research, where the speakers are excluded that have any other conditions, which are known to affect speech production organs or the neurological mechanisms related to speech,
says that they excluded (only) smokers. It should be added that this study comes from the voice pathology research and concerns with the description of the patterns in the whole population rather than their speaker-dependent detection. Likely, this disregard of the exclusion criteria is by design, since the studies aim at building a speaker-dependent predictive model relying on the features responsible for the BG swings and assuming that the distributions for a speaker are stationary, even though if permanently distorted by some pathology. In such a case, temporary conditions, such as flue or intoxication, are of relevance rather than static deformations of voice patterns due to chronic neurological or complications of diabetes.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
To be able to continuously monitor BG from speech in real-life applications, it is necessary to do recording with a smartphone or a wearable device.
Computational approaches
The study designs differ according to the initial assumption why BG affects voice (will be further detailed in Table 7) resulting in five approaches coming from different research communities. Below the existing systems are catalogued with respect to different options chosen in their respective system designs. The acoustic parameters and the computational analysis are further detailed in Tables 5 and 7, respectively.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
From the Fourier spectrum of the voice, ie, values in the coordinates of intensity vs frequency, the coefficients are determined for the solution of equations describing a physical system.
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
The Fourier spectrum. (The voice is transformed into spectrum, sound spectrum peaks in the areas of low (100-1,500 Hz) and high (7k-10k) frequencies are sampled, intensities of the selected peaks are determined by frequency, a ratio between the peaks of the selected low and high frequencies is obtained.)
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Short-term features from formants, pitch, articulation rate (eg, number transitions between voiced and unvoiced sounds), intensity, number speech errors, response time, nonfluency, speech quality, “S” sounds are shifted to “SH”, “R” goes to “L”, “EZ” goes to “ES”, delayed responsiveness.
Number of fundamental periods, time of fundamental periods, fundamental frequency, energy, amplitude of fundamental frequency, indicator of voiced probability, simple voice quality, relative average perturbation, shimmer, amplitude perturbation quotient, F1-F4 frequencies, harmonic perturbation quotients, residual to harmonic ratio, unharmonic to harmonic ratio, subharmonic to harmonic ratio, noise to harmonic ratio, F1 to F4 harmonic to all energy ratio.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
the elasticity coefficient changing in response to the glucose concentration in blood is thought to be the underlying reason of voice changes, and an approach from computational physics was followed
•
Approach IV: the spectral features based on the Fourier transform with the classification implementing a statistically averaged functional dependency.
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
In a more recent work, statistical approximation was replaced with an analytical solution.
•
Approach V: a set of differential equations was proposed describing a physical model of the speech organs, which links the BG concentration and the coefficients from the Fourier transform.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
via a statistical analysis of acoustic features, it was concluded that acoustic parameters in hypo- and hyperglycemia patterns in voice were significantly altered. Predictive models that would work on new samples were reported in.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
as a result of a large-scale study resulting in 98% accuracy for 7,000 subjects.
MISSING VALUES
The Transparency and Reproducibility matrix (Table 8) was built from the evidence tables (TABLE 2, TABLE 3, TABLE 4, TABLE 5, TABLE 6, TABLE 7). This matrix permits to give a different weight of consideration to the information reported in different sources. The idea of it is that, if some aspect is lacking description (such as type of diabetes, speech unit, way of recording speech material, experimental details, speaker dependency or independency of the model, or computational approach), "missing" was placed in the corresponding cell.
TABLE 8The Transparency and Reproducibility Matrix
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
This approach is radically different to the rest of the literature, since it models the speech tract with a set of differential equations in place of statistics or machine learning. The problem that
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
explicitly states that the method has a know-how and lacks a detailed experimental section. The other path is a a better understood strategy and, for example, with deep learning should give an accurate approximation (discussed in the “Future Work” Subsection).
BIASES OF DATA COLLECTION
There are application-specific issues in data collection that, unless properly addressed, can undermine the conclusions of the study.
CBGM values need to be correctly retrofitted
The options of BG measurement are either obtaining it from a drop of blood by means of a capillary finger-pricks with a lancet and then estimating a glucose value with a BG meter (BG readings) as was done in
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
Table 2. The first option is an instantaneous and reliable way to the measure glucose value. As for the second option, currently available CGM systems have needle-type sensors that are inserted in the subcutaneous system and remain there for several days. Unfortunately, it is not trivial to get the values of the glucose concentration from a CGM trace, and a retrospective “retrofitting” algorithm is needed. CGM data are usually noisy and characterized with a significant bias.
C. Cobelli, S. Del Favero, A. Facchinetti, et al., “Retrospective retrofitting method to generate a continuous glucose concentration profile by exploiting continuous glucose monitoring sensor data and blood glucose measurements”, patent US 2019/0223807.
CGM readings are affected by a physiological and a technical time delay, when compared to BG readings. The accuracy of CGM systems depend on several factors, one of which is the rate of change in BG concentrations. Inaccuracy at rapidly changing BG concentrations is, in part, caused by the fact that CGM systems measure glucose values in the interstitial tissue, not on capillary blood. A physiological time delay (of less than 10 minutes) is thus expected under such conditions. Moreover, it appears that physiological delay can work both ways: with BG changes preceding interstitial glucose changes as well as interstitial glucose changes preceding BG changes.
Further time delay is introduced through the sensor architecture, since glucose has to pass through the membranes surrounding the electrodes, and more delay is added, which is equal to the running time of the algorithm of the CGM system. Based on the above, we recommend that in the data collection protocol, it should be explicitly explained that the values in the presence of rapid glucose fluctuations were not taken and how other CGM issues were accounted for.
Subjects must be blinded to their glucose value
In the literature, the speech sample and the BG were said to be simultaneous. Indeed, technically they have the same time stamp, but that would be a nontrivial multitasking, and we assume that reading a sentence precedes prickling the finger with a lancet or vice versa. The order is of significance. Suppose the subject first takes the glucose value, introduces the numeric value in the mobile app and then reads a prompted sentence. Then, there is a danger that the resulting voice pattern of hypoglycemia is overlaid with the pattern of emotion: the subject receives bad news about a much feared condition of hypoglycemia or annoying hyperglycemia (with a need to take an insulin shot, which is obtrusive and possibly expensive). The voice patterns for the normal BG values can be mixed with relief. The resulting system can detect the affect induced by the faulty data collection design, instead of the BG.
Another connected problem stemming from the fact that the subjects were not blinded to their glucose value is a possible involuntary manipulation of the voice data, especially by the diabetic colleagues, who volunteered for the study and were familiar the working hypothesis of generalizable voice patterns in hypo- and hyperglycemia.
In all the referenced studies with BG a necessary detail of the protocol is missing, mining endocrinology research we found it expressed as follows: “The subjects were blinded to their actual glucose value”.
The research on BG estimation has its specific error measure: Parkes Error (PE). Since the clinical consequence of any particular error depends on the absolute value of both predicted and actual values and not on the percentage of deviation, from mid 1980s, statistical metrics such as accuracy or linear regression score were considered to be inadequate for reporting errors in BG estimation and evaluating new methodologies.
In PE, the risk plane, on which the X-axis is the actual BG and the Y-axis is the measured BG, is divided into eight concentric zones (Figure 1). There are five risk levels associated to all the possible errors: A is the zone corresponding to clinically accurate measurement, i.e., the one with no effect on clinical action, B is the zone corresponding to altered clinical action, but no effect on clinical outcome, C is the zone corresponding to wrong actions that are likely to affect clinical outcome, D is the zone corresponding to altered clinical action with a significant clinical risk, and E is the zone corresponding to altered clinical action with dangerous consequences. The PE is defined in a slightly different way for T1D and T2D and was based on the aggregated expert knowledge of 100 endocrinologists and refined the previous version of the same idea initially proposed in the 1980s. It has become part of the standard for BG measurement ISO 1519 2013/2018.
FIGURE 1Definition of Parkes Error, reproduced from ref.
At the moment, none of the noninvasive technologies for glucose monitoring is sufficiently reliable and convenient and their development remains to be an active research field. An interested reader can refer to Bolla and Priefer
for an overview of current and future noninvasive devices. Yet, a system implementing the fusion of the three modalities had a much better success rate than any of the constituent technologies alone.
Fusion of features from complementary sources enhances predicitive models in general, beyond pattern recognition from voice, eg. in the prediction of the biochemical activity
Also a recent patent lays emphasis on the fusion of CBGM (a method with a significant delay but accurate) and BG estimation from voice (instantaneous).
Another possibility is a fusion of invasive and noninvasive methods: based on speech analysis, alert the patient when to do an additional invasive BG measurement.
DISCUSSION
The present review has covered the literature on BG estimation from voice.
1)
All the studies on glucose estimation from voice report positive results: from a nonrandom nature of the relation between the acoustic patterns and BG value in few subjects to 98% correct estimation in 7,000 subjects. Due to the novelty of the field and the authors being unaware of the published studies (eg, several claim they are the first ones), there are five types of computational system designs that come from different research fields.
2)
Unlike other vocal biomarkers, generalization beyond one speaker in instantaneous glucose estimation can be impossible, because in endocrinology research the trend is that many aspects regarding diabetes are highly individual and average responses are useless.
3)
The community of glucose estimation from voice should be aware of the specific practices in endocrinology regarding success/error measure, when it comes to comparing different methods, as well as avoid the bias of data collection, namely, (i) the subjects must be blinded about their BG; (ii) if CBGM values are used, then BG traces need to be correctly retrofitted.
CONCLUSIONS
We have reviewed and discussed the usefulness of voice as a biomarker of glucose variability taking into consideration variations across individuals. The two basic mechanisms for the association between BG level and voice are glucose-induced changes in vocal fold elasticity and/or anxiety. The relation between glucose level and voice is nonrandom and, even if future research shows that glucose level cannot be estimated from voice alone with sufficient accuracy, there is a lot of active research on noninvasive estimation of BG with fusion of complementary technologies, and voice is one more source of information to gain additional accuracy. As a technical contribution, the article discussed the defects of the currently used protocols and catalogued the existing systems.
Future Work
In the authors’ opinion, unless the details of study
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
depending on the size of the database either train a deep neuronal network (several thousand voice samples) or a classical classification algorithm such as the Support Vector Machine or Naïve Bayes (several hundred voice samples), with appropriate problem formulation, e. g.
feature extraction, and training/testing of the classification function) are available from the first author on request.
Many factors can affect the voice, for example, fatigue, alcohol and other drugs, stress, circadian rhythm, temporary mucus, infection and so on. This phenomenon, when the distributions shift due to a global change, is referred to as “context drift.” There are diverse ways to handle this variability and the fact that a drift has happened is detectable. Also, to the extent possible, the vocal biomarker should be built with the features that are robust to context drift.
ACKNOWLEDGMENTS
The corresponding author is pleased to acknowledge being part of the “Scalable resource-efficient systems for big data analytics”funded by the Knowledge Foundation, Sweden (20140032).
AUTHOR DISCLOSURE STATEMENT
The authors declare no conflict of interest.
REFERENCES
Muaremi A.
Arnrich B.
Tröster G.
Towards measuring stress with smartphones and wearable devices during workday and sleep.
HALEF: an open-source standard-compliant telephony-based modular spoken dialog system: a review and an outlook.
in: Geunbae Lee G. Kook Kim H. Jeong M. Kim J.-H. Natural Language Dialog Systems and Intelligent Assistants. Springer International Publishing,
Switzerland2015: 53-61 (ch. 5)
Towards disorder-independent automatic assessment of emotional competence in neurological patients with a classical emotion recognition system: application in foreign accent syndrome.
Scientific solutions for the parameter’s automation in biochemical and biomechanical processes of the operational estimation of blood glucose from human voice.
Theory Pract Modern Sci.2016; 7(in Russian): 214-226
Y. Ulanovsky, A. Frolov, A. Kozlova, “Method of non-invasive determination of glucose concentration in blood and device for the implementation of thereof”, Patent WO2014/049438.
P. R. Michaelis, “Detection of extreme hypoglycemia and hyperglycemia based on automatic analysis of speech patterns”, US patent US 7, 925,508 B1, 2011.
C. Cobelli, S. Del Favero, A. Facchinetti, et al., “Retrospective retrofitting method to generate a continuous glucose concentration profile by exploiting continuous glucose monitoring sensor data and blood glucose measurements”, patent US 2019/0223807.
30334-9&id=fx1.jpg){kind=link}
30334-9&id=gr1.jpg){kind=link}