Predicting glycemic control status and high blood glucose levels through voice characteristic analysis in patients with cystic fibrosis-related diabetes (CFRD)

Cystic fibrosis-related diabetes (CFRD) is associated with reduced life expectancy in adults with cystic fibrosis (CF). Voice analysis may be a convenient method for diagnosing and monitoring CFRD. This study aims to determine the relationship between voice characteristics and markers of glucose and glycemic control and to identify if voice analysis can predict high blood glucose levels and glycemic control in adults with CFRD. We conducted a prospective cross-sectional study in adults with CF from March to December 2021. We recorded 3-second voice samples of a sustained /a/ vowel and analyzed voice characteristic using the Computerized Speech Lab with the Multi-Dimensional Voice Program. In female participants with CFRD, the noise-to-harmonic ratio was significantly lower in those with HbA1c ≥ 7. Furthermore, fundamental frequency variation was significantly lower in both male and female participants with CFRD who had a glucose level of 200 mg/dL or higher at the time of collection. This finding was also associated with a high level of point-of-care glucose. The human voice has potential as a non-invasive tool for measuring glucose levels and glycemic control status in CFRD patients in the future.

www.nature.com/scientificreports/ Changes in blood glucose levels may cause changes in the elastic properties of laryngeal soft tissue, resulting in changes in the tissue compliance based on Hooke's law of physics 7 . Studies in voice acoustic analysis have demonstrated significant differences in fundamental frequency, jitter, shimmer, and noise-to-harmonic ratio when comparing healthy and patients with diabetes 8,9 . Neurological, vascular, and muscular changes are associated with impaired glycemic control. Prior studies indicated differences in voice parameters among adults with CF who have and have not been diagnosed with CFRD 10,11 . Thus, voice acoustic analysis may potentially be used to screen for CFRD screening and to monitor home glucose values. The effects of high blood glucose concentrations on voice characteristics in patients with CFRD are poorly understood.
This study aims to determine the relationship between voice characteristics and markers of glucose and glycemic control. The secondary goal is to determine if voice analysis can predict high blood glucose levels and glycemic control.

Methods
Study participants. This was a prospective pilot study with CF at the Emory Adult CF Center recruited between March and December 2021. The study was approved by the Emory University Institutional Review Board. All participants provided signed informed consent before participating in the study. The inclusion criteria included adults with CF between the age of 18 and 70 years, a diagnosis of CF based on CFTR gene mutation or sweat chloride testing, and a diagnosis of cystic fibrosis-related diabetes (CFRD) based on a standard OGTT. The upper limit of 70 years was determined based on voice changes found in the elderly 12 . Exclusion criteria included pregnancy, breastfeeding, uncontrolled hypertension (systolic blood pressure > 140 mmHg or diastolic blood pressure > 90 mmHg), acute myocardial infarction or stroke within 6 months, history of thyroid disease, history of substance abuse, neurological disorder, active mental disorders, active smoking or former smokers who had quit smoking less than 6 months, speech disorders, history of abnormal voice such as hoarseness of voice, upper respiratory tract infection and acute pulmonary exacerbation within the previous 2 weeks. We excluded individuals with uncontrolled hypertension due to the potential change in voice parameters 13,14 and individuals with a history of thyroid disease due to potential laryngeal apparatus distortion 15 . Study protocol. Participants who met the eligibility criteria were interviewed using case record forms to collect demographic data and the duration of the disease. Relevant medical information, including results from the OGTT, hemoglobin A1c and serum glucose, was extracted from the medical record and from the participant questionnaire. Glycemic control was defined as the HbA1c level using 7% as a cutoff point on the day of the voice recording or within the previous three months.
Glucose measurements. Point-of-care blood glucose testing (POCT) was conducted at the participant's clinic visit using the OneTouch Verio Flex glucometer (Malvern, Pennsylvania). Capillary glucose measurement was performed by pricking a sterile lancet on one of the participant's fingertips with a sterile lancet to obtain a drop of blood for use in the glucometer.
Voice recordings. During an outpatient clinic appointment, the participant was requested to produce a sustained vowel, /a:/, at their usual frequency and strength after taking a deep breath, in order to obtain maximum phonation time without using expiratory reserve air. Patients were seated in a soundproof room with a headmounted mouth microphone positioned 4 cm from the lips and 45 to 90 degrees from the front of the mouth. All voice datasets were amplified and analog-to-digital conversed via USBPre2 (Sound Devices LLC, Reedsburg, Wisconsin). Using voice recorder software (Apple, Cupertino, California), voice samples are saved as .wav files. An initial 3.5 s recording of the vowel /a:/ emission was collected for voice acoustic analysis; the commencement of emission was omitted to prevent interference from voice attachment on data processing. Voice acoustic assessment. Acoustic parameters were analyzed using Computerized Speech Lab (CS) model 4500 (PENTAX Medical, New Jersey, USA). CSL is considered the gold standard system for acoustic analysis and has been validated to assess voice pathology in many controlled trials 9,16 . The Multi-Dimensional Voice Program (MDVP), a computerized voice analysis system, in conjunction with CSL, is a versatile voice-processing and spectrographic analysis software package ideally suited for use in determining acoustic parameters including fundamental frequency (F0), fundamental frequency variation (vF0), jitter, shimmer, relative average perturbation (RAP), and noise-to-harmonic ratio (NHR). F0, measured in Hertz (Hz), is defined as the number of times a sound wave produced by the vocal cords repeats during a given time period. It is also the number of cycles of opening/closure of the glottis. vF0, measured in %, represents the relative standard deviation of the period-to-period calculated F0.
Jitter, measured in %, refers to frequency variation from cycle to cycle. It is affected mainly by the lack of control of the vibration of the vocal cords. Shimmer, measured in %, relates to the amplitude variation of the sound wave. It changes due to the reduction of glottal resistance and mass lesions on the vocal cords and is correlated with the presence of noise emission and breathiness 17 . RAP, measured in %, is a quantitative measure of the voice while producing a sustained vowel. It minimizes the effects of slow changes in the F0 by averaging across three successive vibratory cycles to provide an estimate of the value of the middle period had there been no perturbation present 17 . NHR is a general evaluation of noise present in vocalization. It is an assessment of the ratio between the non-periodic components, which are the glottal noise, and periodic components, which are the vibration of the vocal cords. The quality of the recording of each patient's voice, namely the sustained /a:/, was blindly evaluated by a speech therapist (NK). www.nature.com/scientificreports/ Statistical analysis. All study data were checked for normality and presented as mean (SD) if they were normally distributed, or median (interquartile range, IQR) if they were not normally distributed. For baseline characteristic data in Table 1, the independent t test or Mann-Whitney U test was used to compare all continuous variables and the Chi-square test was used to compare categorical variables between the two groups. Multivariate analysis and logistic regression analysis were performed to identify independent variables associated with CFRD. A two-tailed p-value of less than 0.05 was considered statistically significant. Statistical analyses were performed using Stata version 15 (StataCorp LLC, College Station, Texas).

Results
Demographics of study participants. A total of 43 participants were recruited in this study. There were 25 participants with CFRD who had an HbA1c < 7% and 18 participants with CFRD who had an HbA1c ≥ 7%. Baseline characteristics are presented in Table 1. There were no significant differences in the participant demographics between the two groups. The participants with CFRD who were classified in the group of HbA1c < 7% had a mean HbA1c level of 5.9 ± 0.4% and a mean POCT glucose level of 123 ± 33 mg/dL. As expected, the participants with CFRD who were classified in the group of HbA1c ≥ 7% had a mean HbA1c level of 9 ± 2.2% and a mean POCT glucose level of 183 ± 87 mg/dL.
Voice characteristics as a marker of glycemic control. When stratified by sex in Table 2, female participants with CFRD who had an HbA1c ≥ 7% had a significantly lower NHR than female participants with CFRD had an HbA1c < 7% (0.13 ± 0.02 vs 0.15 ± 0.03, p = 0.049). However, there was no significant difference in male participants with CFRD regarding glycemic control status. We performed multivariate linear regression analysis in Table 3 to determine NHR as a dependent variable, was significantly associated with glycemic control (p = 0.041) and female sex (p = 0.047). NHR decreased by 0.0058 for every 1% increase in HbA1c in participant To determine vF0 and RAP as dependent variables significantly associated with glucose level, we conducted multivariate linear regression analysis in Table 6. Data showed that the duration of CFRD and POCT glucose levels were statistically related to vF0, p = 0.019 and p = 0.012 respectively. However,

Discussion
The principal goal of this study was to determine whether voice parameters differed by glycemic control status and high blood glucose level in people with CFRD. We found that, in female participants with CFRD, NHR was significantly lower in those who had an HbA1c ≥ 7%. NHR was also associated with HbA1c level when adjusted for age, BMI, duration of diabetes diagnosis, and glucose level. In addition, vF0 was significantly lower in both male and female participants with CFRD who had a glucose level of 200 mg/dL or higher at the time of collection. The vF0 value was also associated with a high level of point-of-care glucose. Other voice parameter changes including F0, Jitter, Shimmer and VTI were not significantly correlated with HbA1c level and glucose level in adults with CFRD.
There are a number of potential mechanisms for the differences in NHR in females with CFRD regarding their HbA1c level. NHR has been reported to assess vocal fold behavior 17 . Studies have shown diabetes-related decrements in laryngeal function in the phonatory apparatus in diseases such as myopathy and neuropathy 18,19 . Glycemic status may lower NHR by contributing to instability in vocal fold vibration and thus interfering with the periodicity of vibration 20 . Changes in voice parameters may be more prominent in females compared to males with CFRD. There is no clear mechanism why females with CFRD have prominent changes in voice parameters compared to males. A study in adults without CF demonstrated that laryngeal geometry, hormone differences, and behavior characteristics of women were prone to have more voice disorders than men 21 . Changes in voice characteristics including vF0 and RAP mainly reflect laryngeal muscle behavior which differs between sexes. vF0 has been used to measure the ability to keep the laryngeal muscles in a fixed position for vowel prolongation 22 while RAP has been used to measure the intrinsic stability of the voice in terms of contraction of laryngeal muscles during sound production 17 . High glucose levels may decrease the strength of intrinsic laryngeal muscles can result in a decrease in vF0 and RAP in our study. Moreover, poor nutrition due to malabsorption in CF may cause decreased laryngeal muscle mass leading to laryngopharyngeal tract hypomobility. Our study did not find any association between BMI and HbA1c level well as BMI and glucose level in participants with CFRD. Although BMI grossly represents nutritional status, the measurement for laryngeal muscle mass needs to be examined in future studies to confirm this hypothesis.
Changes in other voice characteristics have been studied in participants with CF. One particular characteristic is the voice turbulence index (VTI) 17 which measures the relative energy level of high-frequency noise. This characteristic correlates with the turbulence caused by incomplete or loose adduction of the vocal folds and Table 6. Multivariate linear regression analysis with fundamental frequency variation (vF0) and relative average perturbation (RAP) as outcomes. Significant values are in bold.  23 . Our prior study focused on detecting CFRD using changes in voice characteristics and showed significantly higher VTI in males with CFRD compared to male CF patients without CFRD 10 whereas this study demonstrated that there is no significant difference of VTI in individuals with CFRD according to glycemic control status or POCT glucose level. This can be explained by gastrointestinal acid reflux in individuals with CFRD leading to laryngopharyngeal reflux 24 . However, there is no specific sex differences in prevalence of acid reflex endoscopy-based study in CFRD.
The number of studies on the voice characteristic analysis in non-CF diabetes is limited. Three studies in adults with type 2 diabetes had mixed results. A study in India demonstrated a reduction of NHR, RAP and shimmer in participants with diabetes compared to participants without diabetes 8 . A study in Thailand showed significantly lowered F0 in females with diabetes and no change was detected in RAP as compared to females without diabetes 9 However, the study in Turkey found a significantly higher of absolute Jitter in participants with diabetes compare to healthy volunteer. There is also a significant lowering maximum phonation time in individuals with diabetes neuropathy 25 . There is only one study in pediatric individuals between 6 and 18 years that found no significant difference in acoustic voice quality index in children with type 1 diabetes compared to age and sex-matched healthy participants 26 . Further study on voice characteristic analysis are needed to understand more about effect of diabetes in voice.
Developing this new voice technology may improve screening for CFRD. One advantage of this technology is that glycemic control can be indirectly associated with the voice as a non-invasive technique. Although not studied directly in this study, future applications could involve the use of this technology to approximate realtime markers of glycemic control without requiring an invasive procedure such as a finger prick. Other potential applications include monitoring of long-term CFRD glycemic status, which would save the time in clinic and/ or the need for venipuncture for a patient. Also, this technique can be used in telehealth visits, which further provides another potential use of this technology.
This study of voice characteristic analysis provides an initial step in the understanding of the physiology of voice changes occurring in patients with CFRD. Several studies show voice signals have the potential to be used to predict abnormal blood glucose levels in non-CF diabetes 27,28 . Our previous study in deep learning-based voice technique for CFRD screening demonstrated the potential of employing artificial intelligence to analyze this data 29 . Future directions of this work include utilizing artificial intelligence methods, developing a mobile application for monitoring glucose levels and long-term glycemic control in CFRD, and assessing the technological feasibility and the cost of the equipment.

Limitations
Our study provides important data regarding the relationship between voice characteristics and glycemic control in adults with CF; however, there are a few limitations. One limitation is the small number of samples in the different glycemic control categories which may not enough to detect the difference between the range of HbA1c level Future studies need to have a larger number of subjects in different categories of glycemic control levels to have sufficient statistical power to determine the relationship with various voice characteristics. Our study may not be generalizable to other forms of diabetes as we only focused on CFRD. Another limitation is the lack of laryngeal imaging which would be useful in interpreting the results of the study. To better understand the functionality of vocal folds, a laryngeal imaging technique like videostroboscopy could be a part of a future study. Finally, the study was done in a soundproof room under extremely controlled conditions. Future studies need to focus on voice recordings performed in an ambient environment to develop the practical clinical application for this technology.

Conclusions
We found differences in voice parameters that distinguished normal glycemic control from inadequate and poorly controlled CFRD in patients with CF. There is potential to use the human voice as a non-invasive tool for glycemic control assessment in patients with CFRD. NHR was significantly associated with poor glycemic control status while vF0 was associated with high point-of-care glucose levels in participants with CFRD. Larger sample sizes examining voice characteristics and laryngeal imaging techniques are warranted to further investigate the feasibility of using voice as a biomarker for glycemic control in CFRD.