Voice parameters for difficult mask ventilation evaluation: an observational study.

BACKGROUND: Mask ventilation (MV) is an essential component of airway management. Difficult mask ventilation (DMV) is a major cause for perioperative hypoxic brain injury; however, predicting DMV remains a challenge. This study aimed to determine the potential value of voice parameters as novel predictors of DMV in patients scheduled for general anesthesia.
METHODS: We included 1,160 adult patients scheduled for elective surgery under general anesthesia. The clinical variables usually reported as predictors of DMV were collected before surgery. Voice sample of phonemes ([a], [o], [e], [i], [u], [ü], [ci], [qi], [chi], [le], [ke], and [en]) were recorded and their formants (f1-f4) and bandwidths (bw1-bw4) were extracted. The definition of DMV was the inability of an unassisted anesthesiologist to ensure adequate ventilation during MV under general anesthesia. Univariate and multivariate logistic regression analyses were used to explore the association between voice parameters and DMV. The predictive value of the voice parameters was evaluated by assessment of area under the curve (AUC) of receiver operating characteristic (ROC) curves of a stepwise forward model.
RESULTS: The prevalence of DMV was 218/1,160 (18.8%). The AUC of the stepwise forward model (including o_f4, e_bw2, i_f3, u_pitch, u_f1, u_f4, ü_bw4, ci_f1, qi_f1, qi_f4, qi_bw4, chi_f1, chi_bw2, chi_bw4, le_pitch, le_bw3, ke_bw2, en_pitch, and en_f2, en_bw4) attained a value of 0.779. The sensitivity and specificity of the model were 75.0% and 71.0%, respectively.
CONCLUSIONS: Voice parameters may be considered as alternative predictors of DMV, but additional studies are needed to confirm the initial findings. 2021 Annals of Translational Medicine. All rights reserved.

Introduction

Mask ventilation (MV) is a crucial aspect of airway management in general anesthesia (1). Successful MV is extremely important to maintain airway patency for optimal oxygenation of the tissues. Ability to ventilate the patient with a mask is also a core rescue skill during difficult or failed tracheal intubation situations (2). Difficult mask ventilation (DMV) is defined as a situation in which it is not possible for the anesthesiologist to provide adequate ventilation, due to inadequate mask seal, excessive gas leak, or excessive resistance to the ingress or egress of gas, according to the American Society of Anesthesiologists (ASA) Task Force on Management of the Difficult Airway (3). The reported incidence of DMV varies widely (0.08–15%) due to the lack of a precise and objective definition (4-9). It is believed that DMV is the major contributing factor for severe anesthesia-related complications including death or hypoxic brain injury (10,11). Therefore, the accurate prediction of DMV is of significant importance. Previous studies have reported that age >55 years, body mass index (BMI) >26 kg/m2, presence of a beard, lack of teeth, and history of snoring are 5 independent risk factors for DMV (4,6,7,12). In addition, previous neck radiation, thick or short neck, modified Mallampati class 3 or 4, and limited mandibular protrusion are also known as predictors for DMV (13). Unfortunately, the diagnostic accuracy of the above predictors has remained poor. In a large cohort study, DMV was unanticipated in 808 of 857 (94%) cases (14). Nørskov et al. (15) developed a scoring system [named the difficult facemask (DIFFMASK) score] based on 13 different predictive factors of DMV, which yielded a sensitivity of 85% and specificity of just 59%. Thus, further studies are needed to investigate accurate predictors for DMV.

The human voice is complex signal of sound emanating from our vocal cords, which is not only a primary mean of communicating with each other, but also contains a large amount of characteristic information of the human body (16). Prior studies have suggested that voice parameters can reflect anatomical features of the upper airway, including the opening of the mandible, the length of mandible, maxilla, and the laryngeal tube, and the volume of the oral and pharyngeal cavities (17,18). Some studies have found that patients with obstructive sleep apnea (OSA; a disease associated with DMV) show significant changes in acoustic parameters compared to healthy controls, and acoustic analysis showed potential in the assessment of OSA (19-21).

Formant is one of the most important parameters in the acoustic analysis, representing the vocal tract resonance, and relating to the upper airway anatomy (18,19). At present, only studies reported by de Carvalho et al. (22,23) have provided some evidence supporting an association between formant frequencies and difficult airway. Considering the influence of different races on pronunciation manner and the incidence of DMV, and the scarcity of literature on the voice features used for MV evaluation, we conducted this study in the Chinese population to confirm the initial findings by de Carvalho et al. (22), and to further clarify the potential value of voice analysis of both vowels and consonants to predict DMV. We hypothesized that voice features may be considered an alternative valuable predictor, in addition to already known factors for DMV in patients scheduled for general anesthesia. We present the following article in accordance with the STROBE reporting checklist (available at https://dx.doi.org/10.21037/atm-21-6274).

Methods

Patients

This observational study was conducted between December 2020 and July 2021 after obtaining approval from the Ethics Committee of Shanghai Ninth People’s Hospital (No. SH9H-2020-T386-1). The protocol is registered in ClinicalTrials.gov (trial registration No. NCT 04458220). The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013).

The inclusion criteria for the study were patients scheduled for elective surgery under general anesthesia, aged 18 years and above and native Mandarin speakers. The exclusion criteria were as follows: ASA physical status class IV; contraindications for MV (rapid sequence induction, awake intubation in the case of expected difficult airway); participants with a previous history of speech disorders, vocal cord disease, cleft palate, craniofacial syndromes, hearing or perception abnormalities, mental or central nervous system disease, or those who had participated in other relevant clinical investigation in the past 3 months. Informed consent was provided by each participant before their inclusion.

Preoperative airway assessment

During the preoperative visit, the demographic properties of age, gender, weight, height, and body mass index (BMI) and the following information were recorded: lack of teeth, presence of beard, history of snoring, history of neck radiotherapy, inter-incisor gap (IIG), modified Mallampati test (MMT), thyromental distance (TMD), mandibular protrusion, neck circumference (NC), and cervical spine mobility. To reduce measurement bias, all researchers received repeated educational sessions prior to this trial.

The IIG refers to the maximal distance between the upper and lower incisors, which is used to assess the mouth opening capacity. The MMT was performed with the patient seated and their head in full extension, tongue out, and without pronunciation (24). The TMD refers to the distance between the uppermost border of the thyroid cartilage and the mentum, and was measured with the neck extended and the mouth closed (25). Mandibular protrusion assesses the range of movement of the mandible by asking patients to move their lower teeth past their upper teeth (25). The NC was measured at the level of the cricothyroid membrane. Cervical spine mobility (25) was measured by asking the patient to fully extend and flex the head and neck, and classified as grade 1 >90°, grade 2 =90°, or grade 3 <90°.

Voice data and analysis

Recording

The voice samples were selected from the Chinese phonetic alphabet (https://github.com/Samurais/DaCiDian). All participants were asked to phonate the Chinese vowels: [a], [o], [e], [i], [u], and [ü]; consonants: interdentals [ci], post-alveolar [chi], alveolar [le], alveo-palatals [qi], velars [ke]; and nasal vowel [en] (based on the anteroposterior position of the tongue and the location of air friction distributed along the vocal tract). All recordings were performed with the participant in a seated position in a quiet room in comfortable conditions. A 16-bit hand-held recorder with a sampling frequency of 44.1 kHz (Sony, ICD-PX470) was used to record the participants’ speech. A distance of 5 cm was maintained between the mouth and the recorder for each recording. The duration of each sound and the interval between sounds was approximately 1 second.

Voice parameters abstraction

The voice samples were first identified and aligned using a Kaldi chain (26) model trained on the AISHELL-2 (27) dataset, which enabled the exact position of syllable in the speech samples to be determined. Conventionally, formant frequencies refer to more intense frequencies within the envelope of the spectrum, which are computed based on the number of energy maxima per period, and bandwidth is measured as the width of a formant between 2 points that are 3 dB below the peak on either side of it. In this study, the Praat software program (https://www.praat.org/) was used to extract the pitch (fundamental frequency), the first 4 formant frequencies (f1–f4), and bandwidths (bw1–bw4). For all the above features, the mean value along with the time of the clip where the syllable was located was taken as the final result. This voice analysis was conducted by a specialist in speech technology who was blinded to the airway evaluation and MV classifications.

Induction of anesthesia and MV evaluation

An anesthesiologist with more than 3 years of experience was assigned to airway management. During induction of anesthesia, the patient’s head and neck were placed in a sniffing position. Preoxygenation of each patient was performed for 4 min by mask with 100% O2. Each patient was routinely monitored during the whole procedure by electrocardiography, peripheral oxygen saturation (SpO2), and end-tidal carbon dioxide. General anesthesia was induced with a combination of midazolam 0.05 mg/kg, fentanyl 2–4 µg/kg, propofol 2–2.5 mg/kg, and rocuronium 0.6 mg/kg. Evaluation of MV was made after induction of anesthesia. According to Langeron et al., MV is considered difficult in the case of: (I) inability for the unassisted anesthesiologist to maintain oxygen saturation >92% using 100% oxygen and positive pressure ventilation, (II) important gas flow leak, (III) necessity to increase the gas flow to greater than 15 L/min and to use the oxygen flush valve more than twice, (IV) no perceptible chest movement, (V) performing a 2-handed MV technique, and (VI) operator change.

Statistical analysis

Measurement data were expressed as mean ± standard deviation, and categorical variables were expressed as frequency (%). The hypothesis was tested using one-way analysis of variance (ANOVA), the Mann-Whitney U test, and Fisher’s exact probability method. Univariate and multivariate logistic regression models were established using the general linear model to evaluate the relationship between the voice features and DMV grade. Covariables for multivariate regression including gender (male, female), age (as a continuous variable), height, weight, BMI, IIG, NC, and snoring history. Stepwise logistic model and the ROC curve was used to analyze the predictive value of voice parameters to predict DMV, expressed as the AUC with its 95% confidence interval (95% CI). A P<0.05 indicated statistical significance. The dataset was randomly split into the training and testing sets at a ratio of 8:2, and the random seed was 2021. The data analysis was performed using the R project software program (R 4.0.4; https://cran.r-project.org/bin/windows/base/old/4.0.4/).

Results

Baseline characteristics

A total of 1,270 eligible patients were initially recruited. After excluding cases with missing data related to voice parameters and with contraindications for MV, 1,160 cases were subjected to statistical analysis (). A summary of perioperative baseline characteristics by group is presented in . Difficulties of MV were encountered in 218 of 1,160 (18.8%) participants. Between 2 groups, statistically significant differences were revealed in age, gender, BMI, MMT, NC, mandibular protrusion, cervical spine mobility, history of snoring, and neck radiotherapy, while no difference were observed in IIG, TMD, beard, and no teeth.

Figure 1

Flow chart of the study.

Table 1

Baseline demographic characteristics and preoperative airway assessment variables of patients scheduled for general anesthesia

Risk factors	Overall (n=1,160)	Easy MV (n=942)	Difficult MV (n=218)	P value
Age (years)	34.62±12.72	33.18±11.79	40.81±14.62	<0.001
Gender				<0.001
Female	702 (60.5)	634 (67.3)	68 (31.2)
Male	458 (39.5)	308 (32.7)	150 (68.8)
BMI (kg/m2)	21.86±3.30	21.34±3.02	24.12±3.50	<0.001
Snoring history				<0.001
Yes	447 (39.0)	309 (33.3)	138 (63.9)
No	698 (61.0)	620 (66.7)	78 (36.1)
Neck radiotherapy				0.001
Yes	11 (1.0)	4 (0.4)	7 (3.2)
No	1,144 (99.0)	934 (99.6)	210 (96.8)
Beard				0.464
Yes	3 (0.3)	2 (0.2)	1 (0.5)
No	1,155 (99.7)	939 (99.8)	216 (99.5)
No teeth				0.163
Yes	4 (0.3)	2 (0.2)	2 (0.9)
No	1,153 (99.7)	937 (99.8)	216 (99.1)
IIG (cm)	4.06±0.98	4.06±0.97	4.03±1.03	0.656
TMD (cm)	9.36±1.61	9.38±1.61	9.28±1.65	0.401
NC (cm)	34.85±3.84	34.14±3.47	37.89±3.88	<0.001
MMT				0.001
1	358 (31.2)	311 (33.4)	47 (21.8)
2	274 (23.9)	226 (24.3)	48 (22.2)
3	395 (34.5)	308 (33.1)	87 (40.3)
4	119 (10.4)	85 (9.1)	34 (15.7)
Mandibular protrusion				0.010
1	1,045 (90.9)	860 (92.2)	185 (85.6)
2	81 (7.0)	56 (6.0)	25 (11.6)
3	23 (2.0)	17 (1.8)	6 (2.8)
Cervical spine mobility				<0.001
1	1,121 (98.0)	919 (98.9)	202 (94.0)
2	21 (1.8)	9 (1.0)	12 (5.6)
3	2 (0.2)	1 (0.1)	1 (0.5)

Data are shown as number (percentage) or mean ± SD. MV, mask ventilation; SD, standard deviation; BMI, body mass index; MMT, modified Mallampati test; NC, neck circumference; IIG; inter-incisor gap; TMD, thyromental distance.

Univariate and multivariate logistic regression

Univariate logistic regression analysis revealed that 69 voice parameters differed significantly between the difficult group and easy group with P values <0.05 (Table S1).

Univariate logistic regression analysis of clinical airway assessment variables and DMV () showed that age, gender (male), BMI, snoring history, neck radiotherapy, NC, MMT (grade 3/4), mandibular protrusion (grade 2), and cervical spine mobility (grade 2) were significantly associated with DMV (P<0.01).

Table 2

Univariate logistic regression analysis of airway assessment variables and DMV

Variables	OR	95% CI	P value
Age (years)	1.043	1.032–1.055	<0.001
Gender
Female	Reference
Male	4.541	3.320–6.266	<0.001
BMI (kg/m2)	1.290	1.230–1.355	<0.001
Snoring history	3.550	2.612–4.854	<0.001
Neck radiotherapy	7.783	2.329–29.941	0.001
Beard	2.174	0.101–22.792	0.527
No teeth	4.338	0.518–36.319	0.143
IIG (cm)	0.966	0.832–1.124	0.655
TMD (cm)	0.961	0.877–1.053	0.400
NC (cm)	1.295	1.241–1.353	<0.001
MMT
1	Reference
2	1.405	0.907–2.179	0.127
3	1.869	1.273–2.772	0.002
4	2.647	1.595–4.367	<0.001
Mandibular protrusion
1	Reference
2	2.075	1.244–3.377	0.004
3	1.641	0.585–4.009	0.304
Cervical spine mobility
1	Reference
2	6.066	2.534–15.045	<0.001
3	4.550	0.179–115.369	0.285

DMV, difficult mask ventilation; OR, odds ratio; CI, confidence interval. BMI, body mass index; MMT, modified Mallampati test; NC, neck circumference; IIG; inter-incisor gap; TMD, thyromental distance.

All the significant risk factors with P values <0.05 found by univariate analysis were subjected to multivariate analysis to identify independent predictors for DMV. shows that 9 voice parameters (o_f3, e_f3, e_f4, i_f2, i_f3, u_f4, ci_f3, ci_f4, en_f3) were independently related to DMV (P<0.05), after adjusting for gender, age, height, weight, BMI, IIG, NC, and snoring history.

Table 3

Multivariable logistic regression analysis of voice parameters and DMV

Voice parameters	OR	95% CI	P value
o_f3	0.999	0.997–1.000	0.031
e_f3	0.999	0.997–1.000	0.019
e_f4	0.999	0.998–1.000	0.040
i_f2	0.999	0.998–1.000	0.019
i_f3	0.998	0.997–0.999	0.002
u_f4	0.998	0.997–1.000	0.009
ci_f3	0.998	0.997–1.000	0.018
ci_f4	0.998	0.997–1.000	0.023
en_f3	0.999	0.998–1.000	0.029

DMV, difficult mask ventilation; OR, odds ratio; CI, confidence interval.

Establishment of the assessment model

We used forward stepwise logistic regression for the whole sample to further construct the predictive model for DMV. The 20 selected voice parameters (including o_f4, e_bw2, i_f3, u_pitch, u_f1, u_f4, ü_bw4, ci_f1, qi_f1, qi_f4, qi_bw4, chi_f1, chi_bw2, chi_bw4, le_pitch, le_bw3, ke_bw2, en_pitch, en_f2, en_bw4) are presented in .

Table 4

The forward stepwise model of voice parameters for predicting DMV in patients scheduled for general anesthesia

Voice parameters	OR	95% CI	P value
o_f4	1.002	1.000–1.004	0.015
e_bw2	0.999	0.997–1.000	0.032
i_f3	0.998	0.997–0.999	0.001
u_pitch	0.989	0.980 –0.999	0.026
u_f1	0.996	0.993–0.999	0.004
u_f4	0.998	0.997–1.000	0.023
ü_bw4	1.001	1.000–1.003	0.036
ci_f1	0.995	0.992–0.998	0.001
qi_f1	1.004	1.001–1.007	0.006
qi_f4	0.999	0.997–1.000	0.070
qi_bw4	1.002	1.000–1.003	0.032
chi_f1	1.004	1.002–1.007	0.002
chi_bw2	0.999	0.997–1.000	0.055
chi_bw4	0.999	0.997–1.000	0.021
le_bw3	0.999	0.998–1.000	0.096
le_pitch	1.014	1.007–1.023	< 0.001
ke_bw2	1.002	1.000–1.003	0.022
en_pitch	0.987	0.977–0.997	0.010
en_f2	0.999	0.998–1.000	0.127
en_bw4	0.999	0.998–1.000	0.046

DMV, difficult mask ventilation; OR, odds ratio; CI, confidence interval.

The ROC curve analysis showed that the AUC of the forward stepwise model for predicting DMV was 0.779 (95% CI: 0.710–0.848) (), The highest point of the Youden index was designated as the threshold, which revealed the sensitivity and specificity of 71.0% (95% CI: 63.5–86.5%) and 75.0% (95% CI: 63.9–77.6%), respectively.

Figure 2

ROC curve analysis of the forward stepwise model of voice parameters for predicting DMV. ROC, receiver operating characteristic; DMV, difficult mask ventilation.

Discussion

Ventilation is frequently initiated manually with bag and face mask in the delivery of general anesthesia, which is the most basic, yet the most essential technique of ventilation before endotracheal intubation or insertion of any airway device. However, the prediction of DMV is still poorly addressed in current airway management research (12,14,28). Meanwhile, previous studies have focused primarily on physical signs of difficult airway without analyzing changes in voice. The main result of this study was that voice features can have a role in the prediction of DMV. The experimental results showed that both formant frequencies and bandwidths can make significant contributions to the evaluation of MV difficulty.

Regarding formant frequencies, our results showed that the first 4 formants presented a correlation with DMV. Formant frequencies are the distinguishing frequency components of human articulation that are dependent on the shape and length of the vocal tract in addition to the movement of the articulators (29,30). The OSA, known as an independent risk factor for DMV, has been revealed to display obvious alterations in formant frequencies due to the deposition of redundant tissues in the upper airway and reduced posterior airway space behind the base of the tongue (19,21). Macari et al. (18) previously reported a significant negative relation between the formant frequencies F3 and F4 and the length of the mandible and maxilla. This may be the underlying theoretical basis for the relationship between formant frequencies and MV difficulty.

When considering the results for bandwidths, we found there is an association between bw2/bw3/bw4 and DMV. Based on the pronunciation process, excessive tissue compliance would tend to increase the sound damping within the vocal tract. The measurement of formant bandwidths would provide an estimate of vocal tract compliance (20). The most common pathogenesis of the DMV is excessive pharyngeal compliance which would be expected to result in upper airway collapse or obstruction occurring after induction of general anesthesia (2). Results in the study by Robb et al. (20), which were extracted from the analysis of sustained phonations of vowel sounds, showed a significantly wider formant bandwidth in OSA patients compared with non-OSA participants, which provides an anatomic interpretation for our results.

Similar to other studies, our study also suggested that age, gender, BMI, and snoring are risk factors related to DMV. These risk factors also significantly affect the pronunciation process. Different speech features have been reported to be correlated with age, gender, height, and weight, due to the changes in the anatomy and physiology of the speech production system. Most reports have agreed that formant frequencies are negatively correlated with age and body size (31-33). For obese patients, abnormal fat deposition in the upper airway may interfere with voice production (34). Moreover, previous studies have provided evidence for an association between OSA and abnormal or particular speech features, which has mainly been attributed to alterations in upper airway anatomy such as craniofacial abnormalities, dental occlusion, relaxed pharyngeal soft tissues, large tongue base, and so on (19,21). Hence, when evaluating the association between voice parameters and DMV, we adjusted for age, gender, height, weight, BMI, IIG, NC, and snoring history in the multivariate regression. The results showed that only formants of [o], [e], [i], [u], [en], and [ci] were independently correlated with DMV.

Our study differs from the previous work by de Carvalho et al. (22). To our knowledge, this study was the first attempt to analyze the association between voice features and DMV in a Chinese population. There are discernably different cranial and facial structures between Western and Asian populations, which would lead to variations in vocal tract structure and pronunciation manners. Our results showed the applicability of acoustic evaluation in the prediction of DMV among Chinese participants. Furthermore, our speech samples consisted of both vowels and consonants. Vowels are the most frequently discussed syllables in the literature, which are produced without obvious tongue movement, representing resonances of the vocal tract, whereas the consonants are primarily determined by the relative positions of articulator including the tongue, teeth, lip, and soft palate (35,36). Our results demonstrated that both vowels ([o], [e], [i], [u], [ü], and [en]) and consonants ([ci], [chi], [le], [qi], and [ke]) are important predictors for DMV. Up to now, there is no universal definition and detection equipment for DMV, because DMV is not a disease, neither is it just one particular anatomical characteristic of patient physiognomy, it is a clinical situation mostly requiring a clinician’s judgment. Considering that ASA Task Force’s general definition (3) was vague and partly subjective, while the definition by Han et al. (37) was too stringent that it may result in underestimating the incidence of clinically significant difficult ventilation, in this study we rated MV according to Langeron et al. (4). Thus, different DMV definitions will lead to variations in incidence among different studies, and also affect the comparability of the model performance.

There were still some deficiencies in this study. The results presented in this paper were limited to the voices of Chinese speakers, so comparisons with other languages will require a more careful analysis of language-dependent acoustic traits. Another limitation in our study was that all patients’ voice were recorded in the sitting position. It is known that the patients were mostly placed in a supine position during induction of anesthesia, therefore changes of voice parameters measured in 2 distinct positions should be considered in further study to improve the accuracy of the predictive model. Moreover, our study was conducted in the context of elective intubation in the operating room, and voice recording requires the patients to be awake and cooperative, thus our method cannot be used to predict difficult airway in critically ill patients and in the case of emergency surgery. As the study was conducted on a general adult population, further studies including a varied age groups are needed to determine whether the relationship between DMV and voice parameters found in this study applies to other groups. In terms of technical factors, values of acoustic measures may vary with the software used. In these cases, other acoustic software programs would have to be used to analyze voice data.

In summary, the voice parameters may prove valuable information in MV evaluation of patients scheduled for general anesthesia. Compared to other currently used methods, voice parameter-based methods are promising as they are nonintrusive and can work as an automated and objective method to simultaneously analyze several voice parameters and thus may promote the development of and telematic airway evaluation. However, external, prospective validation is needed to fully realize the applicability of this novel approach in the assessment of difficult airway.

The article’s supplementary files as