Halitosis, a widespread health concern, has been extensively documented by researchers across the world. Recent studies have indicated a prevalence ranging from 22% to 50% [1] with an estimated incidence of 31.8% (95% confidence interval 24.6%–39.0%) [2], suggesting a growing trend of halitosis-related issues worldwide. This condition bears significant social and psychological repercussions for affected individuals. Currently, the assessment of causative agents is pivotal in halitosis examination, with volatile sulfur compounds (VSCs) identified as key contributors to unpleasant oral odors. The three primary VSCs in the oral cavity—hydrogen sulfide (H2S), methyl mercaptan (CH3SH), and dimethyl sulfide [(CH3)2S]—serve as reference substances in halitosis tests [3]. In dental clinics and university hospitals, organoleptic test and instrumental measurements constitute the highest standard for halitosis assessment in general clinical practice [4]. However, organoleptic test is susceptible to examiner fatigue and variability in judgment criteria due to mental and health conditions [5]. By contrast, gas chromatography (GC-MS) and sulfide monitors, which are commonly employed as instrumental tools, offer an objective and quantitative approach to halitosis assessment [6]. Nevertheless, instrumental methods possess certain drawbacks—GC-MS, being sizable and expensive, and persisting issues in both procurement and maintenance. Additionally, the complexities in equipment startup procedures further limit their widespread adoption. Consequently, the availability of facilities equipped with GC-MS is constrained. By contrast, sulfide monitors, which are smaller and more economical, present a viable alternative. However, their efficacy is compromised compared to GC-MS considering their inability to differentiate between gas types or react with components other than VSCs [6–8]. Some instruments integrate GC-MS and sulfide monitoring by combining a sulfide monitor with a gas component-separating column. This integration helps achieve both miniaturization and accuracy. Nonetheless, these devices are not sufficiently compact for chairside measurements in dental offices. Moreover, their operational duration exceeds that of standard sulfide monitors, presenting a notable disadvantage [9].

The primary challenges in current instrumental halitosis measurement can be categorized into three aspects: size, detection capability, and accuracy. To address these challenges, we have developed an AI-based, compact, and high-precision device capable of measuring three key VSCs similarly to GC-MS. We have developed an AI-based halitosis measurement device named 'Kunkun dental' to address these challenges. Our objective is to tackle the miniaturization issue by crafting a measurement apparatus utilizing a compact semiconductor gas sensor. In terms of accuracy, we have identified two key aspects: the inability to differentiate between gas types and reactivity to components other than VSCs. Kunkun dental is equipped with multiple gas sensors, each providing waveform data reflecting its reaction to odor components. These waveform datasets are then utilized, and multiple features are extracted and fed into a machine learning model to determine VSC concentration. This model was developed based on data from over 200 halitosis samples, and it incorporates waveform data from Kunkun dental alongside precise VSC measurements obtained through GC-MS analysis (see figure 1) [10]. The device's accuracy is bolstered by amalgamating numerous features acquired from the four sensors and leveraging AI technology, thereby overcoming the limitations of gas sensors, which solely capture gas intensity information. Moreover, non-VSC odors commonly found in the oral cavity, such as food and tobacco, are isolated and incorporated into the training data to prevent interference with concentration estimations. Furthermore, when these components are detected at high levels, which may have a significant impact on the VSC concentration estimates, the results are flagged as errors to counter the issue of reactivity to non-VSC components. However, it is worth noting that the accuracy of VSC concentration detection in Kunkun dental samples is yet to be verified in comparison to the highly precise GC-MS method utilized for halitosis diagnosis in university hospitals.

Figure 1. Construction process of AI judgment system in Kunkun dental.

Download figure:

Standard image High-resolution image

The aim of this study was to assess the accuracy of Kunkun dental by analyzing the breath of halitosis patients using both Kunkun dental and GC-MS. We sought to evaluate the agreement in the three VSC concentrations mentioned earlier.

2.1. Subject

The study participants were halitosis patients aged 20 years or older who visited the Fresh Breath Clinic of Tokyo Medical and Dental University Hospital between October 2022 and March 2023. Halitosis patients are defined as individuals who emit an unpleasant odor through the oral or nasal cavity that exceeds socially acceptable levels and may cause discomfort to others. This definition applies regardless of whether the cause is oral, systemic, or psychological [11]. Written informed consent was obtained from all patients, and 72 patients who provided consent were initially included in the study. After excluding individuals with instrumental measurement errors, 68 patients were ultimately included in the study. The study protocol received approval from the Ethics Committee on Human Research of Tokyo Medical and Dental University (approval number: D2019-094), and the research was conducted in compliance with the principles of the Declaration of Helsinki and the guidelines outlined in the STROBE statement. All subjects arrived at the hospital in morning baseline conditions, and halitosis measurements were conducted. To ensure consistency, participants were instructed to refrain from consuming strongly scented foods the day before measurement, using perfumed cosmetics on the day of measurement, and engaging in activities such as brushing, rinsing their mouths, eating, drinking, and smoking on the day of the examination [12, 13]. Halitosis patients underwent GC-MS measurements first, followed immediately by Kunkun dental measurements.

2.2. GC-MS

Halitosis measurements were conducted by quantifying the concentration (ng/10 ml) of VSCs, namely H2S, CH3SH, and (CH3)2S, using GC-MS for halitosis assessment (GC-8A, Shimadzu, Kyoto, Japan) [14–16]. Prior to measurements, patients were instructed to maintain mouth closure for 3 min and breathe quietly through their noses. A disposable tube was then inserted directly into the patient's mouth, and 20 ml oral air was aspirated using a syringe connected to the automatic inlet of the GC-MS. VSC concentration was determined from the peak area of each gas, with substances measured including H2S, CH3SH, and (CH3)2S. The units were assessed in ng/10 ml and subsequently converted into ppb under 20 °C temperature conditions. The olfactory thresholds proposed by Tonzetich [17] and Komori [18], which were H2S 1.5 ng/10 ml, CH3SH 0.5 ng/10 ml, and (CH3)2S 0.4 mg/10 ml, were used as each criteria for classifying normal and bad odors. The threshold values converted into ppb are 106 ppb for H2S, 25 ppb for CH3SH, and 8 ppb for (CH3)2S.

2.3. Kunkun dental (figure 2)

Immediately following the GC-MS measurements, Kunkun dental assessments were conducted. Kunkun dental is equipped with four semiconductor gas sensors primarily composed of metal oxides, along with a casing, battery, and other control devices. Semiconductor gas sensors are widely used for gas detection and are capable of detecting gases with concentrations above a few ppb [19, 20]. Due to their high sensitivity to volatile organic compounds, these sensors are well-suited for detecting low-concentration gases, such as VSCs found in the oral cavity [21]. Kunkun dental incorporates an AI-based prediction model for VSC concentration. This prediction model is constructed using training data from over 200 patients, combining multiple regression analysis with rule-based AI.

Figure 2. Kunkun dental equipment and accessories.

Download figure:

Standard image High-resolution image

Participants were instructed to wait for 30 s with their mouths closed to stabilize their oral conditions. Subsequently, 20 ml of exhaled air was aspirated using an all-plastic syringe (Henke-Sass, Wolf, Tuttlingen, Germany), and the entire volume was injected into Kunkun dental. The waveform data obtained were then input into a machine learning model to derive the concentration of each gas. A mechanism was incorporated to identify instrumental measurement errors caused by deviations from the training dataset due to injection errors or abnormal sensor conditions during measurement, and to account for odors other than VSC components (such as alcohol, tobacco, and food). After excluding four individuals with instrumental measurement errors, the data from 68 participants were considered valid. All analyzer concentrations were calculated in the order of ppb.

2.4. Analysis

To show the descriptive distributions, the VSC concentrations obtained through GC-MS and Kunkun dental were stratified by gender (male and female) and age (39 years and younger, 40–59 years, and 60 years and older) Pearson's product-moment correlation coefficients were also calculated to assess the accuracy of Kunkun dental measurements compared to GC-MS. Subsequently, the measurements were categorized into malodor and non-malodor groups using the threshold mentioned in the section 2.2 and the sensitivity and specificity were determined for Kunkun dental by considering GC-MS as the reference standard. Finally, the results of Kunkun dental were sorted into tertiles: low-, medium-, and high-concentration groups. The mean GC-MS value of VSC for each group was calculated, and analysis of variance was performed. Each group was defined as follows: the low-concentration group comprised 33.82% of the lower 23 data points, the medium-concentration group included 32.35% of the middle 22 data points, and the high-concentration group consisted of 33.82% of the upper 23 data points. Statistical analyses were conducted using R 4.3.1 (Ihaka and Gentleman, 1996) with a significance level set at p < 0.05.

3.1. Measurement results of each halitosis measuring device (table 1)

The study comprised 68 subjects, including 28 males and 40 females, with a mean age of 45.35 ± 14.64 years (ranging from 22 to 77 years). The results from GC-MS indicated a mean value of 303.22 ± 296.79 ppb for H2S (with a maximum value of 1577 ppb and minimum value of 0 ppb), 69.75 ± 91.84 ppb for CH3SH (with a maximum value of 511 ppb and minimum value of 0 ppb), and 14.65 ± 21.40 ppb for (CH3)2S (with a maximum value of 72 ppb and minimum value of 0 ppb). Of the subjects, 49 (72%) had H2S levels above the cognitive threshold, 46 (68%) had CH3SH levels above, and 24 (35%) had (CH3)2S levels above. Kunkun dental measurements yielded mean values of 341.77 ± 401.02 ppb for H2S (with a maximum value of 2739 ppb and minimum value of 0 ppb), 54.25 ± 95.38 ppb for CH3SH (with a maximum value of 611 ppb and minimum value of 0 ppb), and 14.83 ± 10.37 ppb for (CH3)2S (with a maximum value of 61 ppb and minimum value of 0 ppb).

Table 1. Measurement results of each halitosis measuring device.

   GC-MSKunkun dental   H2SCH3SH(CH3)2SH2SCH3SH(CH3)2S  nMeanSDMeanSDMeanSDMeanSDMeanSDMeanSDTotal68303.22296.7969.7591.8414.6521.4341.77401.0254.2595.3814.8310.37GenderMale28332.94345.5976.94107.8116.8923.71396.4530.567.05128.4915.8512.81Female40282.41197.8664.7173.7313.0819.37303.53265.5945.2959.9414.128.1Age≦3917166.6256.928.8548.012.638.28225.26196.1129.9842.1310.98.1640–5938349.52220.3272.5761.4418.9420.03286.04242.8539.8849.7915.558.16≦6013473.63440.94158.3164.6428.9231.18810.85733.19161.37188.6321.8616.033.2. Correlation coefficient of each halitosis measuring instrument (figure 3)

Pearson's product-moment correlation coefficients for GC-MS and Kunkun dental confirmed significant correlations: H2S had a correlation coefficient of 0.719 (p < 0.001), CH3SH had a correlation coefficient of 0.821 (p < 0.001), and (CH3)2S had a correlation coefficient of 0.637 (p < 0.001). These findings confirm a significant correlation for all three gases.

Figure 3. Distribution of GC-MS and Kunkun dental measurements.

Download figure:

Standard image High-resolution image 3.3. Kunkun dental GC-MS VSC concentrations by tertiles (analysis of variance)

The average GC-MS results corresponding to each group when Kunkun dental measurements were stratified into tertiles revealed significantly elevated VSC concentrations in the group with higher Kunkun dental readings (figure 4). For H2S, significant differences were observed between the low-concentration group and the medium-concentration group, as well as between the low-concentration group and the high-concentration group. For CH3SH, a significant difference was found between the high-concentration group and the low-concentration group. For (CH3)2S, significant differences were observed between the low-concentration group and the medium-concentration group, as well as between the low-concentration group and the high-concentration group (table 2).

Figure 4. GC-MS concentration averages for the three groups.

Download figure:

Standard image High-resolution image

Table 2. Results of analysis of variance among the three groups of Kunkun dental score.

 H2S (GC-MS)CH3SH (GC-MS)(CH3)2S (GC-MS)The three groups of Kunkun dental scoreMeanSDp-valueMeanSDp-valueMeanSDp-valueLow concentration group148.55154.17 *33.3237.6 *2.428.01 *Medium concentration group261.24187.82 39.6541.71 11.2516.47 High concentration group498.03373.3 134.97124.15 30.1225.11 

*p < 0.05

3.4. Sensitivity/specificity of each halitosis measuring Kunkun dental

The sensitivity and specificity of H2S in Kunkun dental were 0.86 and 0.53, respectively. For CH3SH, the sensitivity and specificity were 0.59 and 0.73, respectively. Additionally, for (CH3)2S, the sensitivity and specificity were 0.71 and 0.77, respectively (table 3).

Table 3. Sensitivities and specificities in Kunkun dental.

 H2SCH3SH(CH3)2SSensitivity0.860.590.71Specificity0.530.730.77

In this study, we verified the agreement between VSC concentrations measured by Kunkun dental and the precise VSC concentrations measured using GC-MS. The results revealed a strong correlation ranging from 0.637 to 0.821 for all three VSCs.

The sensitivity and specificity of the three VSCs were validated, demonstrating a range of 0.59 to 0.86 for sensitivity and 0.53–0.77 for specificity, which values are comparable to similar studies [22].

One limitation of this study was that some sensitivity and specificity values were below 0.6. This may be attributed to the machine learning model being designed to maximize correlation coefficients with GC-MS without considering cognitive threshold information. Future improvements in sensitivity and specificity are expected by developing a machine learning model that incorporates cognitive threshold information. However, according to other studies, commonly used VSC monitors for instrumental measurements exhibit sensitivity ranging from 0.79 to 0.88 and specificity from 0.61 to 0.73 when compared to GC-MS measurements, showing no significant differences with Kunkun dental [22]. Notably, the VSC monitors used for comparison detect only the total VSC values. In contrast, Kunkun dental enables component-by-component evaluation, allowing for a more detailed analysis of halitosis.

Another analysis compared the readings of each group when the Kunkun dental data were divided into three groups. The results indicated that for all components, higher concentrations of Kunkun dental extracts corresponded to higher GC-MS concentrations. However, there were no significant differences in the low- and medium-concentration groups of CH3SH. This discrepancy may be attributed to the disparity in concentration ranges between the two groups: while the maximum value for H2S in GC-MS measurements ranged from 0 to 1577 ppb, CH3SH exhibited a smaller interval between each group, with a maximum value of 511 ppb and a minimum value of 0 ppb, rendering it more susceptible to errors. Therefore, increasing the number of samples is necessary in future studies to minimize error influences and ensure accurate analysis.

Despite the validation of Kunkun dental samples against GC-MS in this study, several methodological issues persist. All subjects in this study were assessed under waking conditions, potentially resulting in elevated VSC concentrations due to halitosis. Future validations should encompass a broader range of subjects to enhance the applicability of the findings. Additionally, this study included only new patients, and the changes in VSC concentrations before and after halitosis treatment were not investigated. As halitosis can fluctuate throughout the day, it may be imperative to verify the measurement accuracy in the same patients under various conditions in future research. Four subjects were excluded from the data owing to errors; however, as all subjects were assessed under waking conditions, it is improbable that odors other than VSCs caused the errors. Operational errors or abnormal sensor conditions were the more likely contributors. Nonetheless, identifying the precise cause from the sensor information remains challenging, necessitating further analysis in future studies. Furthermore, in comparison with other studies, the patients in this research exhibited generally lower VSC concentrations [22]. This discrepancy may be attributed to patient selection biases or difference of measurement conditions. Further validation with a broader range of subjects is necessary to reassess the population distribution. Additionally, improvements such as retraining the AI model using other data could enhance its accuracy and reliability.

The results of this study demonstrate that the newly developed Kunkun dental system is capable of measuring VSC concentrations and halitosis with high accuracy. This finding suggests that Kunkun dental effectively addresses accuracy issues in measuring VSC concentrations, in addition to concerns related to miniaturization, gas type differentiation, and reactions to components other than VSCs. Utilizing a semiconductor gas sensor and a simple structure that eliminates the need for gas separation devices, Kunkun dental achieves compactness while ensuring rapidity, overcomes the drawbacks associated with existing halitosis measurement devices, such as weight and measurement duration, highlighting its efficacy in halitosis testing.

The AI-based judgment system introduced in this study draws upon the technology and knowledge of 'Kunkun Body', initially developed as a body odor analyzer. By acquiring new training data and system redesign, it was tailored to specialize in VSCs associated with halitosis [23, 24]. This technology, capable of detecting three types of VSC concentrations in breath or multiple components in complex gas mixtures, holds great promise in identifying odor components derived from other diseases in breath and biogas. Consequently, it may find applications across various domains in the future.

This study validated the Kunkun dental system's accuracy in measuring VSCs, showing strong correlations (0.637 to 0.821) with GC-MS. Sensitivity and specificity ranged from 0.59 to 0.86 and 0.53–0.77, respectively, indicating the potential for reliable halitosis screening. However, the study identified needs for further refinement in the machine learning model and broader validation to enhance its utility in detecting disease-related odor components.

No new data were created or analyzed in this study.