Cervical cancer is the second most common malignancy in women and one of the leading causes of death, especially in developing countries. Its estimated incidence is 348,809 cases, with a mortality rate of approximately 60% (207,210 deaths) [1]. Cervical cancer-related deaths are expected to reach 12 million by 2030. Approximately 18,000 new cases of cervical cancer are reported annually in Indonesia, with a first-year mortality rate of 75%. The majority of newly diagnosed patients at an advanced stage at the time of diagnosis (70% of cases) and those with terminal cancer account for this death rate [2].

Cervical cancer can be prevented and controlled by screening and human papillomavirus (HPV) vaccination. Effective screening programs, such as Papanicolaou smear, have significantly reduced cervical cancer rates in developed countries [3]. In South Korea, the incidence of cervical cancer is decreasing annually owing to screening and prophylactic vaccination [4]. Cervical cancer remains an extremely important disease in developing countries [5]. However, such programs are not widely available in developing nations. The World Health Organization (WHO) advocates for national programs that use visual inspection with acetic acid (VIA) examinations to detect and manage precancerous lesions, particularly in low-resource countries. Therefore, sexually active women with a positive VIA test result should undergo cryoablation [6]. Meta-analysis showed an VIA sensitivity of 49-98% and specificity of 75-91% [7].

Several undeveloped and developing countries have failed to meet the WHO recommendations for screening 80% of the targeted population for cervical cancer. There is an urgent need for a reliable and convenient detection method for volatile biomarkers of cervical cancer. Volatile organic compounds (VOCs) are present in biological matrices such as urine and blood under various conditions, including cancer [8]. Detection methods such as gas chromatography-mass spectrometry (GC-MS) and electronic noses have been developed. An electronic nose, which uses arrays of chemical sensors, employs artificial intelligence to differentiate and classify chemical vapor samples using multivariate statistical techniques [9].

Persistent HPV infection may cause biological changes in the cervical epithelium, leading to VOC production. In this study, we developed a novel strategy for detecting cervical cancer using VOC sensing. Subsequently, we aimed to quantitatively study volatile biomarkers in the urine samples of women with cervical cancer and healthy controls using an electronic nose.

This quantitative and experimental case-control study included 200 urine samples from HPV-negative patients. This study was conducted in Indonesia between May and July 2022. The eligibility criteria for the cervical cancer group were as follows: 1) patients with a histopathological diagnosis of cervical cancer; 2) patients who had not undergone or were not undergoing any standard treatment, such as radiation therapy, concurrent chemoradiotherapy, or surgery; 3) patients who had not been cytologically diagnosed according to lesion type (Bethesda classification: low grade squamous intraepithelial lesion or high grade squamous intraepithelial lesion); and 4) patients who signed an informed consent form.

The eligibility criteria for the healthy control group were as follows: 1) women negative for HPV infection and 2) patients who signed an informed consent form. The exclusion criteria were as follows: 1) insufficient sample collection; 2) death during follow-up; and 3) comorbidities (liver disease, diabetes mellitus, hypercholesterolemia, and other malignancies). Our study included cervical cancer patients admitted to a tertiary university hospital. Control participants were recruited through cervical cancer screening programs at primary healthcare centers and underwent a thorough clinical assessment to evaluate potential risk factors, followed by a physical examination using visual inspection with acetic acid. This was complemented by an HPV DNA test to confirm the absence of HPV infection both clinically and at the molecular level. The collected samples reflected the clinical settings and study design. As inpatients, multiple urine samples were collected from cervical cancer participants at various time points during their hospital stay. This strategy was intended to capture temporal variations in urinary VOCs, potentially associated with disease progression, treatment response, and physiological changes. Healthy control participants were outpatients who provided only a single urine sample per individual, as their VOC profiles were presumed to be more stable over time. Samples from both groups were collected using standardized protocols to ensure consistency and reliability of VOC analysis.

Informed consent was obtained from all participants and the study was approved by the Medical and Health Research Ethics Committee. In accordance with the journal’s guidelines, we will provide our data for independent analysis by a team selected by the editorial team for the purpose of additional data analysis or for the reproducibility of this study in other centers, if requested.

We obtained 95 urine samples from healthy controls and 235 urine samples from cervical cancer patients. The age of the study participants ranged from 20 years to 58 years, as shown in Table 1.

The data were projected onto the first two principal components (PC1 and PC2) to capture the differences among the groups from the PCA plot and the most salient variance within the dataset (Fig. 2). PC1 accounted for 76.83% of the observed variance, while PC2 accounted for 9.22% of the observed variance. This revealed a degree of overlap, suggesting that PCA may not optimally differentiate between these classes. However, a clear separation between the “negative” and “positive” classes was evident along linear discriminant 1 (LD1).

The performance of the XGBoost model (Python) on the training data describes key metrics such as precision, recall, and F1-score for each class, as well as the overall accuracy and weighted averages. The model achieved high precision and recall for both classes, as listed in Table 2.

The model maintained a high level of performance and achieved an overall accuracy of 0.89. The results indicated that the model is effectively generalizable, demonstrating its robustness and reliability in real-world scenarios. Although the precision and recall for the “negative” class using the test data were slightly lower than those using the training data, they remained respectable, indicating that the model can still effectively identify the majority of negative instances. The continued strong performance for the “positive” class further emphasizes the model’s ability to accurately classify positive instances.

Fig. 3A, B show the GC-MS analysis of two samples from the healthy group, which revealed that carbon monoxide and nitrogen compounds were the predominant components detected at a retention time of two seconds. Fig. 3C, D show the components identified in two samples obtained from the asymptomatic individuals. GC-MS identified carbon monoxide, nitrogen, and nickel tetracarbonyl in the first, second, and third hits, respectively.

The GC-MS analysis of two samples from cervical cancer patients revealed the presence of carbon monoxide, nitrogen, and nickel tetracarbonyl over a retention period of two seconds (Fig. 4A, B). Table 3 summarises identical and varying VOCs between groups. The GC-MS analysis of urine samples from two other patients diagnosed with cervical cancer revealed the presence of carbon monoxide, nitrogen, and trisilaxane (Fig. 4C, D). Comprehensive overview of global feature importance in the XGBoost model (Python) to validate the LDA results. The SHAP plot revealed the dominance of sensors S3, S2, and S10. Conversely, sensors S4 and S6 demonstrated relatively low SHAP values, implying a less significant role in the decision-making process of the model.

Fig. 5A, which consists of 5X, shows the force plot of the impact of individual sensors on the XGBoost model’s (Python) prediction and Fig. 5 presents a compelling visualization of the complex interplay between individual sensor responses and the XGBoost model’s (Python) prediction for both negative and positive classes. Fig. 5 shows that S1 exerted a strong negative influence, pushing the prediction to be significantly lower. Conversely, sensors S3 and S2 counteracted this effect, driving the prediction toward the positive class. Despite their opposing influences, the combined effect of all the sensors resulted in a final prediction (f(x)=-1.56), which remained within the negative range.

The bottom plot, representing a positive sample, reveals a different pattern of sensor contributions. Sensor S10 was the primary driver of positive classification and exerted a strong positive influence. S2 also contributed positively, albeit to a lesser extent. In addition, S3 exerted a slightly positive influence. This contrasting behavior highlights the complex relationships between the sensor responses and sample classification.

Additionally, Fig. 5B, using boxplot analysis, further highlights these findings by providing the distribution and variability of the sensor responses, aiding in the identification of sensors with higher discriminatory power. The boxplot distribution of the sensor responses for the positive and negative samples illustrates the maximum response increase from the baseline for each sensor (S1-S10) in the electronic nose array. This analysis displays the range and variability of the maximum response increase for each sensor. By comparing the positive and negative samples, distinct patterns emerged across the different sensors. Notably, sensors S2, S3, and S10 exhibited a clear separation between the median responses of the positive and negative samples. Conversely, sensors S4, S5, and S6 showed substantial overlap in their interquartile ranges, indicating a lower discriminatory power.

The presence of outliers, depicted as individual points beyond whiskers, highlighted the variability within each sample group. Furthermore, the varying lengths of the boxes and whiskers across different sensors emphasized the differences in the dynamic ranges and sensitivities of the sensors.