Cervical cancer (CC) is the fourth most common cancer among women worldwide and is a global public health issue closely related to women’s health ^{[1, 2]}. According to a World Health Organization (WHO) survey in 2022 on cervical cancer, there were about 660,000 new cases and about 350,000 deaths ^[3]. The incidence of cervical cancer is increasing among young individuals ^{[4, 5]}, placing a considerable burden on women’s health and socioeconomic development ^[6]. The progression from HPV infection to the development of cervical cancer is a prolonged and potentially reversible pathological process ^[7]. Therefore, early screening for cervical cancer and timely intervention are crucial for reducing both its incidence and mortality ^[8].

Cervical intraepithelial neoplasia (CIN) is a precancerous condition that occurs before the development of invasive cervical cancer, representing a continuous process in cervical ^[9]. CIN is classified into three grades: CIN 1, CIN 2, and CIN 3. While the majority of CIN 1 lesions may regress spontaneously, a subset of CIN 2 and CIN 3 lesions carry a risk of progression to invasive cancer ^[10].

In addition, histopathological biopsy is considered the gold standard for the clinical diagnosis of cervical cancer due to its high accuracy ^[11]. However, it involves the collection of cervical tissue samples, which carries a risk of secondary injury from infection, is relatively costly, and requires a high level of diagnostic expertise ^[12]. Therefore, it is not suitable for large-scale population screening ^[13]. Prior to histopathological biopsy, the use of auxiliary diagnostic tools to predict the risk level of cervical cancer is of critical importance. This approach facilitates the early identification and treatment of high-risk individuals, thereby reducing unnecessary time and financial burdens while improving the accuracy and cost-effectiveness of screening.

In recent years, AI-assisted diagnostic and therapeutic technologies have emerged as a prominent area of research. Extensive studies have applied statistical analysis and artificial intelligence to investigate risk factors and predict the occurrence of various diseases. Leveraging these technologies, researchers have conducted in-depth analyses of disease risk, developed predictive models, explored novel influencing factors, and continuously optimized tools for risk assessment and clinical decision support. Several studies on CC-related ML models based on public datasets have emerged. Yuan Li et al. ^[14] conducted a large-scale prospective cohort study to develop and validate a prognostic risk prediction model for overall survival in patients with cervical cancer. By integrating clinical characteristics, treatment information, and inflammatory biomarkers, their nomogram demonstrated good discrimination and calibration, with C-index values of 0.769 and 0.779 in the training and validation cohorts, respectively, and showed superior prognostic performance and clinical net benefit compared with the FIGO staging system. Lu Zhen et al. ^[15] applied unsupervised and supervised machine learning approaches to identify data-driven clinical subgroups for cervical cancer prevention using large-scale population-based electronic health records. They identified five distinct phenotypic subgroups with significantly different risks of CIN2+, CIN3+, and cervical cancer, and demonstrated that the proposed subgroup-based prevention strategy achieved robust external and diagnostic validation, enabling risk-stratified screening and targeted prevention beyond conventional guideline-based approaches. Mavra Mehmood et al. ^[16] proposed a method called “CervDetect”, which is based on four target parameters (biopsy, cytology, Schiller, and Hinselmann) and 32 risk factors collected from the UCI cervical cancer dataset. The method uses the Random Forest algorithm to select important feature variables, followed by a shallow neural network to predict cervical cancer. Ruan Yuxin et al. ^[17] developed a machine learning–based nomogram using U.S. NHIS data to predict cervical cancer risk, achieving good discrimination and calibration (AUC = 0.82) and demonstrating clinical utility for risk assessment and prevention. Karimi S et al. ^[18] applied the SMOTE method to oversample the imbalanced UCI cervical cancer dataset and used data mining techniques, including Boosted Decision Tree, Decision Forest, and Decision Jungle, to predict cervical cancer. Among various screening methods, the Boosted Decision Tree showed the best performance under the Hinselmann screening method, achieving an AUC of 0.978, which was significantly superior to the other classifiers.

Therefore, we conducted a retrospective cohort study based on real-world clinical data from hospitals and developed a predictive model for the early diagnosis of cervical cancer using routine physical examination indicators. This study aims to identify key risk factors associated with the development of CIN and provide a novel approach for early screening in high-risk populations. The findings not only offer theoretical support for effective cervical cancer prevention but also hold significant clinical application value. Additionally, by integrating clinical data from both affected and healthy individuals, we applied oversampling and undersampling techniques to balance the dataset and employed Tabnet to construct a predictive model. This model serves as a valuable reference for the prevention and management of cervical cancer in high-risk populations.

Methods

2.1 Study subjects

The study included women who underwent cervical biopsy at a hospital in Jiangsu Province from 2018 to 2022. All data were anonymized, ensuring that no patient-identifiable information was used. The dataset primarily comprised age, TCT, HPV, multiple infection, FRD, cotton-tipped swab, and cervical pathological tissue biopsy results. Of these variables, the cervical biopsy results served as the outcome measure. Ultimately, 570 participants were included from an initial cohort of 597, based on the following inclusion criteria: (1) women aged between 18 and 100 years; (2) completion of cervical pathological tissue biopsy with reliable results.

2.2 Data preprocessing

The original dataset may contain missing values, outliers, or imbalanced sampling, which necessitates data preprocessing to ensure high-quality and reliable input for subsequent analyses. The analysis process was shown in Figure 1. First, samples with missing outcome variables were excluded. Next, samples with more than 30% missing data were removed. After data preprocessing, categorical features were converted into numerical values using label encoding to enable processing and interpretation by machine learning models. Label encoding creates a mapping that assigns a unique integer to each distinct categorical label. To enable machine learning models to process and interpret categorical features, non-numeric categorical labels were converted into numerical values using label encoding. TCT results were interpreted following the TBS cervical cytology classification ^[19]. Results were categorized as negative for intraepithelial lesions or malignancy (NILM), abnormal results include atypical cytology of undetermined significance (ASC-US), low-grade squamous intraepithelial lesion (LSIL), atypical squamous cells-high grade (ASC-H), high-grade squamous intraepithelial lesion (HSIL) ^[20]. HPV genotypes were categorized according to their oncogenic potential for cervical cancer ^[21], including negative, high-risk HPV (HR-HPV) types (HPV 16, 18, 31, 33, 35, 39, 45, 51, 52, 58, 59, 66, and 68), and low-risk HPV (LR-HPV) types (any positive genotypes not classified as HR-HPV) ^[22]. Multiple infection results were categorized as negative, single infection (infection with a single HPV genotype), and multiple infection (infection with two or more HPV genotypes) ^[23]. FRD was categorized into two groups: no lesions (the swabs were brown, green or colourless), with intraepithelial neoplasia (the swabs were dark green, black or blue) ^[24]. The results of cotton-tipped swab test were categorized as negative, suspicious and positive. Based on cervical pathological tissue biopsy results, CIN 2 and CIN 3, as well as cervical cancer (CC), were collectively classified as CIN 2 or higher (CIN 2+), whereas CIN 1 was categorized as CIN 2 or lower (CIN 2−) ^[25]. Finally, missing values in the dataset were imputed using the Multiple Imputation by Chained Equations (MICE) method. This approach constructs multiple regression models to predict and replace missing values based on information from other variables, thereby maximizing data completeness and preserving statistical properties, while avoiding information loss or bias that could result from simply removing incomplete cases.

Figure 1. Analysis process of this study. PPV, positive predictive value; NPV, negative predictive value; AUC, area under the receiver operating characteristic curve.

The dataset was split into a training set and a test set in an 8:2 ratio. All subsequent sampling algorithms for handling the imbalanced dataset were applied only to the training set.

2.3 Handling Imbalanced Datasets

Imbalanced datasets represent a common challenge in machine learning and artificial intelligence. When one class significantly outnumbers the other, models tend to be biased toward the majority class, resulting in suboptimal performance. Sampling algorithms, as data-level solutions, address this issue by employing oversampling or undersampling techniques to achieve a more balanced distribution between positive and negative samples.

Oversampling is a technique that randomly duplicates a portion of samples from the minority class and adds them to the training dataset. This increases the number of minority class samples to a level comparable to that of the majority class, thereby reducing class imbalance. Synthetic Minority Over-sampling Technique (SMOTE) is an oversampling-based method for handling imbalanced data. It addresses the class imbalance by generating synthetic samples to bridge the gap between the minority and majority classes. The specific procedure of SMOTE is as follows:

1. For any minority class sample, the Euclidean distance is used to calculate its distance to other samples in the minority class set. All distances are then sorted, and thenearest samples toare identified as its k-nearest neighbors.

2. The sampling ratiois set according to the proportion of minority to majority samples. For any minority class sample, a sampleis randomly selected from its k-nearest neighbors.

3. For each randomly selected neighbor, a new sample is constructed with the original sample according to the following formula.

 (1)

Here,  represents the Euclidean distance between each feature of the minority class instance and each feature of its neighbor.

Undersampling techniques reduce the number of majority class samples to make the class distribution in the dataset more balanced, thereby improving the model’s focus on the minority class and its classification performance. TomekLinks is a commonly used undersampling method that belongs to data cleaning techniques. Its main purpose is to improve the classification boundary by removing overlapping samples between classes. By definition, if a majority class sample and a minority class sample are each other’s nearest neighbors, they form a TomekLinks pair. Once two samples constitute a TomekLinks pair, it means that either one of the samples is noise, or both samples lie on the decision boundary. Its basic idea is: given a pair of samples , if there does not exist any sample  such that condition  or  holds, then  and  form a TomekLinks pair. Specifically, TomekLinks refer to a pair of samples that are mutual nearest neighbors but belong to different classes. Such samples are usually located near the class boundary or may be anomalies. By removing the majority class samples within these TomekLinks, class overlap and noise can be effectively reduced, thereby enhancing the model’s ability to recognize boundaries, achieving sample balance and improving classification accuracy.

2.4 Model development and evaluation

In this study, TabNet was selected as the baseline model for structured data classification tasks. TabNet is a deep neural network architecture specifically designed for tabular data ^[26]. Compared with conventional deep learning models, TabNet not only possesses strong modeling capabilities but also retains a degree of interpretability similar to decision tree models, allowing for intuitive visualization of each feature’s contribution to the model’s predictions. To further enhance TabNet’s performance on structured data classification tasks, we employed Particle Swarm Optimization (PSO) for automated hyperparameter tuning. By integrating TabNet with PSO, the model can effectively handle complex tabular data while achieving superior classification performance without compromising interpretability.

To evaluate the performance of the prediction model, we used metrics including accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, kappa, and AUC. By comparing the best models obtained from different algorithms in terms of discrimination and calibration performance, we ultimately identified the most suitable model for the target population.

2.5 Statistical analysis

In this study, continuous variables were represented using mean (standard deviation), while categorical variables were presented using counts (percentages). For each variable, t-tests or Mann-Whitney tests were used for continuous variables, and chi-square tests or Fisher’s exact tests were used for categorical variables. All statistical analyses were performed on R software (version 4.4.3). Two-tailed p-values less than 0.05 were considered statistically significant.

Results

3.1 Clinical characteristics of the participants

Among the 570 subjects included, 268 (47.02%) were classified as CIN 2+ and 302 (52.98%) as CIN 2-. Baseline characteristics were presented in Table 1. The study population was split into a training set (n = 513) and a test set (n = 57). Statistically significant differences between the two groups were found for TCT, HPV, multiple infections, FRD, and cotton-tipped swab results (p < 0.05) (Table 2).

Table 1 Baseline characteristics of the participants

Table 2 Characteristics of the training and test cohorts

TCT (ThinPrep cytological test), HPV (human papillomavirus), FRD (folate receptor-mediated tumor detection), Multiple infection (the result of determining how many HPV genotypes (one or multiple) the patient is infected with), Cotton-tipped swab (the assessment result of the cotton-tipped swab). Continuous variables are described using the median, while categorical variables were described in terms of frequency (percentage).

Univariate logistic regression analysis revealed that all variables were significant (p < 0.05). Those variables were then entered into a multivariate logistic regression model, which identified age, TCT, HPV, multiple infections, and cotton-tipped swab as independent predictors of CIN 2+ (Table 3).

Table 3 Univariate and multivariate logistic regression analysis

TCT (ThinPrep cytological test), HPV (human papillomavirus), FRD (folate receptor-mediated tumor detection), Multiple infection (the result of determining how many HPV genotypes (one or multiple) the patient is infected with), Cotton-tipped swab (the assessment result of the cotton-tipped swab).

3.2 Model development and evaluation

In this study, we compared the binary classification performance of six different data processing methods combined with the TabNet model, including: the original TabNet, SMOTE+TabNet, RandomOverSampler+TabNet, TomekLinks+TabNet, TomekLinks+SMOTE+TabNet, and SMOTE+TomekLinks+TabNet (Table 1). The results demonstrated that the SMOTE+TomekLinks+TabNet model achieved the best overall performance, with an AUC of 0.780, an accuracy of 0.737, and a well‑balanced sensitivity (0.722) and specificity (0.750). Its F1‑score was 0.722, indicating stable performance in identifying both positive and negative samples.

In contrast, the TomekLinks+SMOTE+TabNet model exhibited high sensitivity (0.870) but low specificity (0.583), suggesting a strong ability to identify positive samples while producing a considerable number of false positives. The original TabNet model achieved the highest specificity (0.783), but its sensitivity was only 0.667, implying a risk of missed diagnoses. The other methods (SMOTE+TabNet, RandomOverSampler+TabNet, TomekLinks+TabNet) performed slightly worse than SMOTE+TomekLinks+TabNet in overall metrics.

The ROC curves for both the training and testing sets (Figure 2) further validated these findings. The SMOTE+TomekLinks+TabNet model maintained a high area under the curve (AUC = 0.780) on the testing set, with the curve positioned close to the upper-left corner. This indicates that the model consistently delivered strong classification performance across varying threshold values. Overall, in handling class‑imbalanced data, the SMOTE+TomekLinks+TabNet model demonstrated a balanced trade‑off between sensitivity and specificity, representing the most stable and robust approach in this study.

Table 4 The performance of model

PPV, positive predictive value; NPV, negative predictive value.

Figure 2. ROC curve.