OBOA Journal

Introduction

Cervical intraepithelial neoplasia (CIN) represents a spectrum of premalignant changes, and stratification into CIN1, CIN2, and CIN3 informs both surveillance and treatment strategies. However, current diagnostic pathways heavily rely on colposcopy and histopathology, which are invasive and resource-intensive. In Eastern Europe, the prevalence of HPV infection is estimated at 21%, with the region showing the highest global rates of highgrade cervical dysplasia (~4.3% overall in Europe) [1-3]. Among women with normal cytology in Central and Eastern Europe, HPV prevalence remains substantial at 12.6%, with HPV16 being the most frequently detected genotype, particularly in high-grade lesions [3].

Interobserver variability remains a major challenge in the histopathological diagnosis of cervical intraepithelial neoplasia (CIN), particularly in the evaluation of lower-grade lesions. Substantial disagreement among pathologists has been welldocumented, especially in differentiating between CIN I and CIN II. In contrast, diagnostic agreement tends to be higher for invasive carcinoma and CIN III. This is quantitatively reflected in reported kappa (κ) values, which measure interrater reliability: κ = 0.496 for CIN III, compared with κ = 0.172 and κ = 0.175 for CIN II and CIN I, respectively [4-6]. A key contributor to this variability is the morphological overlap between reactive epithelial changes and low-grade neoplastic processes, particularly CIN I. Histological features can often mimic one another, and even among experienced observers, the distinction may be ambiguous. As a result, both overdiagnosis and underdiagnosis of CIN I are common, which has implications for both patient anxiety and resource utilization [4-6].

To address these challenges, several emerging technologies have shown promise. Biomarker-based profiling—particularly the use of immunocytochemical markers such as p16, Ki-67, and DNA methylation signatures—has demonstrated the ability to enhance risk stratification and improve the specificity of CIN grading [7, 8]. These tools offer the potential to reduce overtreatment by identifying lesions with low malignant potential. In parallel, advances in artificial intelligence have introduced new diagnostic approaches. Deep learning models have achieved diagnostic accuracies as high as 90.8%, offering a reproducible and scalable method to support pathologist decision-making [9].

Recent advancements in machine learning (ML) and deep learning have markedly improved the diagnostic performance for cervical intraepithelial neoplasia, surpassing traditional approaches in many contexts. Diagnostic accuracy is influenced by multiple factors, including the nature of the input data, model architecture, and the robustness of the validation technique employed.

Ensemble convolutional neural networks (CNNs) trained on digital histopathological images have achieved exceptionally high accuracy, with reported rates up to 94.6% for CIN grading, underscoring their potential as decision-support tools in pathology workflows [10, 11].

Multimodal models combining clinical variables with colposcopic imaging data have further enhanced prediction capabilities. For instance, a convolutional neural network, which integrated both visual and clinical information, achieved a diagnostic accuracy of 92.3%, highlighting the value of contextual clinical data in improving classification outcomes [12]. In the context of highgrade squamous intraepithelial lesions (HSIL), the Swin-B model reached an accuracy of 91.4%, supporting its use in triaging potentially severe lesions [13].

Beyond imaging, ML models trained on molecular data have also shown strong discriminatory power. A Random Forest algorithm applied to DNA methylation profiles for identifying CIN2+ lesions achieved an area under the ROC curve (AUC) of 0.90, outperforming traditional methods such as HPV genotyping and cytological screening [14]. Similarly, a Naive Bayes classifier using methylation data demonstrated an AUC of 0.88 and a specificity of 93.9%, reflecting its utility in high-specificity applications [14]. Moreover, a recent meta-analysis of 77 studies evaluating artificial intelligence–assisted cytology reported pooled diagnostic accuracies ranging from 90% to 94%, further confirming the consistency and reliability of AI-enhanced diagnostic systems in cervical cancer screening programs [15].

The present study aimed to evaluate the predictive performance of various supervised machine learning models in differentiating between CIN1, CIN2, and CIN3 lesions, utilizing a dataset comprising HPV genotyping, cytology findings, and p16/Ki-67 dual staining.

Materials and methods

Sampling

Study Population

This retrospective study included 98 women who underwent colposcopy-guided cervical biopsies at tertiary care centers from Romania between January 2022 and December 2024. Patients were eligible for inclusion if they were adult women (aged ≥18 years) who underwent colposcopic evaluation and had histopathological confirmation of cervical intraepithelial neoplasia (CIN1, CIN2, or CIN3). Inclusion required availability of complete clinical data, including HPV genotyping results, cytological findings, and immunocytochemical staining (CINtec® for p16/Ki-67). Patients were excluded if they had a history of cervical cancer, previous hysterectomy, immunosuppressive disorders, or incomplete clinical or histopathologic data. Additionally, pregnant women and individuals with co-existing gynecologic malignancies were excluded to reduce confounding. The final dataset comprised 98 patients who met all eligibility criteria and were included in the analysis. Patients were classified into three histologically confirmed groups: CIN1 (n = 53), CIN2 (n = 36), and CIN3 (n = 9). Demographic and clinical variables, including smoking status, sexual history, HPV infection, cytological findings, and immunocytochemical markers, were collected from their medical records.

Feature Selection and Data Pre-processing

Clinical and laboratory variables with potential relevance to cervical neoplasia were extracted for analysis. These included age, residence (urban or rural), smoking history, number of sexual partners, age at sexual debut, parity, combined oral contraceptive (COC) use, sexually transmitted disease (STD) history, prior cervical treatment, HPV genotype (16/18, other high-risk, or lowrisk), CINtec® dual-staining status (p16/Ki-67), and Pap smear results (ASCUS, LSIL, HSIL, NILM). Continuous variables such as age were normalized, while categorical variables were onehot encoded. Missing data were imputed using median values for continuous variables and the most frequent category for categorical variables.

Machine Learning Models

To evaluate the predictive capacity of clinical and laboratory features in classifying CIN grade, supervised machine learning models were developed separately for each CIN category (CIN1, CIN2, CIN3). The following classifiers were employed: Logistic

Regression, Random Forest, Support Vector Machine (SVM),

Naive Bayes, k-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), and CatBoost. Model development was performed using the scikit-learn and xgboost libraries in Python (version 3.11).

Handling Class Imbalance

Given the class imbalance—especially the low number of CIN3 cases—Synthetic Minority Oversampling Technique (SMOTE) was applied to the training data within a stratified k-fold crossvalidation framework (k = 5). The oversampling was applied only on the training set in each fold to avoid data leakage.

Hyperparameter Tuning

Hyperparameters for each model were optimized using grid search with cross-validation. For logistic regression, L2 and L1 regularization parameters (C values) were tuned. For random forest, the number of estimators and tree depth were varied. SVM models were optimized for the penalty term (C) and kernel type (linear or radial basis function). KNN models were tuned by varying the number of neighbors (k). Gradient boosting models (XGBoost and CatBoost) were optimized for learning rate, depth, number of estimators, and iterations as appropriate.

Evaluation Metrics

Model performance was evaluated on a hold-out test set, with metrics including precision, recall, F1-score, overall accuracy, and area under the receiver operating characteristic curve (ROC AUC). Metrics were calculated for each class (0 and 1), with particular focus on the performance for the positive class (presence of the respective CIN grade). Macro and weighted averages were also computed. The best-performing parameters and metrics for each classifier were recorded and are presented alongside visual representations.

Ethical Approval

This study was conducted in accordance with the principles of the Declaration of Helsinki. Institutional Review Board (IRB) approval was obtained from the Institutional Ethics Committee of Clinical Hospital of Obstetrics and Gynecology ,,Buna vestire” Galati (No. 115/05.01.2021). No financial incentives were provided for participation, and patient care was not influenced by study involvement.

Results

Patient characteristics

Ninety eight patients were included in the study and their clinical characteristics are presented in Table 1. The presence of multiple sexual partners showed a statistically significant association with higher CIN grades (p = 0.022). Only 1 patient (1.9%) in the CIN1 group reported multiple partners, whereas 6 (16.7%) in CIN2 and 2 (22.2%) in CIN3 reported the same behaviour. Early sexual debut (defined as ≤15 years) was more frequently reported in CIN2 and CIN3 groups, though the difference was not statistically significant (p = 0.165).

A significant association was observed between sexually transmitted disease history and CIN grade (p < 0.001). No patients in the CIN1 group had a history of STD, whereas 1 (2.8%) in CIN2 and 3 (33.3%) in CIN3 did.

HPV 16/18 infection was significantly more common in highergrade lesions (p < 0.001). Twelve patients (22.6%) with CIN1 had HPV 16/18, compared with 18 (50.0%) in CIN2 and 8 (88.9%) in CIN3. Low-risk HPV types were not significantly associated with CIN severity (p = 0.367), found in 10 (19.2%), 3 (8.6%), and 1 (11.1%) patients with CIN1, CIN2, and CIN3, respectively.

CINtec+ (p16/Ki-67 dual staining) positivity was strongly associated with higher CIN grade (p < 0.001). It was present in 6 (11.32%) of CIN1 cases, 6 (16.7%) of CIN2, and 5 (55.6%) of CIN3. Low-grade squamous intraepithelial lesion (LSIL) findings were more common in higher CIN grades (p = 0.034), reported in 24 (45.3%) CIN1 cases, 22 (61.1%) CIN2 cases, and 8 (88.9%) CIN3 cases. High-grade squamous intraepithelial lesion (HSIL)

results were significantly more common in CIN2 and CIN3 groups compared to CIN1 (p < 0.001). Only 2 patients (3.8%) in CIN1 had HSIL, while 16 (44.4%) in CIN2 and 5 (55.6%) in CIN3 did.

Table 1: RT-PCR kits used for the diagnosis of COVID-19included in this study.

The distribution of age across the three categories of cervical intraepithelial neoplasia is presented in Figure 1. Among patients with CIN1, the median age was 35 years (interquartile range [IQR], 30–45 years). A similar age distribution was observed in the CIN2 group, with a median age of 34 years and an IQR of 30–45 years. In contrast, patients diagnosed with CIN3 were younger, with a median age of 28 years (IQR, 24–35 years), indicating a shift toward younger age in those with higher-grade lesions. There was no statistically significant difference between groups regarding age distribution (p= 0.12).

Figure 1: The distribution of age across the three categories of cervical intraepithelial neoplasia (CIN).

The performance metrics of the evaluated algorithms for CIN1 prediction is presented in Table 2 and in Figure 2. In evaluating the performance of machine learning algorithms for the prediction of CIN1, the Random Forest model demonstrated the highest overall performance. It achieved a precision of 0.67, recall of 0.67, and F1-score of 0.67 for the positive class (CIN1), with an overall accuracy of 63% and a ROC AUC of 0.658. The optimized model used no maximum depth restriction and 100 estimators. The K-nearest neighbors (KNN) classifier also showed acceptable performance, with a precision of 0.58, recall of 0.73, and F1-score of 0.65 for CIN1, an accuracy of 56%, and a ROC AUC of 0.539.

Support Vector Machine (SVM) achieved moderate values, with a precision of 0.54, recall of 0.47, and F1-score of 0.50 for the CIN1 group, an accuracy of 48%, and a ROC AUC of 0.444. Logistic regression had a lower performance with a precision of 0.50, recall of 0.40, and F1-score of 0.44 for CIN1, resulting in an overall accuracy of 44% and the lowest ROC AUC of 0.389, despite using a regularization parameter C = 1 and L2 penalty.

Naive Bayes performed modestly with a precision of 0.50, recall of 0.67, and F1-score of 0.57 for CIN1, yielding an accuracy of 44% and ROC AUC of 0.417. While both logistic regression and Naive Bayes offered high recall or precision individually, their overall discrimination was inferior to Random Forest and KNN.

 Table 2: Performance metrics of machine learning models for CIN1 prediction.

Figure 2: Performance metrics of machine learning models for CIN1 prediction.

The performance metrics of the evaluated algorithms for CIN2 prediction is presented in Table 3 and in Figure 3. Logistic regression with L2 regularization (C = 0.01) exhibited the strongest overall performance, achieving an accuracy of 0.89, a precision of 0.83, a recall of 0.71, and an F1-score of 0.77. Its discriminative ability, as indicated by the ROC AUC, was 0.8393, suggesting a good balance between sensitivity and specificity.

CatBoost, a gradient boosting model, also showed high performance with an accuracy of 0.85, precision and recall both at 0.71, and an F1-score of 0.71. It yielded the highest ROC AUC among all models at 0.8464, indicating excellent predictive power. This model was tuned using a depth of 5 and 100 iterations.

Random Forest, configured with a maximum depth of 10 and 50 estimators, reached an accuracy of 0.78, with precision, recall, and F1-score each at 0.57. Its ROC AUC of 0.8036 indicated solid model discrimination.

The Support Vector Machine with a radial basis function kernel (C = 10) achieved an accuracy of 0.74, precision of 0.50, recall of 0.71, and an F1-score of 0.59, with a ROC AUC of 0.7893. The k-Nearest Neighbors (KNN) model (k = 5) provided similar performance, with an accuracy of 0.78, precision of 0.56, recall of 0.71, and an F1-score of 0.62. Its ROC AUC was 0.7750, indicating moderate predictive power.

XGBoost, another ensemble learning model, achieved an accuracy of 0.74, with precision of 0.50, recall of 0.57, and F1-score of 0.53. Its ROC AUC stood at 0.7964, which was comparable to other ensemble approaches. In contrast, Naive Bayes demonstrated the weakest performance, with an accuracy of 0.37, precision of 0.27, and an F1-score of 0.41, despite a relatively high recall of 0.86. Its ROC AUC was 0.6893, suggesting limited discriminatory capability.

Table 3: Performance metrics of machine learning models for CIN2 prediction.

Figure 3: Performance metrics of machine learning models for CIN2 prediction.