A Machine Learning Model Based on Routine Preoperative Laboratory Parameters for the Auxiliary Differentiation of HSIL and Cervical Cancer

1. Introduction

Cervical cancer is one of the most common malignancies affecting women worldwide [1,2]. Its development usually follows a stepwise progression from cervical intraepithelial neoplasia to invasive cancer [3,4]. HSIL is an important precancerous lesion. Timely detection and treatment may prevent progression in some patients [3,5]. However, HSIL and cervical cancer are not simply adjacent stages of the same disease process. They differ in preoperative assessment, treatment planning and follow-up. Patients with HSIL are usually treated with local excision, conization or close surveillance. Patients with cervical cancer require tumor staging, imaging assessment and individualized treatment [5,6]. More accurate preoperative differentiation is therefore important. It helps define the lesion and guides subsequent treatment.

Current screening and diagnosis of cervical lesions rely mainly on HPV testing, cytology, colposcopy and histopathology [5,6,7]. HPV testing is sensitive and helps identify high-risk individuals. Cytology and colposcopy can detect abnormal lesions and guide biopsy. Histopathology remains the gold standard for diagnosis. These methods still have practical limitations. Cytology and colposcopy can be affected by sampling quality, operator experience and subjective interpretation. Biopsy is invasive and may be affected by sampling site, lesion heterogeneity and the extent of local disease. Access to organized screening and diagnostic services also remains uneven across settings [7,8]. For some patients, a single local examination may not fully reflect disease status. Objective and readily available supplementary information may improve preoperative differentiation between HSIL and cervical cancer.

Routine laboratory parameters are obtained during routine clinical care. They are accessible, inexpensive and relatively objective. A single routine parameter is usually not specific for cervical cancer. However, progression from a precancerous lesion to invasive cancer may involve systemic changes in inflammation and coagulation [9,10]. Lipid metabolism and broader metabolic reprogramming may also change during malignant progression [11,12]. Serum inflammatory markers and plasma proteomic profiling also suggest that systemic alterations may accompany increasing severity of cervical lesions [13,14]. These findings are consistent with the concept that cancer progression involves multiple biological processes rather than one isolated pathway [15]. Such changes may not appear as marked abnormalities in one parameter. They may instead appear as small, scattered and interrelated changes across several parameters. Traditional statistical methods have limits in detecting nonlinear multivariable patterns. Machine learning can integrate multidimensional information and identify potential interactions. This provides a practical way to reuse routine laboratory data.

Machine learning has been widely used in medical diagnosis, risk stratification and prognosis prediction. In cervical lesion research, models have used routine blood indicators, colposcopic images and clinical parameters for risk prediction [16,17,18]. Other studies have used cytology images, HSIL-specific management variables and HPV genotype information to improve lesion classification or risk stratification [19,20,21]. Recent work on AI-assisted cytology and systematic reviews also supports the potential of artificial intelligence in cervical lesion diagnosis [22,23]. Colposcopic lesion segmentation models provide another example of image-based auxiliary diagnosis [24]. Most studies have focused on screening, CIN grading or HSIL+ risk assessment. Few studies have directly used routine preoperative laboratory parameters to distinguish HSIL from cervical cancer. Routine laboratory parameters do not come directly from local cervical lesions. However, they can provide systemic information. If these weak signals are integrated by machine learning, they may offer an additional reference for preoperative differentiation. Therefore, this study aimed to extract combined features from routine preoperative test data. The goal was to explore a low-cost and accessible auxiliary stratification method that can fit into existing clinical workflows.

2. Materials and Methods

2.1. Study Population

We retrospectively reviewed age and preoperative laboratory data from patients at the Affiliated Cancer Hospital of Xinjiang Medical University. Eligible patients were pathologically diagnosed with HSIL or cervical cancer between April 2024 and June 2025. A total of 370 patients with HSIL and 332 patients with cervical cancer were included. The inclusion criteria were as follows: ref. [1] pathologically confirmed HSIL or cervical cancer; ref. [2] complete clinical information and peripheral blood parameters; and [3] no antitumor therapy before enrollment. The exclusion criteria were as follows: ref. [1] concurrent malignant tumors; and [2] infection, hematological disease, autoimmune disease, or severe hepatic or renal insufficiency. The study was approved by the hospital Ethics Committee (K-2025095).

2.2. Statistical Analysis

Descriptive statistics and normality testing were performed using SPSS 23.0. Model development and evaluation were performed in R software (version 4.4.2). Normality was assessed using the Shapiro-Wilk test. Normally distributed variables are presented as mean ± standard deviation and were compared using the independent-samples t-test. Homogeneity of variance was assessed using Levene's test. Welch's t-test was used when this assumption was violated. Non-normally distributed variables are presented as median and interquartile range and were compared using the Mann-Whitney U test. A two-sided P value <0.05 was considered statistically significant.

2.3. Model Construction

Stratified random sampling was performed using the rsample package in R. This approach was used to maintain a balanced distribution of the outcome between the training and test sets. The 702 patients were divided into a training set (n = 491) and a test set (n = 211) at a 7:3 ratio. The training set was used for model construction, hyperparameter tuning and internal validation. The test set was used to assess model performance. A stepwise feature-selection strategy was used to reduce model complexity. LASSO regression was first applied to reduce variable dimensionality. The optimal regularization parameter lambda was selected by 10-fold cross-validation using minimum binomial deviance [25]. Pearson correlation analysis was then performed. Variables with stronger outcome relevance and lower intervariable redundancy were retained. The selected features were used to build five models. Random forest was included as an ensemble tree-based method [26]. XGBoost and LightGBM were included as gradient-boosting methods [27,28]. Support vector machine and K-nearest neighbors were included as comparator algorithms [29,30].

2.4. Model Evaluation

Accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score were calculated in the test set. ROC curves and AUC values were used to assess discrimination. Calibration curves and comparisons of classification metrics were also generated. SHAP analysis was used to improve model interpretability. Mean absolute SHAP value was used as the feature-importance metric.

Figure 1. Study workflow for machine learning-based differentiation between HSIL and cervical cancer. A total of 702 pathologically confirmed patients were enrolled, including 370 patients with HSIL and 332 patients with cervical cancer. Patients were divided into a training set (n = 491) and a test set (n = 211) using stratified random sampling. Feature selection was performed using LASSO regression and Pearson correlation analysis. Five machine learning models were developed and evaluated in the test set. The best-performing model was interpreted using SHAP analysis.

Figure 1. Study workflow for machine learning-based differentiation between HSIL and cervical cancer. A total of 702 pathologically confirmed patients were enrolled, including 370 patients with HSIL and 332 patients with cervical cancer. Patients were divided into a training set (n = 491) and a test set (n = 211) using stratified random sampling. Feature selection was performed using LASSO regression and Pearson correlation analysis. Five machine learning models were developed and evaluated in the test set. The best-performing model was interpreted using SHAP analysis.

3. Results

3.1. Clinical Features and Variable Selection

A total of 702 female patients were included. There were 370 patients with HSIL and 332 patients with cervical cancer. All patients had complete age and preoperative laboratory data. Forty-one variables met the predefined screening criteria and entered feature selection. In the training set, LASSO regression retained 25 candidate features with nonzero coefficients at lambda = 1.0 (lambda.1se). Pearson correlation analysis was then used to reduce redundancy. Variables were retained when |r|>0.2 with the outcome and when the pairwise correlation coefficient was |r|<0.2. Ten features were selected as model inputs: P4, TT, CA125, Lp(a), PT, AFP, HDL-C, HCO₃^⁻, GGT and age. The selected variables are summarized in Table 1, and the comparison of all 41 candidate variables is provided in Supplementary Table 1.

Figure 2. Feature selection using LASSO regression. (A) LASSO coefficient profiles of candidate variables. (B) Ten-fold cross-validation curve for binomial deviance. The vertical dashed lines indicate the selected lambda values.

Figure 2. Feature selection using LASSO regression. (A) LASSO coefficient profiles of candidate variables. (B) Ten-fold cross-validation curve for binomial deviance. The vertical dashed lines indicate the selected lambda values.

3.2. Development and Evaluation of Diagnostic Models

After grid search and 5-fold cross-validation, five machine learning models were compared. All models showed discriminatory ability for differentiating HSIL from cervical cancer. Random forest achieved the highest AUC in the test set. The test-set AUC was 0.977. The positive predictive value, negative predictive value, F1 score and accuracy were 0.988, 0.815, 0.876 and 0.886, respectively. The optimal random forest hyperparameters were mtry = 1, trees = 200 and min_n = 30. The detailed performance metrics are shown in Table 2.

Figure 3. Performance of five machine learning models for differentiating HSIL from cervical cancer. (A) ROC curves of the models. (B) AUC comparison in the test set. (C) Calibration curves. (D) Comparison of classification metrics across models. AUC, area under the receiver operating characteristic curve.

Figure 3. Performance of five machine learning models for differentiating HSIL from cervical cancer. (A) ROC curves of the models. (B) AUC comparison in the test set. (C) Calibration curves. (D) Comparison of classification metrics across models. AUC, area under the receiver operating characteristic curve.

3.3. SHAP Analysis of the Random Forest Model

SHAP analysis showed that each variable contributed differently to random forest output in the test set. In the SHAP swarm plot, variables were ranked as follows: progesterone, thrombin time, CA125, Lp(a), prothrombin time, alpha-fetoprotein, HDL-C, serum bicarbonate, GGT and age. Progesterone had the largest influence on prediction. TT, CA125, Lp(a), PT and AFP also made substantial contributions.

Figure 4. SHAP interpretation of the random forest model. (A) SHAP beeswarm plot showing the direction and magnitude of feature effects. (B) Mean absolute SHAP value ranking of the 10 input variables. SHAP, SHapley Additive exPlanations.

Figure 4. SHAP interpretation of the random forest model. (A) SHAP beeswarm plot showing the direction and magnitude of feature effects. (B) Mean absolute SHAP value ranking of the 10 input variables. SHAP, SHapley Additive exPlanations.

4. Discussion

This study included 702 patients with pathologically confirmed HSIL or cervical cancer. We developed a diagnostic model based on routine preoperative laboratory parameters. After LASSO regression and Pearson correlation analysis, 10 variables were retained. The random forest model showed good performance in the test set. Its AUC was 0.977, and the positive predictive value, negative predictive value, F1 score and accuracy were 0.988, 0.815, 0.876 and 0.886, respectively.

Among the five algorithms, random forest achieved the highest AUC. This may be related to the structure of the data. Random forest constructs multiple decision trees and repeatedly partitions samples. It can capture nonlinear relationships and interactions among variables. This may be useful for the multivariable, nonlinear and weak-signal patterns in this study. We also used SHAP to interpret the model. SHAP is based on Shapley values from cooperative game theory. It quantifies the contribution of each feature to model output and helps explain the model at both the global and individual levels.

The variables identified in this study reflect systemic differences between patients with HSIL and those with cervical cancer. Progesterone reflects endocrine status and may be influenced by ovarian function, the menstrual cycle and the hormonal environment. TT and PT reflect different stages of coagulation. Patients with cervical cancer may have greater coagulation fluctuation or inflammation-related coagulation changes than patients with HSIL [10]. CA125 and AFP are not specific markers for cervical cancer. CA125 may be associated with tumor burden and selected pathological features in cervical cancer. AFP is mainly used as an adjunctive marker for hepatocellular carcinoma and some germ cell tumors [31,32]. Lp(a) is associated with lipid transport, endothelial function and chronic inflammation. HDL-C participates in reverse cholesterol transport and has anti-inflammatory and antioxidant effects [11,12]. HCO₃⁻ reflects acid-base balance and metabolic compensation. GGT is related to hepatobiliary metabolism, glutathione metabolism and oxidative stress [33]. Age may reflect differences in persistent HPV infection risk, endocrine status and immune response. Together, these routine indicators describe the preoperative systemic status of patients with HSIL and cervical cancer across several dimensions, including endocrine function, coagulation, metabolism and oxidative stress.

The random forest model performed well in differentiating HSIL from cervical cancer. Its predictive performance was relatively high compared with previous studies. Yue et al. developed an SVM model based on routine hematological and biochemical indicators for high-grade CIN risk stratification and reported an AUC of 0.75 [16]. Yuan et al. developed a colposcopic image-based CIN classification model with an AUC of 0.93 [17]. Li et al. combined HPV status, cytology and transformation-zone type to predict HSIL+ risk in patients with LSIL and reported an AUC of 0.936 [18]. Other studies have used cytology images, HSIL-specific risk-stratification variables and HPV genotype information for cervical lesion classification and risk assessment [19,20,21]. AI-assisted cytology studies and systematic reviews also support the diagnostic potential of artificial intelligence in cervical lesions [22,23]. Recent studies have examined occult microinvasive or invasive cervical cancer among patients with HSIL [34,35]. These comparisons should be interpreted carefully because target populations, predictors and validation designs differ. This is also emphasized in current guidance for clinical prediction models using artificial intelligence [36]. Colposcopic segmentation studies provide further context for image-based auxiliary diagnosis [24]. Other localization and deep-learning studies also support model-based cervical lesion stratification [37,38]. Studies on HPV viral load also support model-based risk stratification for cervical lesions [39]. Compared with models based on colposcopic images, HPV genotyping or cytology, our model used routine preoperative laboratory parameters and age. These indicators are obtained in routine clinical practice and are relatively objective. The model should therefore be regarded as a supplement to current clinical workflows. It is not a substitute for histopathology, colposcopy, imaging assessment or clinical judgment. Its value lies in integrating subtle laboratory signals that are difficult to interpret from single parameters. It may provide an objective and quantifiable reference for auxiliary differentiation between HSIL and cervical cancer.

Several limitations should be acknowledged. First, this was a single-center retrospective study. All cases came from the same hospital. Testing platforms and clinical procedures were relatively consistent. This may reduce measurement variability, but it may also limit generalizability to other centers, regions and platforms. For machine learning models, sample diversity directly affects stability and generalization. The current results therefore require validation in larger multicenter cohorts. Future reports should also follow current guidance for clinical prediction models using artificial intelligence. Second, this study mainly included age and routine preoperative laboratory parameters. HPV genotyping, TCT results, colposcopic findings and imaging data were not integrated. The model therefore reflects the classification ability of routine laboratory combinations. It does not represent the full clinical decision-making process. Future studies should integrate HPV, cytology, colposcopy and imaging data. Such models may better reflect real-world practice and may improve preoperative discrimination.

5. Conclusions

This study developed a machine learning model based on routine preoperative laboratory parameters for auxiliary differentiation between HSIL and cervical cancer. The model does not require additional omics testing or complex imaging inputs. Instead, it extracts multimarker information from routinely available clinical data. SHAP analysis provides interpretability support. These findings suggest that routine preoperative test data may contain information useful for distinguishing precancerous lesions from invasive cervical cancer. This approach may supplement current clinical workflows and provide a low-cost auxiliary tool for preoperative risk stratification.