Survival Period Prediction in Cervical Cancer Patients Using the Selective Stacking Technique

Abstract

Objectives

This study aims to increase the effectiveness of cervical cancer treatment by developing a survival prediction model using an innovative ensemble machine learning approach, namely the selective stacking technique.

Methods

Patient data obtained from the Faculty of Medicine, Chiang Mai University, Thailand, were utilized to validate the real-world applicability of the proposed approach. The selective stacking model employed a two-stage machine learning framework in which outputs from base machine learning models were systematically combined through meta-level learning. Importantly, the performance of the proposed model was compared with that reported in previous studies that relied on individual machine learning algorithms as baselines. To provide deeper insight into the predictive mechanisms of the model, local interpretable model-agnostic explanations were applied to assess feature importance and identify the most influential factors contributing to model predictions.

Results

The classification model developed using the selective stacking technique demonstrated a marked improvement in prediction accuracy, achieving an accuracy of 91.41%. The regression model also showed robust performance, with a root mean square error of 18.92 and an r value of 0.669. Feature importance analysis indicated that side effect status involving surrounding organs emerged as the most influential factor in survival prediction.

Conclusions

The selective stacking model exhibited superior predictive performance compared with the base models, suggesting that this approach offers a promising strategy for cervical cancer survival prediction and may support the development of more personalized treatment planning.

I. Introduction

Cancer represents one of the most significant global health challenges, ranking second only to cardiovascular disease [1,2]. This disease arises from the abnormal growth, division, and formation of atypical cells [2]. Among the various cancers affecting women worldwide, cervical cancer is particularly dangerous [3,4]. The progression of cervical cancer is characterized by a prolonged and complex developmental trajectory. Precancerous lesions may take several years to progress into invasive cervical cancer, which substantially complicates timely detection and early-stage intervention [5,6]. As a result, clinical symptoms often appear only at advanced stages, further hindering effective diagnosis and treatment. Given these challenges, accurate prediction of patient survival becomes critically important. Integrating predictive models into treatment planning may help clinicians avoid unnecessary individualized treatments, thereby potentially improving patient outcomes and optimizing healthcare resource utilization [7].

The survival rate of cervical cancer is a complex outcome influenced by multiple key factors, including cancer stage at diagnosis, patient age, overall health status, and treatment modality. Traditionally, survival prediction has relied on statistical methods, with Cox proportional hazards models and Kaplan–Meier analysis serving as the primary analytical tools [8]. These approaches generally evaluate the time from the start of follow-up until the occurrence of a defined event, such as disease onset or patient death. However, accurate interpretation of the effects of multiple influencing factors requires substantial expertise in statistical evaluation, comparison, and analytical methodology [9,10]. Despite their longstanding utility, conventional statistical techniques face notable limitations when applied to non-linear relationships commonly encountered in complex biomedical data [11]. In recent years, machine learning approaches have increasingly been proposed as viable alternatives, demonstrating improved predictive performance relative to traditional statistical methods, particularly in multivariable settings [12]. These algorithms are designed to enhance prediction accuracy and uncover complex patterns within large and heterogeneous datasets, rendering them especially suitable for advanced medical data analysis [13,14].

Previous work by Chanudom et al. [15] demonstrated the potential of machine learning algorithms for predicting survival among cervical cancer patients using medical records from the Faculty of Medicine, Chiang Mai University. Although that study highlighted the advantages of machine learning over conventional statistical methods, it was constrained by a key methodological limitation: individual machine learning algorithms were applied independently, without exploiting their combined predictive strengths. Ensemble learning has since emerged as an advanced machine learning paradigm designed to address this limitation. By integrating multiple predictive models, ensemble approaches establish a unified analytical framework that improves overall performance through the strategic combination of individual model outputs [16–19]. Building on these methodological advances, the present study applies the selective stacking technique (SST), a sophisticated ensemble learning method. By systematically integrating the outputs of multiple classification and regression models, this study seeks to enhance predictive performance in survival period estimation and to address the limitations identified in prior work [15]. In addition, feature importance analysis was conducted to improve model interpretability and to identify factors influencing survival period prediction.

II. Methods

1. Research Framework

The proposed research introduces a novel machine learning approach for survival prediction, referred to as the SST. The overall methodology consists of three distinct phases, as illustrated in Figure 1. Phase 1 focuses on data understanding and preparation to ensure that the dataset is suitable for subsequent modeling. Phase 2 involves the modeling process itself and explains the operational mechanism of SST. Phase 3 is dedicated to feature importance estimation and model explanation using local interpretable model-agnostic explanations (LIME).

Figure 1.

Overview of the selective stacking technique for cervical cancer survival prediction.

2. Dataset

In this study, data from cervical cancer patients were obtained from the Faculty of Medicine, Chiang Mai University. The inclusion criteria required participants to be at least 18 years of age, to have been diagnosed with cervical cancer, to have undergone computed tomography and transabdominal ultrasound-based planning for brachytherapy, and to have received radical radiotherapy between 2008 and 2018. Exclusion criteria were also applied, resulting in the omission of certain individuals from the analysis. Specifically, patients who had undergone re-irradiation or radical surgery in the pelvic region were excluded from the study. The data characteristics of the 295 included patients are reported in Chanudom et al. [15] and comprise information on patient age, tumor size, pathology, tumor stage, chemotherapy method, brachytherapy method, administered radiation dose, side effect status, diagnosis start date, diagnosis end date, date of local recurrence, date of distant recurrence, date of death, and survival status. Side effect status across different anatomical regions was categorized into multiple levels, including no side effects, normal side effects, side effects requiring medication, side effects requiring treatment, side effects requiring long-term treatment or that were difficult to treat, and death resulting from side effects. Clinical outcomes derived from this dataset have been previously reported in international publications [20].

3. SST Modeling

Both classification and regression models were developed in this study. For the regression model, survival time was defined as the target variable and measured in months. For the classification model, survival time was discretized into four categories: <6 months, 6 months to 3 years, 3 years to 5 years, and >5 years. The distribution of records across these target classes was imbalanced. To mitigate this issue, the Synthetic Minority Over-sampling Technique (SMOTE), as proposed by Chawla et al. [21], was applied to balance the class distribution. However, because SMOTE does not natively support nominal attributes, a slight modification was introduced through the use of the “nominal change rate” parameter. This parameter represents the probability that a synthetic sample’s nominal attribute is assigned the value of the nearest neighbor rather than retaining the value of the original instance. In this study, the nominal change rate was set to 0.5, indicating a 50% probability that a nominal attribute value would be inherited from the original sample and a 50% probability that it would be inherited from the neighboring instance. This configuration allows SMOTE to generate synthetic samples with valid categorical combinations by copying values from real observations, thereby facilitating the creation of realistic synthetic data for nominal features.

The stacking technique is illustrated in Figure 2, where S(x_i,y_i)_pxq represents the training dataset composed of feature vector x_i and labels y_i. The variables p and q denote the number of subjects and features, respectively. The dataset S(x_i,y_i)_pxq was used to train the base-level models denoted as M_N Subsequently, the prediction probabilities were used to train the meta-learner model (M_j) or meta-level model. This meta-learning model was trained to identify the optimal combination of base-level models for generating the final prediction result ( $\widehat{y_{final}}$). The probability distribution for each base-level model was established using the following equation:

Figure 2.

Overview of the selective

$$P^M(x)=(P^M(c_1\mid x),P^M(c_2\mid x),P^M(c_3\mid x),\dots ,P^M(c_m\mid x))$$

Where c_m is the possible target, and P^M(c_m|x) is the probability that example x belongs to target c_m by model M_N. The probabilities generated for each target class by each M_N served as meta-level variables [22,23].

For the base-level models, we compared several algorithms, including decision tree (DT), random forest (RF), naive Bayes (NB), artificial neural network (ANN), support vector machine (SVM), k-nearest neighbors (KNN), adaptive boosting (AdaBoost) and gradient boosted trees (GBT).

Model performance was assessed using a 10-fold cross-validation procedure. This approach was selected to maximize utilization of the available data and to ensure a robust evaluation of model generalizability. The complete dataset of 295 records was randomly partitioned into 10 equal-sized subsets. During each iteration, the model was trained on nine subsets and evaluated on the remaining held-out subset. This procedure was repeated 10 times, with each subset serving once as the validation set. The overall model performance was then reported as the average performance across all folds.

Two SST models were constructed using the three highest-performing base-level models, with RF and ANN separately serving as the meta-level model. In addition, a bagging model based on the top-performing base model and a voting ensemble (VOTE) incorporating the same three base models as SST were developed to further compare stacking performance. For the VOTE model, classification was performed by selecting the class receiving the highest number of votes, whereas regression predictions were obtained by averaging the outputs of the base-level models. Parameter optimization was conducted for all algorithms, and their predictive performance was evaluated and compared with that of individual machine learning models employed at the base level.

4. Feature Importance

Providing a comprehensive explanation of survival prediction outcomes, as well as identifying features with the greatest influence, is of considerable interest to medical professionals [24–26]. To address this need, the LIME algorithm was employed. LIME is a model-agnostic approach that explains predictions from classification or regression models by approximating them locally with an interpretable surrogate model through systematic perturbation of input features and observation of resulting changes in predictions. The mathematical formulation of LIME is expressed as follows:

$$interpretation(x)=arg\hspace{0.17em}\underset{v\in V}{min}\hspace{0.17em}L(u,v,{\pi }_x)+\omega (v)$$

where u denotes the initial model, v represents the explainable model, π_x is the function for calculating the weight of each sample from the uniformity of the training samples v, and ω(v) represents the complexity of v and π_x calculated as follows:

$${\pi }_x(z)=\sqrt{e^{-\frac{D_{(x,z)}^2}{{\sigma }^2}}}$$

where x is the datapoint of interest, D_(x,z) is the distance between x and z, which can be calculated by using any function depending on the data type, and σ² is the kernel size, which must be configured manually.

5. Ethical Statement

This study was approved by the Research Ethics Committee, Faculty of Medicine, Chiang Mai University (Reference number: NONE-2565-09166).

III. Results

1. Classification Model Result

Table 1 presents the performance of the classification models, including the model type, target classes, and evaluation metrics. The three most accurate classification models for predicting survival period in cervical cancer patients were RF, ANN (1 hidden layer), and AdaBoost, achieving accuracy values of 87.48%, 87.46%, and 85.49%, respectively.

Table 1.

Performance of the classification models

Model	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
DT	74.95 ± 7.03
<6 months		84.71	94.74	89.44
6 months to 3 years		65.38	44.74	53.13
3 years to 5 years		68.48	82.89	75.00
>5 years		78.67	77.63	78.15
RF	87.48 ± 6.07
<6 months		95.39	100	97.64
6 months to 3 years		92.59	65.79	76.92
3 years to 5 years		84.34	92.11	88.05
>5 years		89.74	92.11	90.91
NB	58.23 ± 7.44
<6 months		100	100	100
6 months to 3 years		66.09	11.84	20.08
3 years to 5 years		48.00	47.37	47.68
>5 years		53.33	73.68	61.87
ANN (1 hidden layer)	87.46 ± 6.65
<6 months		85.90	88.16	87.02
6 months to 3 years		84.06	76.32	80.00
3 years to 5 years		89.74	92.11	90.91
>5 years		89.87	93.42	91.61
ANN (2 hidden layers)	79.88 ± 7.20
<6 months		76.92	92.11	83.83
6 months to 3 years		76.27	59.21	66.67
3 years to 5 years		80.26	80.26	80.26
>5 years		85.90	88.16	87.02
SVM	70.72 ± 9.01
<6 months		87.65	93.42	90.44
6 months to 3 years		61.70	38.16	47.16
3 years to 5 years		58.65	80.26	67.77
>5 years		75.00	71.05	72.97
KNN	72.05 ± 6.77
<6 months		80.46	92.11	85.89
6 months to 3 years		71.70	50.00	58.92
3 years to 5 years		69.41	77.63	73.29
>5 years		65.82	68.42	67.09
AdaBoost	85.49 ± 6.76
<6 months		89.87	93.42	91.61
6 months to 3 years		84.75	65.79	74.08
3 years to 5 years		81.48	86.84	84.07
>5 years		85.88	96.05	90.68
GBT	84.85 ± 5.71
<6 months		87.36	100	93.25
6 months to 3 years		79.37	65.79	71.94
3 years to 5 years		80.52	81.58	81.05
>5 years		90.91	92.11	91.51
VOTE	90.77 ± 4.94
<6 months		89.29	98.68	93.75
6 months to 3 years		95.00	75.00	83.82
3 years to 5 years		89.74	92.11	90.91
>5 years		90.24	97.37	93.67
Bagging	84.53 ± 5.63
<6 months		85.19	90.79	87.90
6 months to 3 years		87.72	65.79	75.19
3 years to 5 years		83.75	88.16	85.90
>5 years		82.56	93.42	87.65
SST with RF	87.81 ± 6.12
<6 months		90.38	96.05	93.13
6 months to 3 years		91.38	69.74	79.11
3 years to 5 years		82.72	88.16	85.35
>5 years		88.10	97.37	92.50
SST with ANN	91.41 ± 5.03
<6 months		90.48	100	95.00
6 months to 3 years		95.08	76.32	84.67
3 years to 5 years		89.74	92.11	90.91
>5 years		91.36	97.37	94.27

Values are presented as mean ± standard deviation.

DT: decision tree, RF: random forest, NB: naive Bayes, ANN: artificial neural network, SVM: support vector machine, KNN: k-nearest neighbors, LR: linear regression, GBT: gradient boosted trees, VOTE: voting ensemble, SST: selective stacking technique.

The three top-performing classification models (RF, ANN with 1 hidden layer, and AdaBoost) were subsequently selected as base-level models. The SST with ANN as the meta-level classification model achieved an accuracy of 91.41%. In addition, SST with ANN demonstrated relatively stable performance, with a standard deviation of 5.03%.

In terms of overall accuracy, SST with ANN, SST with RF, and VOTE outperformed all single-model classifiers, including DT, NB, ANN, SVM, and KNN. When class-specific performance was considered, these ensemble models consistently exceeded the performance of single models across nearly all F1-score categories. Among the ensemble approaches evaluated, SST with ANN delivered the strongest overall performance when compared with the voting ensemble, AdaBoost, and GBT.

It is worth noting that SST with ANN only marginally outperformed the VOTE model in terms of overall accuracy. However, SST with ANN consistently demonstrated higher F1-scores and recall values across most target classes, with particularly strong performance in identifying patients with survival times of less than 6 months. This indicates that SST with ANN is more effective at distinguishing high-risk patient groups, a capability that is especially important for clinical decision-making.

2. Regression Model Result

The performance of the regression models is presented in Table 2. Based on root mean squared error (RMSE) and r values, the three best-performing regression models were ANN with 2 hidden layers, ANN with 1 hidden layer, and linear regression (LR). These models achieved RMSE values of 20.87, 21.04, and 21.40, respectively, with corresponding r values of 0.608, 0.606, and 0.595. The SST with RF regression model achieved the lowest RMSE and the highest r value, at 18.92 and 0.669, respectively, indicating superior performance compared with the individual base models.

Table 2.

Performance of the regression models

Model	MAE	RMSE	Correlation coefficient (r)
DT	18.970 ± 7.381	25.22 ± 11.67	0.352 ± 0.387
RF	15.841 ± 5.511	21.43 ± 8.80	0.591 ± 0.268
ANN (1 hidden layer)	14.996 ± 5.360	21.04 ± 7.79	0.606 ± 0.257
ANN (2 hidden layers)	15.053 ± 5.104	20.87 ± 7.91	0.608 ± 0.250
SVM	17.911 ± 5.376	26.96 ± 10.14	0.000 ± 0.000
KNN	18.882 ± 5.467	25.20 ± 9.18	0.312 ± 0.195
LR	15.064 ± 5.112	21.40 ± 7.72	0.595 ± 0.248
GBT	17.270 ± 5.400	23.26 ± 8.17	0.445 ± 0.181
VOTE	17.335 ± 6.538	21.99 ± 8.83	0.633 ± 0.267
Bagging	16.913 ± 4.607	21.95 ± 7.67	0.606 ± 0.248
SST with RF	14.031 ± 4.809	18.92 ± 7.50	0.669 ± 0.278
SST with ANN	17.735 ± 4.809	25.11 ± 6.93	0.572 ± 0.274

Values are presented as mean ± standard deviation.

DT: decision tree, RF: random forest, NB: naive Bayes, ANN: artificial neural network, SVM: support vector machine, KNN: k-nearest neighbors, LR: linear regression, GBT: gradient boosted trees, VOTE: voting ensemble, SST: selective stacking technique, MAE: mean absolute error, RMSE: root mean squared error.

The RMSE values per patient for the three top-performing base models and the SST with RF model are illustrated in Figure 3. It can be observed that as patient survival time increased, prediction accuracy decreased. This finding suggests that longer survival periods are associated with greater difficulty in accurately predicting survival time. In addition, survival times of less than 4 months also reduce prediction accuracy.

Figure 3.

Comparison of root mean square error values by patient survival time for base models and the the selective stacking technique (SST): (A) ANN (1 hidden layers), (B) ANN (2 hidden layers), (C) LR, (D) SST. ANN: artificial neural network, LR: linear regression.

Although the regression model constructed using SST with RF outperformed the other models overall, analysis of RMSE values by patient survival time, as shown in Figure 3D, indicates that the model remains unsuitable for predicting survival time without clearly defined boundaries. This limitation arose because survival times greater than 5 years continued to reduce predictive accuracy. Similarly, survival times of less than 4 months still posed challenges for accurate prediction, although the magnitude of error in this range was reduced. Nevertheless, the SST model demonstrates a lower overall error than the other regression models.

Figure 4 presents a comparison between actual and predicted survival times for each data point, highlighting that all models experience difficulty in predicting outcomes for patients with longer survival periods. As shown in Figure 4A, 4B, and 4C, corresponding to the three top-performing base models, the gap between actual and predicted values remained relatively small until survival times exceeded 5 years (> 60 months). A similar pattern is observed for the SST model. However, the SST model, shown in Figure 4D, reduced the discrepancy between actual and predicted survival times more effectively, suggesting that SST provides a more accurate survival prediction model for cervical cancer patients and improves overall predictive performance. The comparison between actual and predicted survival times yielded results consistent with the RMSE analysis of patient survival time described earlier.

Figure 4.

Comparison of actual and predicted survival time for base models and the selective stacking technique (SST): (A) ANN (1 hidden layers), (B) ANN (2 hidden layers), (C) LR, (D) SST. ANN: artificial neural network, LR: linear regression.

3. Feature Importance

The feature importance results obtained using LIME at both the global and local weight levels are shown in Figures 5 and 6. The side effect status around the vaginal (Vg) area exhibited the strongest influence on target class prediction, followed by side effect status in the subcutaneous and genitourinary regions.

Figure 5.

Global level feature importance, with features on the Y-axis and importance scores on the X-axis. Vg: vaginal, Subs: subcutaneous, GU: genitourinary, GI: gastrointestinal, Date(LC-FU): the time interval between local control and follow-up, Cx: cervix, Date(DF-FU): the time interval between distant failure and follow-up, FIGO: International Federation of Gynecology and Obstetrics, R: rectum, B: bladder, TECH: brachytherapy, CHEM: chemotherapy, PAT: pathology.

Figure 6.

Representative set of local-level feature importance results, with features on the Y-axis, and importance scores of each target class on the X-axis: (A) target class of <6 months, (B) target class of 6 months to 3 years, (C) target class of 3 years to 5 years, and (D) target class of >5 years. Vg: vaginal, Subs: subcutaneous, GU: genitourinary, GI: gastrointestinal, Date(LC-FU): the time interval between local control and follow-up, Cx: cervix, Date(DF-FU): the time interval between distant failure and follow-up, FIGO: International Federation of Gynecology and Obstetrics, R: rectum, B: bladder, TECH: brachytherapy, CHEM: chemotherapy, PAT: pathology.

As shown in Figure 6, patient age at the start date of diagnosis had a substantial impact on survival period prediction, with lower age generally corresponding to longer survival. This finding is consistent with tumor stage, as reflected by the FIGO (International Federation of Gynecology and Obstetrics) staging system (2018), where these features are commonly associated with reduced survival time from a clinical perspective. Radiation dose administered to the bladder region exhibited a similar effect, with higher doses corresponding to shorter survival periods. This relationship may reflect the association between radiation dose and tumor size. In contrast, higher values of Vg were associated with longer survival times in cervical cancer patients. Such cases may reflect long-term treatment effects that result in side effects across multiple organs. It should be noted, however, that the influence of predictive factors may vary among individual patients, even when factor values are identical.

Side effect status and radiation dose information for specific organs are often excluded from datasets used in traditional statistical survival analyses and are typically reserved for secondary analyses. In this study, however, these variables were intentionally included, as machine learning approaches benefit from incorporating all variables with potential influence on model output. Once the model is adequately trained, features with minimal predictive contribution are implicitly down-weighted or excluded. Moreover, inclusion of these variables provides insight into patient treatment response, which contributes meaningfully to prognosis estimation.

IV. Discussion

For the classification model, Table 1 shows that the recall values of the top three models were lowest for the “6 months to 3 years” target class compared with the other target classes. This finding suggests that these models have a relatively limited ability to correctly identify and retrieve cases belonging to this specific category. In contrast, the models demonstrated high recall values for the “<6 months” and “>5 years” target classes.

Similarly, the precision values indicate that all three models exhibited lower false-positive rates when predicting longer survival periods, particularly for the “>5 years” target class. In addition, all three models demonstrated comparable and consistently high precision values across the remaining target classes. Furthermore, the F1-score, which integrates both recall and precision, further confirmed the strong overall performance of these models.

When the SST technique was applied to the same dataset, the model could accurately identify target classes based on F1-score performance. Moreover, an examination of precision and recall demonstrated that the SST model could predict target classes without relying on strict survival period boundaries, unlike the individual top three models. The recall values showed that the SST-based model had improved retrieval capability for the “6 months to 3 years” and “3 years to 5 years” target classes. For the remaining target classes, the recall values were also higher than those observed for the top three models, indicating an overall improvement in recall performance. An analysis of precision further showed that SST-based models maintained consistently high precision across all target classes. Similar to the top three models with the highest accuracy, SST preserved strong predictive capability and accurately differentiated between classes across all survival categories.

When comparing SST performance with previous research by Chanudom et al. [15], which reported average accuracy, precision, and recall across all classes, several notable differences are observed. In Chanudom et al. [15], the GBT model achieved average accuracy, precision, and recall values of 86.52%, 86.65%, and 88.00%, respectively. In contrast, in the present study, SST with ANN achieved average values of 91.41%, 91.67%, and 91.21%, respectively, as calculated from Table 1. In terms of mean accuracy, precision, and recall, SST therefore outperforms GBT, making SST the more favorable model. With respect to regression models, the best-performing model reported by Chanudom et al. [15] was RF, which achieved an RMSE of 22.33. By comparison, the SST regression model in this study achieved an RMSE of 18.92, indicating superior predictive performance.

By identifying feature importance influencing the classification model for survival period prediction in cervical cancer patients using the LIME technique, this study assessed the extent to which each feature affects individual patients and the overall dataset through local and global weights, respectively. This analysis helps clarify which factors influence model predictions and contribute to longer survival periods.

A key limitation of this study is the reliance on data obtained from a single medical center. Comprehensive survival data are inherently difficult to collect, as survival analysis requires long-term, systematic patient follow-up, and such data become available only after many years of observation. Although use of a single data source ensures consistency in clinical protocols and data recording practices, it may limit the generalizability of the findings and introduce center-specific bias. Accordingly, the results should be interpreted as preliminary and specific to this cohort. Further validation using larger, multi-center or registry-based datasets is required to assess the broader applicability and clinical utility of the proposed predictive model.

As the survival dataset contains relatively few samples representing very short and very long survival times, class imbalance remains an important issue. Although SMOTE was applied to mitigate this imbalance, its reliance on the nominal change rate to handle nominal attributes represents a methodological limitation. Alternative approaches, such as the Synthetic Minority Over-sampling Technique for nominal and continuous features (SMOTE-NC), which is specifically designed for datasets containing both nominal and continuous variables, could be explored to further enhance model performance.

To further improve the regression model, which performs less effectively for very short survival times (under 4 months) and very long survival times (over 5 years), likely due to the limited number of training samples in these ranges, several strategies may be considered. First, a two-stage modeling approach could be implemented. In the initial stage, a classifier could separate patients into groups such as “extreme cases” and “typical cases.” In the second stage, separate prediction models could then be trained for each group, allowing each model to focus on subgroup-specific patterns. Tail-focused loss functions could also be adopted by modifying the training objective so that a greater penalty is assigned to errors involving extreme cases. In addition, data augmentation strategies could improve performance by generating more extreme cases. Algorithms such as the Synthetic Minority Over-sampling Technique for regression with Gaussian noise (SMOGN) are specifically designed to address this type of problem.

The results of this research have potential value in several areas, including treatment planning, development of clinical guidelines, and allocation of healthcare resources. Furthermore, these findings may contribute to enhancing clinical services for cervical cancer patients, improving survival outcomes, and supporting healthcare professionals in treatment- related decision-making.