2.1 Study population

A total of 266 CRC patients diagnosed between October 2016 and September 2019 at the Salah Azaiez Institute of Tunis (Tunisia) were recruited. The patients’ demographic and tumor-related characteristics are presented in Table 1. OS status was assessed over three years following conventional treatment. Ethical approval was obtained from the Institute’s Ethics Committee (ISA/03/2016). All participants provided written consent. All methods followed relevant guidelines and regulations.

2.2 Polymorphism genotyping

Genomic DNA extraction and amplification of HLA-G gene polymorphisms were performed as described in Dhouioui et al., 2022. Briefly, PCR was used for the HLA-G 14-bp Ins/Del (rs371194629) polymorphism, and PCR-RFLP for the HLA-G + 3142C/G polymorphism (rs1063320). Based on Dhouioui’s cases-controls study (Dhouioui et al., 2022) on CRC susceptibility, and work by Gagno (2020) and Gusella (2017) (Gagno et al., 2020, Gusella et al., 2017), a risk score was assigned to each genotype: for the HLA-G 14 bp Ins/Del: Ins/Ins = 0, Ins/Del = 1, Del/Del = 2, with Del/Del as the risk genotype (OR = 1.80, 95 % CI = 1.205–2.664, p = 0.003; (Dhouioui et al., 2022)). For HLA-G + 3142C/G: G/G and C/C = 0, C/G = 1; with C/G as the risk genotype (OR = 1.59,95 %CI = 1.106–2.272,p = 0.013; (Dhouioui et al., 2022)).

2.3 Machine learning models

Decision Tree (DT)-based algorithms were used to predict OS. Decision trees were selected for their effectiveness with missing data and mixed data types commonly found in medical datasets, simplifying preparation. This study focused on HLA-G polymorphisms and integrated clinical data for OS prediction.

The proposed methodology, described in Fig. 1, is a two-step approach combining multiple correspondence factorial analysis (MCFA) and a DT variant. MCFA highlighted correlations between HLA-G polymorphisms and the other features (Table 1), helping improve model performance by identifying influential features, stabilizing tree structure, and reducing training time.

Close

Fig. 1

Workflow of the proposed research study. This figure depicts the workflow that was followed throughout the project implementation. The research was conducted through for phases. During the first phase, the most relevant features for predicting survival status along with adequate ML model were identified. In the second phase, a more refined analysis dealing with the impact of genotypes in informing about overall survival status was carried out according to qualitative and ordinal-based methodologies. Afterwards, we explored the importance of haplotypes that result from different combinations of genotypes to ensure the same predictive task. Finally, the study was extended by incorporating several best features that we obtained from the three previous phasis and investigating the performance of the newly obtained set of data in empowering the ability of the RF model in accurately predicting the overall survival status. We also used the Logistic regression model to further refine outputs and develop conclusive results.

Export to PPT

We evaluated six tree-based models, selecting the top performer among DT, Random Forest (RF), Bagging Classifier (BC), Gradient Boosting Tree (GBT), Stochastic Gradient Boosting Tree (SGBT) and Adaptive Boosting classifier (Adaboost). Except for DT, these are ensemble models combining weak learners into a strong predictor, with DT-based models supporting repeated feature evaluation and scenario exploration for better decision-making (Abdullah et al., 2017).

We also investigated Support Vector Machine (SVM), a supervised algorithm that is suitable for small datasets, effective for medical applications (Venkatesan et al., 2018), and improves specialized recommender systems (Matthew et al., 2022).

2.4 Hyperparameters optimization and model training

When working with ML models, effective parameter optimization technique is important because the default parameters are not always optimal. The idea is to find the best hyperparameters for the data to achieve the optimal performance. A large number of optimization techniques are available in the biomedical literature, including the GridsearchCV (Liashchynskyi and Liashchynskyi, 2019, Kaushik et al., 2020), the evolutionary algorithm (Koohestani et al., 2019, Abdar et al., 2019), the Practical Swarm Optimization (Zomorodi et al., 2021), the Random Search (Liashchynskyi and Liashchynskyi, 2019), and the Bayesian Optimizer (Hosni et al., 2019). In this study, we used the GridSearch Cross Validation (hereafter GridsearchCV) algorithm to optimize key modeling parameters. GridsearchCV systematically explores predefined hyperparameter grids and uses cross-validation to determine the best combination for the training task.

2.5 Forecasting overall survival status

To explore these results, we focus on the relevance of the use of HLA-G polymorphisms in predicting OS status, and clearly distinguishing between ‘Alive’ (Survival) and ‘Dead’ (Mortality) OS status modalities. This is the goal of the second stage of the proposed framework (Fig. 1, block 7).We performed a comparative study in three scenarios: the first including all predictive features (Table 2-scenario 1), the second is linked to HLA-G polymorphisms and the most relevant features upon MCFA findings (Table2-scenario 2), and the third is restricted to HLA-G polymorphisms (Table2-scenario 3).To ensure the robustness of the results, we emphasize the use of pre-specified ML methods.

2.6 Hyperparameter tuning

We used the GridSearchCV algorithm to optimize four key parameters of the DT model: number of leaf nodes, tree depth, minimum samples per leaf, and minimum number samples needed for a split, while also evaluating the Gini and Entropy split criteria.

The number of leaf nodes prevents overfitting by limiting the terminal nodes, while the minimum samples per leaf ensures sufficient data at each leaf to simplify the tree. The tree depth impacts complexity; a deeper tree captures intricates patterns but risks overfitting. The minimum samples for a split ensures enough data at nodes thus setting the threshold for further divisions.

GridSearchCV helps in testing combinations of « max_leaf_nodes », « min_samples_leaf » and « min_samples_split » from 1 to 20, and « max_depth » from 1 to 10. This determines the optimal condition for the tree structure. The final tree model is created from the best parameter combination with the highest cross-validation score using 70 % of the training data.

2.7 Performance evaluation

3.1 HLA-G polymorphisms and clinicopathologic features are key predictive features of overall survival status in patients with CRC

The MCFA was used to investigate associations between OS status and the predictive features with a focus on HLA-G variants. The contribution coefficients of each factor used to construct the factorial axis and information on whether they are significant are provided in the F1 and F2 columns of Table 1. The bolded coefficients indicate features that validate the p-value test and assess the statistical significance of a particular feature on the basis of 5 %-error risk. OS status was found to be significant in terms of the second factorial axis along with the following predictive features: HLA-G + 3142C/G-C/C, HLA-G 14bpIns/Del-Del/Del, tumor location-rectum, histological subtypes-Other, histological subtypes adenocarcinoma, HLA-G 14 bp Ins/Del-Ins/Del, HLA-G + 3142C/G-G/G and grade status-Moderate, Stages-I + II and Stages-III + IV. These factors are the origins of the latent factors given by Axis 2. To gain deeper insights, Table 1 presents information about the linkages between each OS status and the remaining significant key features. These linkages are categorized on the basis of their relationship with the OS status modality. Specifically, features linked to OS Status-“Dead” are bolded whereas those associated with OS Status-“Alive” are written in italics. Accordingly, the HLA-G + 3142C/G-C/C variant was the most commonly associated predictive feature, followed closely by OS Status-“Dead” and HLA-G 14 bp Ins/Del-Del/Del. Conversely, concerning OS Status- “Alive”, both HLA-G 14 bp Ins/Del-Ins/Del and + 3142C/G-G/G were closely linked after the histological subtypes-adenocarcinoma feature.

3.2 HLA-G polymorphisms can efficiently predict overallsurvival status via the random forest classifier

We trained the selected ML classifiers on the dataset, while optimizing hyperparameters. The objective was to investigate the role of HLA-G in predicting OS status in CRC patients. We simulated three scenarios: testing classifiers with all available predictive features, using insights from the MCFA findings which highlight feature linkages, and focusing solely on HLA-G polymorphisms.

In the first scenario, all top-performing models were evaluated (Table 2-scenario 1, Fig. 2a). Notably, mixed findings showed that competing classifiers, particularly SGB and BC, demonstrated similar performance.

Close

Fig. 2

AU-ROC curves for each Machine Learning model. Scenario 1: All predictive features, Scenario 2: HLA-G polymorphisms + most relevant features, Scenario 3: Only HLA-G polymorphisms. The AU-ROC curve measures the discriminatory performance of the machine learning model to correctly predict the right class of patients whether they would grasp living or dead overall survival status, under three different scenarios. As may be noticed from Scenarios 1 and 2, the RF classifier compared more favorable than the remaining peer methods since it yields the nearest score to 1 which suggests that the proportion of true positives is higher than that of false positives. As such, it could be stated that RF enables more effective distinguishing between “Alive” and “Dead” patients compared to that yield by the remaining classifiers. When considering HLA-G polymorphisms as the only set of predictors Scenario 3), the discriminatory performance of the RF decreases to 0.5. This may be explained by the fact that reducing predictive features alters the learning performance of the classifier. The AU-ROC curve is effective in reflecting the discriminatory power of a classifier. However, it should be pointed out that this performance relies upon the interaction between “True positive rate” and “False positive rate” and ignores valuable information yield by the remaining classification metrics “True negative rate” and “False negative”. Thereby, we make use of the confusion matrix to help to capture more comprehensive information. DT: Decision Tree classifier, RF: Random Forest, BC: Bagging Classifier, GBT: Gradient Boosting Tree, SGBT: Stochastic Gradient Boosting Tree, Adaboost: Adaptive Boosting classifier, SVM: Support Vector Machine.

Export to PPT

In the second scenario, the RF performs well in both classification and statistical measures (Table 2-scenario 2). The AU-ROC curves in Fig. 2.b highlight significant differences in classification and AU-ROC performance among the remaining classifiers, likely due to their high sensitivity to available features. This raises questions about the robustness of these models, with RF emerging as the most stable option.

Strong predictive performance was achieved using only HLA-G polymorphisms. To explore this, we conducted a third scenario including only HLA-G polymorphisms in the classification model to predict the OS status. This targeted approach allowed us to assess the significance of HLA-G polymorphisms in predicting patients’ OS status. The classification and statistical results in Table 3 are comparable to those in the second scenario. Nevertheless, there are notable differences in the AU-ROC and RMSE indicators (Fig. 2c). We found significant variability across classifiers with slightly reduced performance overall. Nevertheless, the RF classifier remains the most favorable in this context.

Finally, the RF model excels in predicting the OS status of CRC patients in terms of both classification and statistical metrics. Throughout the three scenarios involving the most relevant input features, we obtained conclusive findings. These refined scenarios contributed to enhancing the discrimination performance of the classification models. Overall, our study demonstrates the superiority of the RF model and highlights the significance of incorporating relevant input features for more accurate CRC risk assessment.

3.3 Performance of the HLA-G genotypes in accurately predicting global overall survival status

To gain deeper insights into the established performance, we conducted a two-step analysis aimed to uncovering the role of HLA-G genotypes in accurately predicting the OS status of patients (Fig. 1, block 2). This analysis involved a two-step approach:

• Qualitative analysis: First, we assessed the predictive power of the HLA-G genotypes for accurately predicting OS status by MCFA (Fig. 1, block3). The goal was to identify the most relevant genotypes among the following options Del/Del, Ins/Del, Ins/Ins, C/C, C/G, and G/G. By evaluating the performance of each genotype, we aimed to identify which had a significant role in predicting patients’ OS outcome.

• Ordinal-based analysis: We assessed the impact of genotype risk factors associated with each HLA-G polymorphism in a process termed “ordinal-based” which involves assigning numerical scores to risk scores and cumulative risk. By quantifying these risk scores, we aimed to gain deeper insights into their contribution to predicting OS status.

3.3.1

3.3.1 Qualitative analysis

3.3.1.1

3.3.1.1 *Del/Del genotype is highly linked to “Dead” survival status and G/G genotype is highly linked with “Alive” survival status

Fig. 3a depicts the result of applying multiple factorial analysis according to which all features were included along with their modalities.

Close

Fig. 3

Investigating the interactions between predictors and the overall survival status using the Factorial Component Analysis (FCA). This is a machine learning method for dimensionality reduction that we explored to identify relevant features for accurately predicting survival status. In Fig. 3.a, we present the results considering all predictive features, which reveal a stronger interaction between HLA-G 14pb Ins/Del and survival status. We found that both G/G and Ins/Ins genotypes have a significant correlation with the “Alive” survival status. This observation is supported by the acute angles formed by the vectors of each genotype and the vector representing the living survival status. Conversely, our results indicate that both C/G and Ins/Del genotypes are significantly associated with the OS_Dead mortality. Remarkably, we observed that the HLA-G 14 bp Ins/Del genotype plays a more crucial role in accurately predicting survival status compared to the HLA-G + 3142C/G genotype. This finding is supported by graphical representations, illustrating that the HLA-G 14 bp Ins/Del genotypes are visually closer to the survival status modalities in Fig. 3.b. In Fig. 3.c, we investigate the interactions between Risk factors and overall survival using the FCA analysis. This figure quotes the FCA results that Investigate the interactions between risk factors of each of HLA-G bp14 Ins/Deland HLA-G + 3142C/G polymorphisms predictors and the survival status. A key finding reveals that HLA-G bp14Ins/Del is more linked to survival status. In particular, a rising value of it informs about worse survival status. Fig. 3.d deals with the interaction between haplotype and survival status using FCA analysis. The FCA shows that the Survival Status target feature is better represented by the second principal component, as it exhibits higher coordinates compared to the first principal component. This indicates that exploring this axis could lead to more reliable interpretations. Accordingly, we observed that the DelG haplotype is negatively related to the Survival status, suggesting a possible link to worse survival outcomes. On the contrary, the InsC haplotype appears to be the most distinct from the target feature and may be associated with better survival status.

Export to PPT

In particular, the “Dead” OS status is strongly correlated with rectal tumor location, stages III + IV, Del/Del genotype, and C/C genotype, although the latter shows a weaker linkage. The angles formed by the Ins/Del genotype with “Alive” or “Dead” OS status are acute, whereas those formed by the C/G genotype are obtuse. This suggests a more substantial interaction between the Ins/Del genotype and OS status.

The HLA-G 14 bp Ins/Del and HLA-G + 3142C/G polymorphisms are expressed according to their respective genotypes. Interestingly, both principal components capture approximately 44.5 % of the total variance in the dataset, with OS status closely aligned with the first component. This indicates that the principal component 1 effectively summarizes the variations of this variable, and thus all associations with other features are meaningful. Specifically, the OS status “Dead” (Mortality) shows strong correlations with the following features: tumor location-rectum, stages III + IV, the Del/Del genotype and the C/C genotype but with a weakest linkage for this last predictive feature.

We observed that the angles formed by the Ins/Del genotype “Alive” or “Dead” OS status are mostly acute, while those formed by the C/G genotypes are more obtuse. These vectors originate from the center of the factorial plane and point towards their respective variable modalities. This finding suggests a more substantial interaction between HLA-G 14pb Ins/Del and OS status. To test this assumption, we replicate the MCFA analysis considering only genotypes as input variables. Fig. 3.a shows that OS status gains importance in the second principal component with the G/G and Ins/Ins strongly linked to the “Alive” OS status, as indicated by the acute angles formed between their vectors and that of “Alive”. Similarly, we assert the potential linkage between genotype Del/Del and OS status “Dead” as they reveal positively correlated according to the first principal component, while it shows a negative correlation with the “Alive” status.

Conversely, the results show that both genotypes C/G and Ins/Del are significantly associated with the OS status-Dead (Mortality). Interestingly, the HLA-G 14bpIns/Del genotype holds more relevance in predicting the OS status than the HLA-G + 3142C/G genotype. This result is supported by graphical representations in Fig. 3.b, which shows the HLA-G 14bpIns/Del genotypes visually closer to the OS status modalities.

3.3.1.2

3.3.1.2 *HLA-G 14 bp Ins/Del is a better predictor of survival status than HLA-G + 3142C/G using the Random Forest classifier

We then evaluated the predictive performance of each genotype for accurately predicting OS status (Fig. 1, block 6). By treating each genotype from both HLA-G polymorphisms as independent input features instead of associating them with specific modalities, we could analyze the impact of each individual genotype on outcomes separately, providing a more comprehensive understanding of their predictive power. The experiment was carried out across five different scenarios each varying the set of input features. Lines 1 to 6 from Table 3 summarize the classification performance (accuracy, precision, recall, F1-score) and statistical error performance (Log ROC, MSE) under different scenarios.

Using only genotypes as learning features led to lower performance compared to scenarios with additional predictors, as ML typically perform better with more data features. However, when including HLA-G 14 bp Ins/Del genotypes with other features, it outperformed its peers, showing more favorable results across all classification and statistical metrics. Thus, HLA-G 14 bp Ins/Del genotypes proved to be more significant for forecasting than HLA-G + 3142C/G genotypes. Comparing the effectiveness of HLA-G14 bp-Ins/Del genotypes across different scenarios (Table 4 − scenarios b and c), the statistical AU-ROC was optimal when all predictors, including the genotypes of HLA-G 14 bp Ins/Del were considered together. However, the RF classification model performed best when these genotypes were considered as modalities rather than separate features. Therefore, we recommend integrating HLA-G14 bp Ins/Del polymorphism genotypes with the remaining features to enhance prediction, classification, and AU-ROC performance.

Table 4 The logistic regression (LR) results on the interaction betweenpredictors and Survival Status.

Scenario	Features	Coefficient in LR Equation	Odds Ratio
A: On the interaction between risk factors of genotypes and Overall Survival Status	Risk factor HLA-G14bp Ins/Del (P1)	0.125 (0.537)	1.130 (1.712)
	Risk factor HLA-G + 3142C/G (P2)	−0.594 (−0.724)	0.552 (1.064)
B: On the interaction between haplotype and Survival Status	Haplotype InsC	−0.446 (−0.542)	0.640 (0.508)
	Haplotype InsG	−0.220 (−0.308)	0.804 (0.734)
	Haplotype DelG	−0.091 (0.234)	0.913 (1.264)
	Haplotype DelC	−0.346 (0.408)	0.708 (1.504
C: Combining key clinical features with genotypes Del/Del	Tumor location	−0.501 (−0.448)	0.606 (1.565)
	Grade	0.029 (0.102)	1.029 (1.107)
	Histological subtypes	0.751 (0.714)	2.118 (0.839)
	Stage	1.633 (1.251)	5.121 (3.495)
	Genotype Del/Del	0.051 (0.379)	1.053 (1.460)
D: Combining key clinical features with genotype G/G	Tumor location	−0.6059 (−0.051)	0.546 (0.950)
	Grade	0.014 (0.242)	1.014 (1.242)
	Histological subtypes	0.5664 (0.187)	1.761 (1.828)
	Stage	1.264 (1.508)	3.540 (2.436)
	Genotype G/G	−0.605 (−0.809)	0.546 (4.519)
E: Combining key clinical features with Haplotypes InsC and DelG	Tumor location	−0.549 (0.003)	0.577 (1.002)
	Grade	0.0244 (0.218)	1.020 (1.244)
	Histological subtypes	0.811 (−0.159)	2.250 (0.853)
	Stage	1.640 (0.778)	5.180 (2.177)
	Haplotype InsC	−0.081 (−0.610)	1.174 (0.543)
	Haplotype DelG	−0.549 (1.744)	0.923 (1.723)
F: Combining key clinical features with Risk factor P1	Tumor location	−0.516 (−0.100)	0.597 (0.923)
	Grade	0.017 (−0.525)	1.020 (0.681)
	Histological subtypes	0.784 (−0.155)	2.190 (0.876)
	Stage	1.610 (1.851)	5.010 (2.299)
	Risk factor P1	−0.107 (−0.011)	0.898 (0.635)

3.4 The InsC haplotype is predictive of favorable overall survival, whereas the DelG haplotypes is associated with a poorer overall survival outcome

This analysis aims to assess the effectiveness of haplotypes derived from HLA-G polymorphisms in predicting the OS status of CRC patients (Fig. 1, Blocks 5 and 6). We used the FCA to explore the correlation between various predictors and the target feature (Fig. 3.d).

We demonstrated that the OS status is best represented by the second principal component, which has higher coordinates than the first principal component does, suggesting that this axis should be explored for more reliable interpretations. The DelG haplotype is closest to the OS status, suggesting that it may be associated with a poor OS status. In contrast, the InsC haplotype appears most distant from the target feature and may indicate a better OS status. To get more rational conclusions about the observed performance, we referred to logistic regression (Table 4, Scenario B). The results show that all haplotypes have negative coefficients, suggesting an inverse relationship with OS status. As the haplotype value increases, the OS status improves.

Notably, the coefficient of the DelG haplotype (−0.09) is smaller than that of InsC (−0.446), and its effect on the log-odds is less significant. The odds ratio for DelG (0.91) is higher than that for InsC (0.64), meaning that the odds of the event decrease more slowly with DelG than with InsC. Thus, an increase in DelG or InsC haplotypes decreases the probability of reaching the target, but the decline is greater for InsC, as it has a larger coefficient.

In conclusion, both the DelG and InsC predictors reduce the likelihood of reaching the target, but the impact of InsC is greater than that of DelG. These findings are consistent with prior with previous results showing that InsC improves OS status, while DelG worsens OS status with its minimal improving effect.

3.5 A combined model integrating clinical features and HLA-G polymorphisms for survival: Del/Del and G/G genotypes associated to key clinical features made correct predictions with 100 % precision score

This section focuses on the benefits of combining each of the top emerging predictive feature with most relevant clinical peers that have been identified through the MCFA analysis within phase 1(Fig. 1). This approach aims to enhance the understanding on the polymorphism’s detection properties and support decision makers in selecting the appropriate prognostic biomarker, including genotypes, associated risk factors or haplotypes. Blocks 6, 7 and 8 in Fig. 1 outline the performed steps. We first designed four scenarios (subset 1–4: Table 4 C-F) given by concatenating the possible combinations between each subset of best polymorphisms and the set of best clinical features: “Grade”, “Stage”, “Tumor location” and “Histological Subtypes”. Subset 1 encompassed the Del/Del genotype of the HLA-G 14 bp Ins/Del polymorphism (Table 4 C). Subset 2 encompassed the G/G genotype of HLA-G + 3142C/G (Table 4D). Subset 3 integrated haplotypes InsC and DelG (Table 4 E). Subset 4 encompassed the risk score associated to HLA-G 14 bp Ins/Del (“Risk factor HLA-G14bp Ins/Del”) (Table 4 F). Afterwards, the RF analysis was conducted using each of the obtained subsets, with results quoted in Table 3 (lines 8, 9, 10 and 11for the respective subsets). The performances for subsets 1, 2 and 3 were similar with a preference to the scenario involving either the Del/Del or G/G genotype. The classification metrics indicated that RF achieved an accuracy of about 75 %, which means it correctly classified around 75 % of the data. Using haplotypes, the accuracy was slightly lower, at around 0.7 %. Interestingly, in the three best scenarios, the RF achieved a 100 % precision score, correctly predicting all positive instances without mislabeling others as “Dead.”. Thus, it could be stated that the RF classifier is perfectly capable of detecting the “Dead” OS status using these combinations of features, without falsely labeling other patients as “Dead”. However, a considerable number of true positives were missed, especially when using the haplotype subset, where only 5.26 % of actual positive instances were identified compared to 14 % for the G/G genotype subset. Overall, we showed incomplete coverage of the “Dead” OS status (Mortality) in both scenarios. Essentially, the weak recall performance reflects the challenge of identifying all patients with a worse OS status, while a high precision indicates that the RF classifier is highly confident in its prediction predicts accuracy when predicting a patient as “Dead”.

The AU-ROC metric indicates the superior discrimination performance of the RF classifier when trained on the subset including Ins/Ins and G/G genotypes compared to competing scenarios. Moreover, this specific combination of features allowed the RF model to achieve the lowest Mean Squared Error (MSE) statistics among all evaluated subsets.

Considering the competing results of the three subsets, we will proceed with a logistic regression analysis to gain additional insights and performance preferences.

Overall, in this context, the trade-off between precision and recall scores is appropriate. Indeed, as we are operating in medical diagnosis context, we are concerned with having trustworthy predictions on “Dead” instances more than merely achieving comprehensive coverage of positive cases.

The high precision score prevents critical situations where false positives could have serious consequences in a clinical practice. The AU-ROC metric enacts that the RF classifier performs better with the Ins/Ins and G/G genotypes, achieving the lowest MSE statistics compared to all remaining peer scenarios. Given the competing results, we performed logistic regression analysis to further explore the performance characteristics and enhance our understanding of how each predictive feature impacts OS status.

The logistic regression model quantifies the impact of each feature on the OS status, with results reported in Table 4 for scenarios C, D, E and F for concatenating subsets 1, 2, 3 and 4 respectively.

The findings reveal a significant linkage of the feature “Stage” to the OS status. Indeed, a-one-unit increase of the “Stage” value brings the odds ratio near the “Dead” OS status by about 443 %.

In scenario “C” the Del/Del genotype is positively associated with the worse OS status with an odds ratio indicating that individuals with this genotype have about 5.3 % higher likelihood of experiencing poor OS. This emphasizes previous findings highlighting the HLA-G 14 bp Ins/Del polymorphism's role in predicting OS, most notably is the linkage of Del/Del and mortality. In scenario “D”, that the G/G is negatively associated with the OS status, with odds ratios indicating that individuals with G/G genotype have about 61.96 % higher likelihood of experiencing better OS status. This supports earlier results regarding the informative properties of the HLA-G + 3142C/G polymorphism.

Scenario “F” combines clinical factors with “Risk factor HLA-G 14 bp Ins/Del”, a risk score inherited from HLA-G 14 bp Ins/Del. Previously, a weak correlation was shown between “Dead” OS status and “Risk factor HLA-G 14 bp Ins/Del”. Logistic regression analysis confirms this with a negative coefficient, suggesting that a one-unit increase in this predictor is associated with approximately 10.2 % odds of experiencing worse OS status.

In scenario “E”, the combination of haplotypes shows that DelG is positively associated with the “Dead” OS status whereas InsC is negatively associated. Particularly, individuals with the DelG haplotype are approximately 17.4 % more likely to experience a worse OS status, whereas those with InsC haplotype have about 8.1 % less likely to experience this OS status. These findings are consistent with previous analyses showing that InsC haplotype tend to be a better predictor of good OS, whereas DelG tends to correlate with worse OS status. Overall, these findings highlight the insights gained in the four phases of our study and clarify the relative impact of each genetic polymorphism. Specifically, the genotype-based study provides the most reliable outcomes as demonstrated by the high odds ratios for Del/Del and G/G. Thus, the predictive properties of these genotypes prove to be the most reliable.

Finally, we highlight that the robustness of our results has been is supported by rigorous external validation procedures, which are shown in Tables 3 and 4 in italics and subsequent to each scenario.