Integrative Analysis of Peripheral Blood Indices for the Renal Sinus Invasion Prediction of T1 Renal Cell Carcinoma: An Ensemble Study Using Machine Learning-Assisted Decision-Support Models

Introduction

Globally, the incidence of renal cell carcinoma (RCC) is increasing annually. According to statistics from the Global Burden of Disease (GBD) database, nearly half of the diagnoses and more than two-thirds of related deaths occur in low- and middle-income countries, where the prognosis is often far from expected.^1,2 Undoubtedly, this increase of RCC is largely attributed to advances in diagnostic imaging techniques, including cross-sectional imaging, especially allowing for the early detection of clinical T1 stage RCC.³ Once diagnosed with T1 stage RCC, clinicians will face a variety of treatment options, including surgical resection and non-surgical treatment.^4,5 For the treatment of stage T1 RCC, according to European and American guidelines, Partial nephrectomy (PN) is currently the standard treatment for T1a tumors (≤4cm) and the first choice for T1b tumors (≤7cm).^6,7 Alarmingly, up to 4% to 13% of patients with T1 RCC treated with PN may have potential adverse pathological features, such as renal or sinus fat infiltration.^8–10 Relatedly, emerging evidence elucidates that renal sinus invasion is tightly associated with inferior oncological outcomes.^9,11 Thus far, developing a predictive model is essential for predicting the risk of renal sinus invasion, but insights into its accuracy and practicality to optimize it are limited.

Peripheral blood is routinely used for preoperative examination in clinics to explore potential predictors, which can improve the convenience of diagnosis. Meanwhile, previous analyses have highlighted that systemic inflammatory markers, hepatic synthetase, and kidney secretion have potential value in predicting the prognosis of various tumors.^12–15 Here we intend to identify the integrative analysis of peripheral blood indices for the renal sinus invasion prediction. Furthermore, the wide usage of various machine learning (ML) approaches has outperformed conventional statistical analyses, which might be increasingly adapted for use in the medical field because of their high accuracy.^16,17 Herein, this study investigated the applicability of machine learning techniques to predict renal sinus invasion in patients with stage T1 RCC.

In the present study, We attempted to integrate these candidate peripheral blood indices to characterize the specific predictive markers that contribute to pathological tumor upstaging, and established multiple predictive models using ensemble analysis, to predict the risk factors for renal sinus invasion. We then analyzed the predictive performance of these machine learning-assisted decision-support models in a deviation cohort and then verified performance in an internal validation cohort.

Patients and Methods

Patients Selection

We retrospectively recruited 1229 cases that had received PN or RN for pathologically confirmed RCC at our center between November 2013 and August 2021, and evaluated their clinicopathological data. The inclusion criteria were as follows: (i) Patients with lesion diameter less than or equal to 7cm, the metastasis of a single regional lymph node cannot be distinguished from that of more than one regional lymph node;^18–20(ii) Patients with complete case data, traceable imaging, pathology, and laboratory examination related indicators;(iii) Patients who underwent RN or PN surgery and had sufficient evidence to prove whether there was renal sinus invasion. The exclusion criteria were as follows: (i) Patients who received preoperative treatment, including thermal ablation, neoadjuvant chemotherapy, or radiotherapy;(ii) Patients with other tumors or psychiatric disorders;(3) Bilateral or multiple renal lesions and patients with defective case data. This retrospective study was in line with the declaration of Helsinki and was ethically reviewed and approved by the Institutional Ethics Committee of the Baotou Cancer Hospital (Reference: 2021-IECBCH027). Since the patient information included in this study was anonymous, the written informed consent of all participants was obtained. The detailed research flow chart was summarized in Figure 1.

Figure 1 The flow chart of patient selection and data process.

Data Collection and Quality Assessment

According to the principle of inclusion of whole candidate variables, we sorted all the variables that can be collected. Among them, if the missing value of the variable exceeds 15%, it will be regarded as an invalid variable and will not be included in the final variable screening. However, provided that the missing range was within 15%, the median or average value filling method will be used to improve the data dimension.²¹ We also collected preoperative routine laboratory measurements, defined as follows: (i) PNI=158-16.6(ALB)-0.78(TSF)-0.20(TFN)-5.80(DHST). ALB: serum albumin (unit: G%); TSF: triceps skinfold thickness (unit: mm); TFN: serum transferrin (unit: mg%), dhst: delayed hypersensitive skin reaction test (dhst = 2 for those with induration diameter >5 mm; dhst = 1 for those with induration diameter <5 mm; dhst = 0 for those without reaction); (ii) HALP = hemoglobin level * albumin level * lymphocyte count/platelet count ratio; (iii) AKI=body mass index*serum albumin/neutrophil-lymphocyte ratio (NLR). NARFIB=neutrophil-to-albumin ratio (NAR) and fibrinogen (FIB) concentration. Combining NAR and FIB, we categorized patients into three “NARFIB” groups: NARFIB1 = NAR <0.083 and fibrinogen <2.28g/L, NARFIB1 = either NAR ≥0.083 or fibrinogen ≥2.28g/L. Finally, a total of 21 variables met the inclusion criteria and were used to build the ML-based model.

Development and Validation of ML-Based Models

We used the caret software package to randomly divide the data set into two parts, 70% for model training and 30% for model testing. A total of five mL-based algorithms were implemented to establish the prediction model. According to the principle of “OOB error”,²² we have gradually screened the model variables, as follows:

The characteristic variable is marked as X and the target variable is marked as Y. The X and Y were evenly divided into two parts, namely X1, Y1, and X2, Y2. Gini index measures the purity of data partition or training tuple set D. Briefly, by sorting the intersection of variable sets, the optimal subset modeling is obtained. The model was evaluated by inspection, discrimination, and calibration. The receiver operating characteristic (ROC) curve was used to evaluate the recognition ability of the prediction model in the training data set and the test data set; The discrimination ability of each model was quantified by the area under the ROC curve (AUC), decision curve analysis (DCA), and clinical impact curve (CIC).

Statistical Analysis

For descriptive analysis, median (IQR) and frequencies (%) were assessed for continuous and categorical variables, respectively. The Chi-square test or the Mann–Whitney U-test as appropriate was used to compare the differences of baseline clinical information between the renal sinus invasion and non-renal sinus invasion cohorts. All analysis was performed using the Python programming language (version 3.9.2, Python Software Foundation, https://www.python.org/) and R Project for Statistical Computing (version 4.0.4, http://www.r-project.org/). All tests were two-sided, and P<0.05 was considered statistically significant.

Results

Baseline Clinicopathological Characteristics of the Study Cohort

The detailed clinical characteristics and pathological baseline data of 1329 patients with T1 stage RCC are summarized in Table 1. Of the patients with T1 stage RCC disease, the rate of renal sinus invasion was 6.98% and 5.01% in the training and validation cohorts, respectively. All baseline characteristic data were compared among the groups of the training and validation cohort. Consistent with the results of the inter-group comparison, as shown in Figure 2A, among the included 21 candidate variables, eight showed significant correlation with renal sinus invasion, namely PLR, NLR, HALP, AKI, PNI, NARFIB, R.E.N.A.L, tumor size, and age. Undoubtedly, systemic inflammatory factors and nutritional parameters suggested that they can be potential predictors of renal sinus invasion.

Table 1 Baseline Demographic and Clinicopathological Features of Included Patients with RCC Diagnosed with and without Renal Sinus Invasion

Figure 2 Variable screening and weight allocation. (A) Correlation matrix analysis of candidate features. (B) The weight distribution of the candidate variables of each ML-based model. Abbreviations: BMI, body mass index; ASA, American Society of Anesthesia; NARFIB, neutrophil-to-albumin ratio (NAR) and fibrinogen (FIB) concentration; R.E.N.A.L, nephrometry score consists of [R]adius [tumor size as maximal diameter], [E]xophytic/endophytic properties of the tumor, [N]earness of tumor deepest portion to the collecting system or sinus, [A]nterior [a]/posterior [p] descriptor and the [L]ocation relative to the polar line; NLR, neutrophil to lymphocyte ratio; PLR, platelet to lymphocyte ratio; NAR, neutrophil to albumin ratio; PNI, prognostic nutrition index; HALP, hemoglobin level * albumin level * lymphocyte count/platelet count ratio; AGR, albumin to globulin ratio; AKI, body mass index*serum albumin/neutrophil-lymphocyte ratio.

Selection of Candidate Variables via Different ML-Based Algorithms

According to the variable screening results of different ML-based algorithms, we applied the intersection of meaningful variables of each model (Figure 2B). The flow sorting in the chord graph indicated the contribution of each variable in the model, which can show that PLR, NLR, HALP, AKI, PNI, NARFIB, R.E.N.A.L, and age occupy a high weight in the five models. Consistent with the chord graph results, we found that the intersection variables mentioned above were also in the top 10 rankings of RFC and DT models (Supplementary Table 1). These results suggested that repeated iterative algorithm analysis, including NAR, PLR, NLR, HALP, AKI, PNI, NARFIB, R.E.N.A.L, and tumor size, can be used to further model construction and predict renal sinus invasion.

Construction of ML-Based Renal Sinus Invasion Predictive Model

As representatives of ML-based algorithms, RFC and DT are commonly used modeling schemes in supervised learning. As shown in Figure 3A, the RFC model was constructed using the following formula: I(X=x_i)=-log₂P(x_i), the I(X) was used to represent the information of candidate variables, and P (x_i) refers to the probability when x_i occurs. Considering the clinical utility, the variables with 10 ranking at the top were determined to construct the optimal RFC prediction model, which included NAR, PLR, NLR, HALP, AKI, PNI, NARFIB, R.E.N.A.L, albumin to globulin ratio(AGR), age, and tumor size. As depicted in Figure 3B, with the addition of inflammatory factor indicators, relevant NLR and NARFIB acted as an irreplaceable weight at the branch of DT. Collectively, using the iterative algorithm analysis of supervised learning, both RFC and DT models were appropriately used to guide renal sinus invasion prediction.

Figure 3 Predictive model visualization based on ML-based algorithm. (A) RFC model. (B) DT model. Note: The candidate factors associated with renal sinus invasion were ordered via RFC algorithm(A) and (B)prediction node and weight was allocated via DT algorithm.

Comparison Between ML-Based Models

Based on the iterative analysis of the baseline characteristics, we further used five supervised learning models for renal sinus invasion risk assessment, attempted to explore whether we can improve the prediction performance. As expected, the RFC model can better distinguish whether patients with LACC showed renal sinus invasion or not. As shown in Figure 4, the DCA also exhibited that the RFC model was equipped with a robust prediction performance in the training and validation cohorts, respectively. Additionally, the AUCs of RFC models reached a plateau when 10 variables were introduced, followed by DT, ANN, SVM, and XGBoost. The detailed predictive performance of ML-based models is summarized in Supplementary Table 2.

Figure 4 Prediction performance of candidate models based on ML-based algorithm. (A) DCA for five ML-based models in the training set. (B) DCA for five ML-based models in the testing set. Abbreviations: RFC, random forest classifier; SVM, support vector machine; DT, decision tree; ANN, artificial neural network; XGboost, eXtreme gradient boosting.

Internal Validation of the Optimal Renal Sinus Invasion Predictive Model

To further validate the performance of the RFC model, we also adopted CIC to evaluate the prediction efficiency, as illustrated in Figure 5A, the CIC demonstrated that the stratification of renal sinus invasion could be distinguished in the training cohorts. These results were also parallel to risk factors of renal sinus invasion delineated in the validation cohorts (Figure 5B), indicating that the selected features were highly relevant to renal sinus invasion.

Figure 5 Prediction performance of RFC model via CIC. (A) Training set. (B) Validation set. Note: The dark green line predicts the probability of poor renal sinus invasion, and the purple line shows how many patients will be at high risk of non-renal sinus invasion.

Discussion

Renal sinus invasion is the main way of extrarenal expansion of renal cell carcinoma, and its prognosis may be worse than perirenal fat invasion.^23,24 Herein, the preoperative evaluation of renal sinus invasion is of great significance for the selection of radical nephrectomy and nephron-sparing surgery and the prediction of postoperative prognosis. In the present study, using five computational ensemble approaches, we identified 10 associated with pathological tumor upstaging, including NAR, PLR, NLR, HALP, AKI, PNI, NARFIB, R.E.N.A.L, AGR, age, and tumor size, which has, to our knowledge, not previously been tested in predicting renal sinus invasion via ML-based algorithm. We also evaluated the predictive ability of the renal sinus invasion. Notably, we observed high concordance in effect estimates between training and validation datasets, highlighting that R.E.N.A.L alone proved to be a weak predictor (receiver operating characteristics area under the curve (ROC-AUC) 0.65) for renal sinus invasion, but the RFC performed substantially better than conventional predictive models, which is the most robust serological index associated risk factor (ROC-AUC 0.924). Meanwhile, adding systemic inflammatory factors and nutritional parameters to the RFC model appreciably improved the prediction accuracy.

On separate iterative analysis, preoperative collectible clinical indicators were used for variable screening. Previous studies have shown that systemic inflammatory factors and nutritional indicators, as candidate predictive markers, were more convenient and cost-effective than the invasive operation or imaging examination, which have been widely used in clinical practice.^25–27 Liu and colleagues et al reported that systemic inflammation response markers and serum lipid parameters in predicting pT3a upstaging may be greatly beneficial to urologists in risk stratification and management decisions, which are similar to the results of this study.²⁸ In contrast, machine learning is superior to the traditional prediction model in screening and using these variables. In this study, we also introduced some classical prediction indicators, including AKI and HALP. Previous studies have shown that HALP can comprehensively reflect the inflammatory response and immune status of tumor patients, which has been reported in the prognosis of various solid tumors, such as renal cell carcinoma, gastric cancer, small cell lung cancer, bladder cancer, pancreatic cancer, rectal cancer, and prostate cancer.^29–33 Meanwhile, Jafri and co-workers et al developed the advanced lung cancer inflammation index to evaluate the systemic nutrition and inflammation index in patients with metastatic non-small cell lung cancer at the time of diagnosis.³⁴ Based on this research idea, we created the AKI score to participate in predicting renal sinus invasion. The results revealed that after adding the AKI score, the robustness and accuracy of the prediction model were significantly improved, suggesting that the role of this index in renal sinus invasion should not be underestimated.

Nowadays, to meet the extensive needs of clinicians for auxiliary early diagnosis, machine learning-based classification is the most popular computer development. As an important branch of supervised learning, the RFC model has been successfully applied to high-dimensional multi-source data reduction in many medical fields.³⁵ Robust and reliable supervised learning classifiers, including support vector machine, random forest, convolutional neural network, and decision tree, have been gradually applied to clinical practice and model construction prediction. Consistent with the results of previous research reports, for feature selection and classification, we find that the RFC model is better than the univariate model β non-iterative linear regression model has more prediction advantages.^36,37 Based on the candidate value variables obtained by multi-layer iteration, we constructed the relevant prediction model. These algorithms were equipped with more accurate and higher diagnosis and prediction performance than the non-iterative algorithm.

Interestingly, the variables obtained by RFC screening are almost the same as those obtained by the DT algorithm, but the potential prediction performance needs to be further optimized, which is suitable for the optimal strategy of clinical individualized diagnosis and treatment. Moreover, CIC indicated that the ML-based algorithm had a similar performance to an ideal model, and the DCA results also demonstrated that our nomogram had a high potential for clinical utility. Taken together, we found that RFC was more suitable for predicting renal sinus invasion via supervised integrated learning, which can timely predict the risk of renal sinus invasion in patients with RCC, to better obtain the guidance of clinical shunt management and avoid misdiagnosis or excessive medical treatment.

Nevertheless, some issues need to be taken into account. First, since the collection of clinical information for the renal sinus invasion of patients with RCC is difficult, at least data integrity and non-selective offset cannot be guaranteed, as well as perform additional external validations. Second, we obtained the most relevant predictive modules for model building, future studies should explore potential markers for renal sinus invasion prediction, such as proteomics, genomics, and provide new clinical auxiliary strategies for T1 stage RCC treatment and renal sinus invasion prediction. Third, despite a well-established vital predictive model via ML-based algorithm, our conclusion may not be persuasive and still needed more clinical practice to verify its feasibility in the future, especially identifying individuals at high risk of renal sinus invasion.

Conclusion

Collectively, our findings provide insights into the candidate predictive markers associated with a high risk of renal sinus invasion via peripheral blood indices, how they can be manipulated to develop ML-based predictive models, and the prediction performance of these unsupervised ensemble models might be at a population-based level. We anticipate that these findings will extend external studies seeking to strengthen valuable adjunct information and guide treatment decisions.