Interpretable Machine Learning Models for Predicting Lateral Pelvic Lymph Node Metastasis in Rectal Cancer: A Chinese Multicenter Retrospective Study

Abstract

PURPOSE

Internal iliac and obturator lymph nodes are common sites of metastasis in rectal cancer. This study developed a machine learning (ML) model using clinical data to predict lymph node metastasis and applied the Shapley Additive explanations (SHAP) method for interpretation.

MATERIALS AND METHODS

Retrospectively, data from patients with rectal cancer at four Chinese centers—who underwent total mesorectal excision and lateral pelvic lymph node dissection without neoadjuvant therapy—were collected. Two centers provided training/test sets (3:1 ratio) and two centers supplied external validation. Lymph node enlargement was determined by imaging and confirmed by pathology. Five ML models were evaluated by AUC, accuracy, and F1 score. Key features included demographics, tumor stage, tumor-to-anal verge distance, imaging measurements, tumor histological differentiation, preoperative carcinoembryonic antigen, and carbohydrate antigen 19-9. SHAP was used to assess feature importance.

RESULTS

Of the 411 cases (174 positives) in the training/test sets and 109 cases (43 positives) in external validation, the random forest (RF) model ranked second in terms of AUC and accuracy in the training set (0.999, 0.995), whereas it achieved the highest AUC and accuracy (0.877 and 0.788) in the test set. In the external validation, the RF model outperformed all other ML models (AUC of 0.899, accuracy of 0.827). Overall, the RF model demonstrates the superior overall performance. According to the SHAP analysis, the most important predictors of internal iliac and obturator lymph node metastasis were, in descending order, the short-axis diameter of enlarged lymph nodes, regional lymph node metastasis, and tumor-to-anal verge distance. At the individual patient level, SHAP force plots provided explanations of the RF model predictions for internal iliac and obturator lymph node metastasis.

CONCLUSION

An interpretable ML model was developed that accurately predicts internal iliac and obturator lymph node metastasis using clinical data. SHAP analysis enhances understanding of feature contributions, supporting personalized treatment planning.

INTRODUCTION

Rectal cancer is one of the most common malignancies globally.¹ Despite significant advancements in screening, surgical techniques, and adjuvant therapies, mortality for rectal cancer remains high, particularly among patients with lymph node metastasis.^2,3 Lymph node involvement, within the lateral pelvic region, plays a critical role in disease progression, prognosis, and treatment decision making.⁴ Among lateral pelvic lymph nodes (LPLNs), metastases to the internal iliac and obturator nodes occur most frequently and correlate with higher recurrence and poorer survival.^5-7 In most studies and clinical guidelines, internal iliac and obturator nodes are classified as regional lymph nodes, whereas metastases to the external iliac and common iliac nodes are considered distant, which necessitates distinct treatment approaches.^6,8

CONTEXT

• Key Objective

• This study aimed to develop and validate a machine learning model based on clinical data to predict the metastatic status of internal iliac and obturator lymph nodes in rectal cancer and to provide interpretability using Shapley Additive explanations (SHAP) analysis.

• Knowledge Generated

• The random forest model achieved the best predictive performance, yielding a receiver operating characteristic AUC of 0.877 in the test set and 0.899 in the external validation cohort. SHAP analysis identified short-axis diameter and regional lymph node status as top predictors.

• Relevance

• This model may help clinicians identify patients at high risk of internal iliac and obturator lymph node metastasis at initial diagnosis, to optimize treatment strategies, and to improve the precision of individualized therapy.

Accurate assessment of internal iliac and obturator lymph nodes is essential for optimal formulating treatment plans, including decisions on neoadjuvant/adjuvant therapies and the scope of surgical dissection.^9-11 Pelvic/rectal magnetic resonance imaging (MRI) is central to preoperative evaluation. Previous research works have demonstrated that measuring the short-axis diameter of LPLNs on pelvic/rectal MRI provides valuable diagnostic information.^12,13 Some investigations have defined separate cutoff diameters for internal iliac versus obturator nodes.^6,14,15 However, clinicians often face difficulty determining whether LPLNs on pelvic/rectal MRI are located in the internal iliac or obturator regions, especially when enlarged nodes span across these areas. Consequently, the applicability of a singular cutoff value in such scenarios is limited. Additionally, certain studies have indicated that the distance between the tumor and the anus also serves as an influential factor.¹⁶ Therefore, current studies that evaluate LPLNs metastasis in rectal cancer based solely on single factors are limited, underscoring the need for multidimensional assessment approaches. Developing a clinical data–driven, multidimensional, and interpretable prediction model is of substantial clinical importance.

Recent advancements in machine learning (ML) have opened new avenues for enhancing the accuracy of predicting internal iliac and obturator lymph node metastasis in rectal cancer.^17,18 ML models are capable of integrating and analyzing extensive clinical data to uncover complex, nonlinear patterns that may remain undetected by traditional methods.^19,20 Leveraging these technologies facilitates the development of more accurate and interpretable prediction models, thereby enhancing preoperative evaluation and optimizing treatment strategies.

This study hypothesizes that a clinical data–based ML model can predict internal iliac and obturator node metastasis, aiding preoperative decisions. We aim to develop and validate an accurate, interpretable ML model that provides reliable preoperative insights, supports personalized treatment planning, and enhances patient prognosis.

MATERIALS AND METHODS

Study Population

This retrospective cohort study aimed to develop and validate an ML model based on clinical data for predicting internal iliac and obturator lymph node metastasis in patients with rectal cancer. Data for this study were obtained from the Cancer Hospital of the Chinese Academy of Medical Sciences/Chinese Cancer Center, Peking University First Hospital, Tianjin People's Hospital, and Peking Union Medical College Hospital. Patients who underwent total mesorectal excision (TME) combined with lateral pelvic lymph node dissection (LPLND) between January 2015 and December 2020 were included. Inclusion criteria were (1) pathological diagnosis of rectal adenocarcinoma, (2) preoperative imaging indicating suspicious internal iliac or obturator lymph node metastasis, (3) receipt of TME with dissection of the internal iliac or obturator lymph nodes, and (4) no neoadjuvant therapy before surgery. Exclusion criteria were (1) lymph node specimens not submitted for separate examination, (2) postoperative pathological reports not documenting LPLNs, and (3) discrepancies between intraoperative lymph node location and the preoperative imaging report. A total of 381 patients met the criteria, including 136 who underwent bilateral LPLND and 245 who underwent unilateral LPLND. In total, 520 samples were used for model training, testing, and validation.

Ethics Approval and Consent to Participate

This study was approved by the Ethics Committee of National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College (NCC2022C-325). All methods were carried out in accordance with the Helsinki Declaration and approved guidelines. All enrolled patients signed informed consent.

Data Collection and Processing

Data were extracted from the hospitals' electronic medical record systems and pathological databases. Potential predictive variables for the diagnostic model were identified based on previous research and included patients' demographic characteristics, clinical characteristics, laboratory test results, and pathological findings.^21-24 Demographic characteristics included patient age, gender, and BMI. Clinical characteristics comprised (1) tumor clinical T stage (AJCC 8th edition); (2) regional lymph node status (peritoneal or mesenteric nodes); (3) tumor location (measured as the distance from the anal verge); (4) preoperative imaging findings, including MRI-based evaluation of primary tumor size, measurement of the short-axis diameter of enlarged lymph nodes, and determination of their location; and (5) the degree of histological differentiation of tumor tissue obtained via pathological biopsy (with low differentiation categorized as the low differentiation group and high differentiation as the high differentiation group). Laboratory test results encompassed preoperative serum levels of carcinoembryonic antigen (CEA) and carbohydrate antigen 19-9 (CA 19-9). Pathological results specifically documented internal iliac and obturator lymph node metastasis status. Missing values were imputed using the median.

Outcome Variables

The outcome variable of this study was the presence or absence of internal iliac and obturator lymph node metastasis, as determined by the postoperative pathological report. For each patient, the largest lymph node in the unilateral pelvic cavity was documented to ensure correspondence between the imaging report and pathological findings. Lymph node location was determined from imaging examinations and categorized as belonging to the internal iliac region, the obturator region, or the junctional region between these two areas.

ML Model Construction

The data set comprised cases from the Cancer Hospital of the Chinese Academy of Medical Sciences/Chinese Cancer Center (320 cases) and Tianjin People's Hospital (91 cases), which were randomly partitioned into a training set (75%) and a validation set (25%) for model development and internal validation. Additionally, patient data from Peking University First Hospital (76 cases) and Beijing Union Medical College Hospital (33 cases) were gathered to form an external test set, which was used to evaluate the model's generalization ability.

We constructed and compared five common ML models: logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), and k-nearest neighbor (KNN). All models were implemented using the Python Scikit-learn library (version 1.0.1). Hyperparameters for the models were optimized via grid search and cross-validation. Model performance was assessed by area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, and F1 score.

Model Interpretation Methods

To enhance model interpretability, we adopted the Shapley Additive explanations (SHAP) method. SHAP uses Shapley values to quantify each feature's contribution, enabling both global and individual-level explanations. The SHAP feature importance plot illustrates the global impact of each feature on the model's predictions, whereas the SHAP dependence plot delineates the specific influence of individual feature values. The SHAP method was implemented using SHAP version 0.40.0 in Python.

Statistical Analysis

Continuous variables are presented as mean ± standard deviation or median (IQR), whereas categorical variables are reported as frequencies and percentages. Group comparisons were conducted using independent-sample t-tests or chi-square tests. All statistical analyses were performed using Python (version 3.7.3) and R (version 4.3.3), with graphical representations generated using Prism (version 9.5.0). A two-sided P value <.05 was considered statistically significant.

RESULTS

Basic Characteristics

A total of 520 surgically resected internal iliac and obturator lymph nodes were included in this study, with 411 cases allocated to the training and testing data set (237 negative, 174 positive) and 109 cases to the external validation data set (66 negative, 43 positive). Table 1 summarizes the demographic and tumor characteristics of both groups. Tumor differentiation differed significantly between cohorts (P = .002), with the positive cohort showing 43.7% poorly differentiated tumors versus 55.3% moderately differentiated tumors in the negative cohort. In the external validation set, tumor differentiation did not differ significantly (P = .180). Regarding clinical staging and tumor markers, clinical T stage and regional lymph node status differed significantly between the groups (P < .001), with a higher proportion of patients exhibiting T4 stage and regional lymph node involvement in the positive group—suggesting an increased risk of tumor aggressiveness and lymph node metastasis. In the training and test data sets, tumor distance from the anus was significantly greater in the negative group than in the positive group (4.98 v 4.20, P = .003). In the external validation data set, although the negative group exhibited a greater tumor-to-anus distance (5.17 v 4.71), the difference was not statistically significant (P = .388). Tumor size did not differ significantly in either cohort (P = .116 and P = .151, respectively). Lymph node diameter was significantly larger in the positive group in both data sets (P < .001), indicating that enlarged lymph nodes are associated with an increased risk of metastasis. Finally, lymph node location differed significantly between the groups in both data sets (P < .001), with internal iliac lymph node metastasis being more prevalent in the positive group (50.6% and 58.1% in the respective data sets).

TABLE 1.

Patient and Tumor Characteristics

Characteristic	Training and Test Sets (n = 411)			External Validation Set (n = 109)
	Negative (n = 237)	Positive (n = 174)	P	Negative (n = 66)	Positive (n = 43)	P
Sex, No. (%)			.080			.421
Male	131 (55.3)	112 (64.4)		43 (65.2)	24 (55.8)
Female	106 (44.7)	62 (35.6)		23 (34.8)	19 (44.2)
Age, years, median (IQR)	63.00 (53.00-67.00)	58.50 (48.00-66.00)	.058	57.00 (48.25-63.00)	58.00 (47.00-67.00)	.737
BMI, median (IQR)	24.59 (22.73-26.45)	23.88 (21.48-26.08)	.080	24.77 (22.37-26.77)	23.88 (21.76-25.30)	.099
Cancer differentiation, No. (%)			.002			.180
Well differentiated	5 (2.1)	10 (5.7)		9 (13.6)	3 (7.0)
Moderately differentiated	131 (55.3)	75 (43.1)		36 (54.5)	18 (41.9)
Badly differentiated	97 (40.9)	76 (43.7)		18 (27.3)	17 (39.5)
Unclear	4 (1.7)	13 (7.5)		3 (4.5)	5 (11.6)
Clinical T stage, No. (%)			<.001			.006
T1	4 (1.7)	3 (1.7)		0 (0.0)	0 (0.0)
T2	32 (13.5)	17 (9.8)		18 (27.3)	2 (4.7)
T3	160 (67.5)	101 (58.0)		36 (54.5)	26 (60.5)
T4	37 (15.6)	13 (7.5)		12 (18.2)	15 (34.9)
Unclear	4 (1.7)	40 (23.0)		0 (0.0)	0 (0.0)
Regional lymph nodes, No. (%)			<.001			<.001
Negative	59 (24.9)	7 (4.0)		21 (31.8)	1 (2.3)
Positive	174 (73.4)	128 (73.6)		45 (68.2)	42 (97.7)
Unclear	4 (1.7)	39 (22.4)		0 (0.0)	0 (0.0)
CEA, mean (SD)	7.69 (12.94)	21.08 (121.49)	.099	28.20 (101.08)	7.77 (5.83)	.205
CA 19-9, mean (SD)	22.09 (53.79)	28.97 (88.68)	.347	17.36 (26.18)	64.11 (259.68)	.162
Distance from the anus, mean (SD)	4.98 (2.70)	4.20 (2.58)	.003	5.17 (2.98)	4.71 (2.09)	.388
Tumor size, mean (SD)	4.63 (1.57)	4.39 (1.46)	.116	4.63 (1.70)	4.09 (2.18)	.151
Lymph nodes diameter, mean (SD)	0.59 (0.25)	1.05 (0.51)	<.001	0.70 (0.29)	1.23 (0.58)	<.001
Lymph nodes location, No. (%)			<.001			<.001
Internal iliac	116 (48.9)	88 (50.6)		22 (33.3)	25 (58.1)
Obturator	97 (40.9)	46 (26.4)		13 (19.7)	14 (32.6)
Junction of internal iliac and obturator	24 (10.1)	40 (23.0)		31 (47.0)	4 (9.3)

NOTE. Grouping according to internal iliac/obturator lymph node metastasis.

Abbreviations: CA 19-9, carbohydrate antigen 19-9; CEA, carcinoembryonic antigen.

Modeling

Variable selection was performed using Lasso regression and Boruta algorithm; Lasso regression excluded CA 19-9 and age (Fig 1A), whereas Boruta excluded CA 19-9, BMI, and gender (Fig 1B). Variables selected by both methods were retained for modeling, including the short-axis diameter of anatomical location of enlarged lymph nodes, clinical T stage and regional lymph node status, tumor diameter, tumor differentiation, tumor-to-anal distance, and preoperative CEA level.

FIG 1.

(A) Lasso feature selection plot, where the x-axis represents the regression coefficients of the features and the y-axis represents the names of the features. In this study, the threshold for the regression coefficient is set at 0.01; variables with regression coefficients below 0.01 are rejected, and those above 0.01 are included. (B) Boruta variable importance plot, with variables ordered by their importance scores along the x-axis from lowest to highest and the y-axis representing the importance scores. Each variable is colored green if accepted or red if rejected. CA 19-9, carbohydrate antigen 19-9.

Model Evaluation

Among the ML models used to predict the metastasis of internal iliac and obturator lymph nodes in patients with rectal cancer, KNN achieved the highest AUC (1.00) on the training set, followed by RF (0.999) and XGBoost (0.991), whereas LR and SVM performed moderately (0.898 and 0.896, respectively; Fig 2). KNN also achieved the highest accuracy (0.997) and F1 score (0.996), followed by RF (0.995 and 0.993, respectively; Table 2). In the test set, AUC values were 0.877 for both the RF model and XGBoost, 0.827 for both KNN and LR, and 0.822 for SVM, indicating that SVM performed the worst (Fig 2B). Regarding accuracy, RF and XGBoost achieved the highest scores (0.788), although RF was slightly inferior to XGBoost in F1 (0.766 v 0.761), whereas KNN recorded the lowest scores across all three metrics (0.740, 0.682, and 0.465; Table 2). In the external validation data set, the RF model demonstrated the highest AUC (0.899) among the five models, followed by XGBoost (0.887), whereas KNN, LR, and SVM achieved AUCs of 0.851, 0.846, and 0.845, respectively (Fig 2C). Moreover, in terms of accuracy and F1 score, the RF model outperformed the others (0.827, 0.786), whereas XGBoost performed worse than both LR and SVM, and KNN recorded the lowest scores (Table 2). Thus, the RF model demonstrated superior overall performance and was selected for subsequent analyses.

FIG 2.

Receiver operating characteristic curves showing the muscle loss predictive performance of the machine learning models in (A and B) the training and test data sets and (C) external validation data set. LR, logistic regression; RF, random forest; XGB, extreme gradient boosting.

TABLE 2.

Evaluation of Model Performance in the Training Test and Validation Set

Train	Precision	F1 Score	Sensitivity	Specificity
RF	0.995 (0.992-0.998)	0.993 (0.990-0.996)	0.990 (0.982-0.998)	0.995 (0.989-1.000)
LR	0.827 (0.801-0.853)	0.771 (0.726-0.817)	0.700 (0.636-0.764)	0.918 (0.917-0.918)
XGB	0.992 (0.987-0.997)	0.985 (0.979-0.991)	0.986 (0.976-0.995)	0.995 (0.990-1.000)
KNN	0.997 (0.994-1.000)	0.996 (0.992-1.000)	0.992 (0.984-1.000)	0.998 (0.995-1.000)
SVM	0.824 (0.821-0.827)	0.758 (0.756-0.759)	0.662 (0.659-0.664)	0.929 (0.912-0.945)

Test	Precision	F1 Score	Sensitivity	Specificity
RF	0.788 (0.760-0.817)	0.761 (0.740-0.782)	0.761 (0.731-0.791)	0.810 (0.785-0.836)
LR	0.750 (0.721-0.779)	0.698 (0.679-0.716)	0.652 (0.630-0.674)	0.828 (0.803-0.852)
XGB	0.788 (0.779-0.798)	0.766 (0.764-0.768)	0.783 (0.775-0.791)	0.793 (0.783-0.803)
KNN	0.740 (0.673-0.808)	0.682 (0.628-0.737)	0.630 (0.610-0.651)	0.828 (0.737-0.918)
SVM	0.760 (0.712-0.808)	0.699 (0.654-0.744)	0.630 (0.586-0.674)	0.828 (0.819-0.836)

Validation	Precision	F1 Score	Sensitivity	Specificity
RF	0.827 (0.760-0.894)	0.786 (0.731-0.841)	0.805 (0.704-0.906)	0.841 (0.794-0.889)
LR	0.817 (0.798-0.837)	0.782 (0.778-0.785)	0.829 (0.690-0.969)	0.810 (0.778-0.841)
XGB	0.808 (0.801-0.815)	0.767 (0.737-0.798)	0.805 (0.735-0.875)	0.810 (0.778-0.841)
KNN	0.779 (0.769-0.788)	0.729 (0.718-0.741)	0.756 (0.637-0.875)	0.794 (0.750-0.837)
SVM	0.817 (0.740-0.894)	0.771 (0.702-0.841)	0.780 (0.655-0.906)	0.794 (0.792-0.796)

Abbreviations: KNN, k-nearest neighbor; LR, logistic regression; RF, random forest; SVM, support vector machine; XGB, extreme gradient boosting.

ML Model Interpretation at the Macro Level

Figure 3A illustrates the importance of the SHAP features for the RF model. On the basis of the mean absolute SHAP values, the most important clinical features, in descending order, were lymph node short-axis diameter, regional lymph node status, tumor-to-anal distance, clinical T stage, and tumor diameter. Figure 3B displays the SHAP summary plot for the RF model. According to the prediction model, a higher SHAP value for a feature corresponds to a greater likelihood of lymph node positivity. For example, larger short-axis diameters correspond to higher SHAP values and greater metastatic probability. The SHAP dependency plot illustrates the effect of individual features on the RF model's predictions: larger short-axis diameters and higher clinical T stage or regional nodal involvement are associated with increased SHAP values, whereas shorter tumor-to-anal distances correspond to higher SHAP values and greater metastatic risk.

FIG 3.

(A) SHAP feature importance shown according to the mean absolute SHAP value of each feature. (B) SHAP summary plot showing the distribution of the SHAP values of each feature. Each dot represents a SHAP value for a feature per patient. The x-axis represents the SHAP value, and the color varying from red to blue represents the feature value from high to low, respectively. SHAP, Shapley Additive explanations.

ML Model Interpretation at the Individual Level

We used SHAP to interpret RF predictions for two patients: patient A (true positive; Fig 4A) with a predicted probability of 91% for lymph node metastasis, and patient B (true negative; Fig 4B) with a predicted probability of 29%.

FIG 4.

SHAP force plot of (A) patient A (true positive) and (B) patient B (true negative). The color represents the contributions of each feature, with red being positive and blue being negative. The length of the color bar represents the contribution strength. CEA, carcinoembryonic antigen; SHAP, Shapley Additive explanations.

Patient A, a 51-year-old woman, had stage III moderately low-differentiated carcinoma. Her preoperative serum CEA was 7.00 ng/mL and CA 19-9 was 6.6 U/mL. MRI showed a 4.5-cm circumferential tumor with subserosal invasion (T4), located 3 cm from the anus; suspicious periampullary and left internal iliac nodes had a maximum short-axis diameter of 0.9 cm. Laparoscopic anterior resection with left LPLND yielded two positive nodes out of two examined. SHAP analysis attributed the high predicted probability to poor differentiation, regional lymph node positivity, advanced T stage, elevated CEA, and larger lymph node diameter.

Patient B, a 64-year-old man, had stage II moderately differentiated rectal cancer. His preoperative serum CEA was 2.76 ng/mL and CA 19-9 was 12.0 U/mL. MRI showed a 4.2-cm circumferential mass invading the submucosal layer (T3), located 8 cm from the anus. No mesenteric nodes were enlarged; left obturator lymph nodes had a maximum short-axis diameter of 0.7 cm. Laparoscopic anterior resection with left lymphadenectomy yielded zero positive nodes out of three examined. SHAP analysis attributed the low predicted probability to absent lymph node involvement, low CEA, moderate differentiation, greater tumor-to-anal distance, smaller lymph node diameter, and lower T stage.

DISCUSSION

To our knowledge, this study is the first to use an interpretable ML model to predict the metastasis of internal iliac and obturator lymph nodes in patients with rectal cancer. By constructing a prediction model based on clinical data, this study provides a more accurate foundation for preoperative treatment decisions. Model variables were initially screened using Lasso regression and the Boruta algorithm. Variables retained by both methods included short-axis diameter and anatomical location of LPLNs, preoperative CEA level, clinical T stage, regional lymph node status, tumor histological differentiation, tumor distance from the anus, and tumor diameter. Five common ML methods were evaluated (LR, RF, XGBoost, SVM, and KNN), and the RF model demonstrated superior overall performance during both model development and validation. External validation was conducted using data from two additional hospitals to ensure the credibility and generalizability of the findings, with the RF model again outperforming the alternatives. Furthermore, SHAP elucidated each feature's contribution, yielding a model with both high predictive accuracy and strong interpretability. SHAP analysis identified the key predictive features—in descending order of importance—as the short-axis diameter of the enlarged lymph node, regional lymph node status, preoperative CEA level, and tumor distance from the anus.

Previous studies have demonstrated that preoperative pelvic/rectal MRI combined with measurement of the short-axis diameter of enlarged LPLNs serves as a simple and rapid predictor of lymph node metastasis.^12,13 Several studies have specifically delineated the short-axis diameter cutoff values for internal iliac and obturator lymph nodes, both before and after neoadjuvant therapy.^6,14,15 In our study, LPLN short-axis diameter was the top predictor in Lasso, Boruta, and SHAP analyses. However, this parameter is not without limitations, such as the potential for false positives due to inflammation²⁵ and false negatives resulting from small lesions.^26,27 For example, patient B's obturator node measured 0.7 cm—above cutoff yet benign. In this case, a combination of negative nodal status, low CEA, moderate differentiation, distal tumor location, smaller lymph node size, and low T stage led SHAP to predict no metastasis, which pathology confirmed. Thus, diameter alone can misclassify patients; models must integrate multiple factors.

Regional lymph nodes (ie, peri- and endo-mesenteric lymph nodes) were assigned high importance in predicting metastasis in internal iliac and obturator lymph nodes, underscoring their critical role in assessing the risk of involvement of LPLNs. Although peri-intestinal and intraintestinal lymph nodes primarily metastasize via the mesenteric vascular route,^28,29 LPLN metastasis is predominantly linked to the internal iliac vascular system, as tumor cells disseminate along the internal iliac arteries and accompany lymphatic channels to reach the internal iliac and obturator nodes.^30,31 However, current statistics indicate that only about 10%-25% of patients exhibit LPLN metastasis in the absence of peri-intestinal or intraintestinal lymph node involvement,^32,33 indicating a strong correlation between these sites and potentially explaining the high predictive weight assigned to regional lymph nodes in the model. Furthermore, elevated preoperative CEA levels were strongly associated with an increased risk of tumor invasiveness and metastasis in rectal cancer, consistent with findings from several studies.

Some investigators have proposed tumor-to-anal verge distance as an indicator of LPLN metastasis risk, with lower located tumors showing higher propensity.¹⁶ In our study, although tumor distance from the anus was incorporated into the model, SHAP analysis revealed that it carried a lower weight compared with regional lymph node status and preoperative CEA levels. Although low rectal cancers are more predisposed to LPLN metastasis, intermediate and even high rectal cancers may also metastasize via lateral drainage pathways,³¹ rendering tumor distance from the anus an insufficient sole predictor. This likely accounts for its lower predictive weight in assessing metastasis in patients with internal iliac and obturator lymph nodes.

In this study, tumor diameter was also identified as one of the factors associated with LPLN metastasis, although its weight was relatively low. Contrary to expectation, larger tumors were linked to lower LPLN metastasis risk. Previous studies have shown that, in colon cancer, patients with tumor diameters >45 mm had a 53% increased risk of lymph node metastasis. However, in rectal and sigmoid colon cancers, tumor size does not appear to have a significant impact on lymph node involvement.³⁴ The pathways of mesorectal versus pelvic lymphatic spread may differ, and further research is required to elucidate the relationship between primary tumor size and LPLN metastasis.

Despite the promising results, this study has several limitations. First, additional factors that could influence lymph node metastasis—such as gene expression profiles of primary tumors and characteristics of the tumor microenvironment—were not incorporated into the model.³⁵ With advances in precision medicine, integrating multidimensional biomarkers is anticipated to further enhance the model's predictive performance. Additionally, as a retrospective study, certain potentially predictive factors—such as MRI assessments of the mesorectal fascia and extramural vascular invasion, lymph node margin clarity, and signal homogeneity—were subject to missing data and therefore could not be included in the model. Future research should explore the integration of more comprehensive imaging data to develop more accurate prediction models that better inform treatment and surgical decisions.³⁶ Third, the model is not applicable to patients who have undergone neoadjuvant therapy, limiting its utility in guiding decisions regarding LPLND.

The ML-based prediction model for internal iliac and obturator lymph node metastasis in rectal cancer developed in this study demonstrates robust predictive ability and interpretability, offering personalized support for clinical decision making. Moreover, the predictive factors are readily obtainable in clinical practice, enabling consideration of more aggressive preoperative neoadjuvant therapy or extensive surgical clearance for high-risk patients, whereas unnecessary extended surgeries—and consequently postoperative complications—can be avoided for low-risk patients. Furthermore, the model's interpretability facilitates improved clinician-patient communication by elucidating the prediction results and the rationale behind treatment choices, thereby enhancing patient compliance. Future work should optimize the model and validate it in diverse clinical settings to refine individualized treatment.

DATA SHARING STATEMENT

The data involved in the article can be shared and contacted by corresponding authors Qian Liu. The code for training, testing and validating the model is detailed in https://github.com/Shawn2025code/LPLN.git

AUTHOR CONTRIBUTIONS

Conception and design: Tixian Xiao, Wei Zhao, Fangze Wei, Fuqiang Zhao, Xin Wang, Qian Liu

Financial support: Qian Liu

Administrative support: Qian Liu

Provision of study materials or patients: Fei Huang, Junge Bai, Xin Wang, Qian Liu

Collection and assembly of data: Tixian Xiao, Wei Zhao, Zhen Sun, Fuqiang Zhao, Fei Huang, Zeyu Wu, Junge Bai, Qian Liu

Data analysis and interpretation: Tixian Xiao, Wei Zhao, Fangze Wei, Fuqiang Zhao, Junge Bai, Qian Liu

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/po/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).

No potential conflicts of interest were reported.