Development of a Machine Learning Model Integrating Pathomics and Clinical Data to Predict Axillary Lymph Node Metastasis in Breast Cancer: A Two-Center Study

Abstract

Background: Accurately assessing the status of axillary lymph nodes (ALNs) is essential for devising optimal surgical plans and making informed treatment decisions in breast cancer (BC) patients.

Aims: This study aims to develop an innovative nomogram based on pathomics to preoperatively predict ALN metastasis (ALNM) in BC.

Methods and results: Our study performed a retrospective analysis on digital hematoxylin and eosin (H&E)-stained images obtained from 407 patients across two institutions who were allocated into a training cohort (TC; n = 203), an internal validation cohort (IVC; n = 136), and an external validation cohort (EVC; n = 68). Initially, the Mann-Whitney U-test and Spearman's rank correlation coefficient were utilized for feature selection, employing the least absolute shrinkage and selection operator (LASSO) regression for further refinement. For the evaluation of the predictive value of ALNM and other clinicopathological factors, we deployed both univariate (ULR) and multivariate (MLR) logistic regression analyses. Among the six machine learning (ML) algorithms, logistic regression, which demonstrated the highest area under the curve (AUC) value, was employed to establish the final nomogram model. The nomogram reliability and stability were assessed by analyzing the AUC of the receiver operating characteristic (ROC) curve, decision curve analysis (DCA), and calibration plots. MLR analysis demonstrated estrogen receptor (ER), human epidermal growth factor receptor 2 (HER2), tumor size, and pathomics features as independent ALNM predictors. The nomogram demonstrated that the AUC in the IVC (0.783) surpassed that of the Path-score model (0.698) (DeLong test, p = 0.008558). Similarly, in the EVC, the nomogram surpassed the clinical model regarding AUC (0.738 vs. 0.574; DeLong test, p = 0.00494). Additionally, DCA analysis indicated a net clinical benefit associated with the nomogram.

Conclusion: Our study demonstrates the effectiveness of pathomics features in predicting ALNM in BC patients. Furthermore, the pathomics-based nomogram offers a valuable tool for personalized treatment planning in this patient population.

Keywords: axillary lymph node metastasis; breast cancer; machine learning; nomogram model; pathomics.

FIGURE 1

Flowchart of the overall construction process of the machine learning models predicting ALNM status. ALNM, axillary lymph node metastasis; IVC, internal validation cohort; EVC, external validation cohort; TC, training cohort.

FIGURE 2

Schematic illustration of the constructed pathomics signature. (A) Representative H&E tile selection. Scale bar: 500, 500, and 200 μm, respectively. (B) The usage of the chosen tile to extract the pathomics feature. Selection of the pathomics feature through (C) LASSO binary regression with (D) 10‐fold cross‐validation. (E) The pathomics feature coefficients are used to construct the Path‐score. (F) Calculation of the pathomics signature relying upon the chosen features. H&E, hematoxylin and eosin; LASSO, least absolute shrinkage and selection operator.

FIGURE 3

The distribution of pathological scores among all patients and across cohorts. (A) Pathomics score for each patient suffering from BC in the total cohort; distribution of pathomics score values of the ALN+/ALN− groups in the TC (B) and IVC (C). ALN, axillary lymph node; BC, breast cancer; IVC, internal validation cohort; TC, training cohort.

FIGURE 4

ROC curves of the six machine learning classifiers that predict the ALNM in the TC (A) and IVC (B). ALNM, axillary lymph node metastasis; IVC, internal validation cohort; ROC, receiver operating characteristic; TC, training cohort.

FIGURE 5

The developed nomogram based on the combined ER, HER2, tumor size, and Path‐score through logistic regression analysis. ER, estrogen receptor; HER2, human epidermal growth factor receptor 2.

FIGURE 6

Validation of the Path‐score‐based nomogram predictive value. The nomogram model's ROC curves in the (A) TC, (B) IVC, and (C) EVC. The nomogram model calibration plots in the (D) TC, (E) IVC, and (F) EVC. The net benefit of nomogram usage is shown by the decision curves in the (G) TC, (H) IVC, and (I) EVC. EVC, external validation cohort; IVC, internal validation cohort; ROC, receiver operating characteristic; TC, training cohort.

FIGURE 7

The ALNET: A simple program interface that predicts ALNM based on tumor size, ER status, HER2 status, and Path‐score. ALNET, Axillary Lymph Node Estimation Tool; ALNM, axillary lymph node metastasis; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2.