Distant metastasis prediction via a multi-feature fusion model in breast cancer

Abstract

This study aimed to develop a model that fused multiple features (multi-feature fusion model) for predicting metachronous distant metastasis (DM) in breast cancer (BC) based on clinicopathological characteristics and magnetic resonance imaging (MRI). A nomogram based on clinicopathological features (clinicopathological-feature model) and a nomogram based on the multi-feature fusion model were constructed based on BC patients with DM (n=67) and matched patients (n=134) without DM. DM was diagnosed on average (17.31±13.12) months after diagnosis. The clinicopathological-feature model included seven features: reproductive history, lymph node metastasis, estrogen receptor status, progesterone receptor status, CA153, CEA, and endocrine therapy. The multi-feature fusion model included the same features and an additional three MRI features (multiple masses, fat-saturated T2WI signal, and mass size). The multi-feature fusion model was relatively better at predicting DM. The sensitivity, specificity, diagnostic accuracy and AUC of the multi-feature fusion model were 0.746 (95% CI: 0.623-0.841), 0.806 (0.727-0.867), 0.786 (0.723-0.841), and 0.854 (0.798-0.911), respectively. Both internal and external validations suggested good generalizability of the multi-feature fusion model to the clinic. The incorporation of MRI factors significantly improved the specificity and sensitivity of the nomogram. The constructed multi-feature fusion nomogram may guide DM screening and the implementation of prophylactic treatment for BC.

Keywords: artificial intelligence; breast neoplasms; early detection; neoplasm metastasis.

Figure 1

Venn diagrams showing intersections between different metastasis types used in our study. There were 26 cases of multiple organ metastases and 41 cases of single organ metastasis. Others include peritoneal (or mediastinal, ovarian, soft tissue) metastasis, pericardial effusion and lemostenosis.

Figure 2

Construction of the clinicopathological-feature alone model. (A) Selection of tuning parameter lambda in the LASSO model used 10-fold cross-validation. The gray line in the figure is the partial likelihood estimate corresponding to the optimal value of lambda. The optimal lambda value of 2.313 was chosen. (B) LASSO coefficient profiles of the eleven selected features. A vertical line was plotted at the optimal lambda value, which resulted in seven features with nonzero coefficients. (C) A nomogram was developed in the training data set with clinicopathological characteristics. Calibration curves and ROC curves of the nomogram for the training set (D, G), validation set (E, H) and total population (F, I).

Figure 3

Construction of the multi-feature fusion model. (A) Selection of tuning parameter lambda in the LASSO model used 10-fold cross-validation. The gray line in the figure is the partial likelihood estimate corresponding to the optimal value of lambda. The optimal lambda value of 2.653 was chosen. (B) LASSO coefficient profiles of the sixteen selected features. A vertical line was plotted at the optimal lambda value, which resulted in ten features with nonzero coefficients. (C) A nomogram was developed in the training data set with clinicopathological and MRI features. Calibration curves and ROC curves of the nomogram for the training set (D, G), validation set (E, H) and total population (F, I).

Figure 4

Receiver operating characteristic (ROC) curves of the nomograms. (A) ROC curves of the clinicopathological-feature alone model and multi-feature fusion model for the total population. (B) ROC curves of the multi-feature fusion model in the training set and calibration set. (C) ROC curve of the multi-feature fusion model in the external validation cohort. (D) Calibration curves of the multi-feature fusion model in the external validation cohort.

Figure 5

Flowchart of the patient selection process in the present study.