JMIR Publications

ABSTRACT

Background:

Breast cancer is the most prevalent form of cancer worldwide, with 2.3 million new diagnoses in 2022. Recent advancements in treatment have led to a shift in the utilization of chemotherapy-targeted immunotherapy from a postoperative adjuvant to a preoperative neoadjuvant approach in select cases, resulting in enhanced survival outcomes. A pathological complete response (pCR) is a critical prognostic marker, with higher pCR rates linked to improved overall and disease-free survival.

Objective:

The objective of this study was to develop robust, machine learning-based prediction models for pCR following neoadjuvant therapy, leveraging clinical, laboratory, and imaging data.

Methods:

A retrospective cohort study was conducted using data from the Taipei Medical University Clinical Research Database (TMU-CRD) from 2015 to 2022. The inclusion criteria were as follows: patients with breast cancer who received neoadjuvant therapy followed by curative surgical resection. Machine learning models were developed using three distinct sets of variables. The model 1 comprised 14 clinical variables, including patient age, tumor size, receptor status, and breast cancer intrinsic subtype. The model 2 expanded on the clinical variables by incorporating additional laboratory data and co-morbidities, resulting in 29 variables. The third model incorporated breast sonography response data in conjunction with the clinical variables from the model 1. Machine learning algorithms, including Logistic Regression (LR), Random Forest (RF), Support Vector Machines (SVM), and XGBoost, were employed for model training. The feature selection process incorporated a machine learning method known as Recursive Feature Elimination with Cross-Validation (RFECV), and the performance of the resulting model was evaluated using accuracy and the area under the receiver operating characteristic curve (AUROC).

Results:

A total of 334 patients were analyzed, with 199 in the non-pCR group and 135 in the pCR group. The application of logistic regression with RFECV was found to demonstrate the optimal performance among the various algorithms that were evaluated in this study. The model 1 attained an accuracy of 0.66 ± 0.02 and an AUROC of 0.73 ± 0.01. The incorporation of laboratory data and co-morbidities in the model 2 did not yield significant enhancement, with an accuracy of 0.67 ± 0.02 and an AUROC of 0.73 ± 0.01. The incorporation of breast sonography response in the model 3 led to a modest enhancement in predictive performance for the sonography group (accuracy: 0.68, AUROC: 0.60) in comparison to the non-sonography group (accuracy: 0.66, AUROC: 0.55). Despite the modest sample size (41 patients) of the model 3, the integration of sonography data appeared to offer additional value in predicting pCR and warrants further investigation.

Conclusions:

This study suggests that incorporating breast sonography into models with clinical and laboratory data may modestly improve pCR prediction. It is important to note that the findings of this study are preliminary and require cautious interpretation. Further studies are required to validate this approach and support its integration into a machine learning-based clinical workflow.