Abstract 147P
Background
Machine learning with radiomics showed great potential to predict response to neoadjuvant systemic treatment (NAST) for breast cancer. However, performance across different ethnicities and potential racial bias remain unclear. We aimed to develop an intelligent algorithm using pretreatment MRI radiomics in addition to clinical variables, to validate their performance in ethnically diverse populations.
Methods
We used institutional data of patients who underwent MRI before NAST. We developed a support vector machine algorithm based on German patients using pretreatment MRI radiomics features in addition to patient and tumor variables to predict pCR status (ypT0 and ypN0). We used N4 bias field correction to maintain the consistency of images acquired by different types of machines and settings. Model performance was validated on an American and Chinese dataset. Findings were compared to the histopathologic evaluation of the surgical specimen. The main outcome measure was the area under the curve (AUC).
Results
We included 656 patients in the German development set, 88.6% (581 of 656) were white people and 34.8% (228 of 656) achieved pCR. The model showed good performance in the development set (AUC: 0.81, 95%CI, 0.71-0.83). Validation in the American population sample (n = 100, 80% white people) showed non-inferior performance compared to the development (AUC: 0.75 vs. 0.81, p = 0.543), but performance in the Chinese population sample (n = 100, 0 white people) decreased (AUC: 0.61 vs. 0.81, p = 0.004). Also within the development set, training performance was descriptively lower for the patients with Asian (n= 59, AUC 0.73 (95% CI 0.56-0.88)), or African (n=16, AUC 0.71 (95% CI 0.44-0.95)) ethnicity.
Conclusions
Racial bias exists in radiomics studies and should be assessed before the global application of AI-based imaging algorithms. Ethnic diversity is crucial to mitigate racial bias when developing such algorithms.
Clinical trial identification
This multi-center and retrospective study was approved by the Ethics Committee of Heidelberg University Medical Faculty (S-092/2022) and Affiliated Hospital of Guizhou Medical University (2023-663).
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.