Author: Lucy Jiang, Chengyu Shi 👨🔬
Affiliation: Department of Radiation Oncology, City of Hope Orange County, Amity Regional High School (10th Grade) 🌍
Purpose: Early-stage breast cancer is common among females, with typically high local tumor control rates (LCR). Brachytherapy is a common way to achieve LCR following surgery. However, the patients may raise concerns about post-treatment toxicity. For forecasting treatment outcomes, we explored deep learning algorithms to develop an Artificial Intelligence (AI)--driven predictor by utilizing clinical and tumor-specific data after brachytherapy.
Methods: A Clinical trial data set (n=120) of early-stage breast cancer after brachytherapy was utilized for model training and evaluation in the MATLAB platform. Input patient data included 50 features regarding tumor size, anatomical locations, hormonal therapy, radiotherapy, etc. The Common Terminology Criteria for Adverse Events (CTC) of treatment outcome was selected as the predicted response. Different training algorithms (such as fine tree and neural network) in classification learner application of Matlab were compared to find the highest prediction accuracy algorithm. Feature importance analysis using the ANOVA statistic method was applied to determine the 5 most impactful inputs on CTC grade. The model was re-trained afterward with the identified 5 features. In addition, the MATLAB App Design interface was used to design an independent app interface using the final trained model.
Results: Preliminary results demonstrated that the fine tree classification algorithm yielded the best predictor accuracy of 94.5%. Feature importance analysis highlights the 5 most impactful factors on CTC Grade ranked in increasing order were progesterone receptor, adjuvant therapy, tumor location 1&2, and tumor size. An App with a simple interface was developed and successfully run to predict the CTC grade.
Conclusion: This research demonstrated the feasibility and accuracy of using deep learning to predict treatment outcomes in early-stage breast cancer patients based on available. An AI-driven predictor (APP interface), shows potential for public use. This study can be extended to other clinical data, broadening its clinical significance.