2185 - Development and External Validation of Predictive Models for Atrial Fibrillation Following Radiotherapy in Non-Small Cell Lung Cancer Patients

S. K. Yoo¹, K. H. Kim¹, J. M. Noh², J. Oh³, G. Yang¹, J. Kim¹, N. Kim², H. Kim¹, and H. I. Yoon¹; ¹Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), ²Department of Radiation Oncology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea, Republic of (South), ³Cardiology Division, Severance Cardiovasular Hospital and Cardiovascular Research Institute, Yonsei University College of Medicine, Seoul, Korea, Republic of (South)

Purpose/Objective(s): Advances in radiotherapy (RT) for lung cancer treatment have enhanced survival rates but also heightened the concern for cardiotoxicity following RT. While previous studies predominantly focused on ischemic heart diseases in RT, our prior study investigated the risk of atrial fibrillation (AF) using clinical, dosimetry, and coronary artery calcification (CAC) data. The current study aims to develop and validate learning-based predictive models, including machine and deep learning techniques, for predicting AF following chemoradiotherapy (CRT) in non-small cell lung cancer (NSCLC) patients. Materials/

Methods: We analyzed 321 and 187 datasets from internal and external NSCLC patient cohorts, noting 17 and 6 cases of AF incidence respectively. The analysis incorporated 159 patient-specific features, including clinical data, dosimetry (dose-volume histograms), and CAC data. We addressed the imbalance between AF and non-AF cases using the Synthetic Minority Oversampling Technique (SMOTE). Our machine learning model used a feature intervention technique for optimal feature representation, selecting the feature with the highest coefficient in each cardiac sub-structure after the feature selection process. On the other hand, the deep learning model was specifically designed with hybrid architectures, enabling it to handle diverse input data types by pairing them with corresponding network architectures. The effectiveness of these models was evaluated using the area under the curve (AUC), with important feature identification using coefficients in classifiers and an open-source library, Captum, for machine and deep learning models, respectively.
Results: In this study, the hybrid deep learning model demonstrated superior performance in the internal cohort, achieving an AUC of 0.823, which surpassed the machine learning model with intervention, which had an AUC of 0.801. In external validation, this deep learning model also showed more consistent results, with an AUC of 0.806, compared to the machine learning models AUC of 0.776. Notably, in both internal and external validations, the identified key features were the maximum dose to the heart and the sinoatrial node (SAN).
Conclusion: In this study, we developed and validated machine and deep learning models for predicting AF in NSCLC patients following CRT, finding the deep learning model to be more accurate and identifying the maximum heart dose and SAN as important feature. Abstract 2185 – Table 1: Performance of predictive models

Validation	Model		Accuracy	Sensitivity	Specificity	AUC
Internal Validation	Machine learning	Without intervention	76.67 (±7.26)	76.67 (±15.26)	76.67 (±8.13)	0.767 (±0.072)
		With intervention	80.01 (±8.51)	81.67 (±13.83)	78.33 (±10.65)	0.801 (±0.086)
	Deep learning		82.48 (±4.49)	83.32 (±10.53)	81.65 (±8.96)	0.823 (±0.045)
External Validation	Machine learning	Without intervention	66.27	66.67	66.26	0.665
		With intervention	72.19	83.33	71.82	0.776
	Deep learning		78.08	83.33	77.90	0.806