Time-dependent prediction model using modified DeepSurv algorithm for dynamic risk assessment of post-operative bone metastases in breast cancer.

Abstract

e13109 Background: Breast cancer (BC) is the most common cancer in women, with bones being the primary site of metastases (~70%). Bone metastases often lead to skeletal-related events (SREs), adversely impacting patients' quality of life and survival. For post-operative BC, timely risk identification, screening, and intervention for bone metastases will reduce SREs and improve survival. However, formulating personized real-time screening recommendations poses a significant challenge. Our study aimed to develop a model to dynamically predict the risk of bone metastases in post-operative BC with the goal of providing an individual screening strategy. Methods: We retrospectively analyzed BC patients aged 20 to 70 years who were firstly diagnosed between 2010 and 2023 and received surgery from 9 medical centers in China (NCT06544668). Muti-center real-world data were collected and divided into 3 groups: Group A (recurrence with bone metastases), Group B (recurrence without bone metastases), and Group C (no recurrence). Univariable, multivariable and correlation analyses were used to identify risk factors related to bone metastases. Then Cox regression, machine learning (Random Forest, Support Vector Machine), and deep learning (DeepSurv) models and the modified version incorporating longitudinal disease trajectories data were constructed respectively. Tenfold cross-validation was used for hyperparameter tuning. A hold-out test was used to evaluate the performance of the models with concordance index (c-index) as an evaluation metric. Results: A total of 3,970 BC patients were enrolled (Group A, 2,078 [52.3%]; Group B, 817 [20.6%]; Group C, 1,075 [27.1%]). The median time of first recurrence from surgery was 44.1 months. In Group A, the median level of ALP gradually increased within the year prior to bone metastases, whereas no increase was observed in the other 2 groups. Univariate analysis showed that bone metastases were significantly associated with post-surgery baseline characteristics (tumor stage, node stage, pathological grade, nerve invasion, neoadjuvant therapy, and Ki-67) and dynamitic characteristics (ALP level and lung metastases). Among the base prediction models, the DeepSurv model demonstrated the most favorable performance for distinguishing the risk of bone metastases in the hold-out test dataset with c-index of 0.67, and 3, 5 years AUC of 0.66 and 0.69. After incorporating longitudinal data, the c-index of the modified DeepSurv model in test dataset was increased to 0.80, and the AUC values at 3, 5 years to 0.81 and 0.79. Conclusions: This study identified new dynamic risk factors for bone metastases in post-operative BC. By incorporating disease trajectory data, the modified DeepSurv model could dynamically predict bone metastases to enable personalized screening. Validation using external test databases is currently underway.