Predictive Modeling of Long-Term Survivors with Stage IV Breast Cancer Using the SEER-Medicare Dataset

Area of Science:

• Oncology
• Biostatistics
• Artificial Intelligence

Background:

• Stage IV breast cancer (BC) survival is limited, necessitating personalized treatment strategies.
• Current models offer minimal survival extension, highlighting the need for advanced predictive tools.

Purpose of the Study:

• To develop high-confidence deep learning (DL) models for predicting individual survival in stage IV BC patients.
• To guide therapy by considering patient-specific circumstances and potential treatment outcomes.

Main Methods:

• Utilized the SEER-Medicare linked dataset (1991-2016) with 14,312 stage IV BC patients.
• Employed DL models (DeepSurv, DeepHit, Nnet-survival, Cox-Time) incorporating time-fixed and time-varying covariates.
• Validated models using random sampling for training, validation, and testing, with fine-tuning via Amazon SageMaker.

Main Results:

• Achieved a prediction error below 10% by integrating time-fixed and time-varying covariates.
• Demonstrated the models' ability to generate accurate, individual patient survival curves.
• Confirmed the proof of principle for DL-based survival prediction in advanced breast cancer.

Conclusions:

• Developed DL models capable of high-confidence survival prediction for individual stage IV BC patients.
• These models can assist in treatment decisions by simulating 'what-if' scenarios.
• The goal is to optimize therapy for extended survival and minimized adverse events in metastatic breast cancer.