| Sumario: | Simple Summary: Most patients with stage IV breast cancer cannot be cured; hence, the goal of therapy is to prolong survival while maintaining a good quality of life. Guidance for treating patients in the nearly limitless scenarios after disease progression following the first few lines of therapy comes from clinical trials conducted in patients who have had limited treatment and do not address these many complex scenarios. We used a form of artificial intelligence called deep learning to investigate patients with stage IV breast cancer from the SEER-Medicare dataset to develop predictive modeling in individual patients. We achieved approximately 90% accuracy in survival predictions in individual patients by combining features from the patient, the cancer, their treatments, and adverse events from treatment. These models can guide decision-making regarding patients' treatment, where guidance is unavailable, to extend their lives while maintaining a good quality of life. Importance: Treatment of women with stage IV breast cancer (BC) extends population-averaged survival by only a few months. Here, we develop a model for identifying individual circumstances where appropriate therapy will extend survival while minimizing adverse events. Objective: Our goal is to develop high-confidence deep learning (DL) models to predict survival in individual stage IV breast cancer patients based on their unique circumstances generated by patient, cancer, treatment, and adverse event variables. We previously showed that predictive DL survival modeling of potentially curable stage I–III patients can be improved by combining time-fixed and time-varying covariates. Here, we demonstrate that DL-based predictive survival modeling in stage IV patients, where treatment does not offer a cure, can generate accurate individual survival predictions by considering subsequent lines of potential treatment to guide therapy. This guidance is rarely obtainable in the nearly limitless scenarios of metastatic disease. Design, Setting, and Participants: We applied the SEER-Medicare linked dataset from 1991 to 2016 to investigate 14,312 unique stage IV patients with 1,880,153 entries. We used DeepSurv- and DeepHit-, Nnet-survival- and Cox-Time DL-based predictive models to consider the combination of time-fixed and time-varying covariates at each visit for each patient. We adopted random sampling to divide the input dataset into training, validation, and testing sets. We verified the models' implementation using the pycox package and fine-tuned the models using the open-source library Amazon SageMaker Python SDK 2.232.2 (software development kit). Our results demonstrated the proof of principle of the models by generating individual patients' survival curves. Conclusions and Relevance: By extending the survival prediction models to consider stage IV BC patients' time-fixed and time-varying covariates, we achieved a prediction error below 10%. Based on their circumstance-specific situations, these models can predict survival in individual stage IV patients with high confidence. The models will serve as an important adjunct to treatment decisions in patients with stage IV BC and test what-if scenarios of treatment or no treatment options to optimize therapy for extending patient lives and minimizing adverse events.
|
|---|