preprints.org > biology and life sciences > life sciences > doi: 10.20944/preprints202411.0041.v1 (registering DOI)

Preprint Article Version 1 This version is not peer-reviewed

Version 1 : Received: 31 October 2024 / Approved: 1 November 2024 / Online: 4 November 2024 (03:44:19 CET)

Importance. Treatment of women with stage IV breast cancer (BC) extends population-averaged survival by only a few months. These investigations will identify individual circumstances where appropriate therapy will extend survival while minimizing adverse events.Objective. Our goal is to develop high confidence deep learning (DL) models for predicting survival in individual stage IV breast cancer patients based on their unique circumstances generated by patient, cancer, treatment and adverse events variables. Our plan is to improve the predictive accuracy of deep learning (DL)-based predictive models by considering time-fixed covariates, i.e., patient and cancer data at the time of diagnosis, together with time-varying events that occur after initial diagnosis. Design, Setting, and Participants. We used the SEER-Medicare linked dataset from 1991-2016 to investigate women diagnosed with stage IV BC who enrolled at 65 years or older for age eligibility. We outlined time-fixed variables, including date of diagnosis, age, race, marital status, breast cancer stage, tumor grade, laterality, estrogen receptor (ER), progesterone receptor (PR), and human epidermal receptor 2 (HER2) status and comorbidity index, prior therapy, adverse events, and changes in comorbidity. We delineated the time-varying covariates at each visit, including administered treatments, adverse events, comorbidity index, and age. We extended four DL-based predictive survival models (DeepSurv and DeepHit. Nnet-survival and Cox-Time) that deal with right-censored time-to-event data to consider both a patient's time-fixed covariates and a patient's time-varying covariates. We predicted the survival of five hypothetical patients to demonstrate the model's utility. We found high concordance between the performance metric, time-dependent concordance, and each of the model's hyperparameters to demonstrated prediction validity.Results. By incorporating time-varying variables with the time-fixed variables, the models reduced the error rates of the concordance index, the most commonly applied evaluation metric in survival analysis, from 28-38% to 2-12%, and significantly improved the integrated Brier score, a metric of the model's discrimination and calibration. Conclusions and Relevance. By combining the consideration of time-fixed with those time-varying variables in our predictive models, we decreased the predictive error rate to under 10% in the predicted survival of stage IV BC patients. Our established models can predict survival in individual stage IV patients with high confidence based on their circumstance-specific situations generated by considering their unique patient, cancer, treatment and adverse events variabels. These models will serve as an important adjunct to treatment decisions in patients with stage IV BC to optimize therapy for extending patient lives and minimizing adverse events.

Deep learning; Breast Cancer; Stage IV; Overdiagnosis and Overtreatment; SEER-Medicare Linked Dataset

Biology and Life Sciences, Life Sciences

* All users must log in before leaving a comment