Predictive Modeling of Dormant Breast Cancer Recurrence Associated with Anxiety and Depression in SEER-Medicare Patients Using Deep Learning

Importance. Up to 50% of patients with localized breast cancer (BC) continue to recur after initial treatment for more than 20 years. Recurrent cancer is not curable, and the goal of therapy is the prolongation of life. Many variables, including adrenergic stimuli associated with anxiety and/or depression (anxiety/depression), have been linked to recurrence. Demonstrating this effect in predisposing circumstances could guide investigations of prevention.Objective. This study aims to develop predictive modeling of breast cancer recurrence to identify specific circumstances in which anxiety/depression are associated with breast cancer recurrence.Design, Setting, and Participants. We used the SEER-Medicare dataset linked dataset to investigate women diagnosed with stage I, II, and III breast cancer enrolled at 65 years or older for age eligibility. We collected data on the date of diagnosis, patient-, cancer-, treatment- and adverse events-associated variables, and diagnosis of anxiety and/or depression. We identified recurrence by a documented new diagnosis of recurrent, contralateral, or stage IV BC, new chemo-, bio-, hormone, or radiotherapy outside the 4-month window after completion of initial therapy. We extended four deep learning (DL)-based predictive survival models (Deep-24 Surv and DeepHit. Nnet-survival and Cox-Time) that deal with right-censored time-to-event data and found greater than 95% concordance in each hyperparameter to demonstrate prediction validity. We generated 96 hypothetical patient recurrence curves distributed into categories of stage I or III, White (W), or African American (AA) without or with a diagnosis of anxiety and/or depression.ResultsThe patient's age, race and comorbidities, cancer stage, and hormone status distributions mirrored previously published values. Recurrence rates and time to recurrence were worse in patients who were in the AA group, ER-/PR-, and higher stage. Approximately 36% of the patients had anxiety and/or depression, with higher prevalence in W patients. Time to diagnosis ranged from 23-35 months. Recurrence rates in these patients were 27.9% higher than in patients without these diagnoses, with greater impact in Stage I, ER+/PR+, W, and Hispanic patients. The hypothetical patient recurrent survival curves we generated by DL-based predictive modeling demonstrated wide ranges of predicted recurrence-free survival curves within each grouping, with median survivals varying by more than a decade within each group. There were some long-term survivors in many of the categories, but most patients with the anxiety/depression group had far fewer long-term recurrence-free survival curves than patients without the diagnosis. These results demonstrate the exceptional variability of median predicted survival rates and their modifications by chronic adrenergic stimuli.Conclusions and Relevance.This study demonstrates, for the first time, that the adrenergic stressors anxiety and depression increase the average population recurrence rate of dormant breast cancer in elderly SEER-Medicare patients with stage I, II, and III disease. The predictive impact of population-derived data is limited in individual patients. However, DL-based predictive modeling that accounts for unique circumstance-specific scenarios is able to model individual patient predictive recurrence and generate predicted survival probabilities and responses to psychological stressors with a 95% degree of confidence. The modeling demonstrates the exceptional variability of individual recurrence probabilities and will enable clinical investigation of the impact of variables in specific scenarios.