Risk prediction models for depression in older adults with cancer

Background: Depression in people with cancer is common and debilitating, but few instruments exist for early identification. We aimed at developing streamlined Risk Prediction Models (Arturo RPMs) for identifying individuals at higher risk for depression. Methods: Predictors of depression were identified from a review of available literature. Then, we used data from the Survey of Health, Ageing and Retirement in Europe (SHARE) prospective study. SHARE recruited community residing adults aged 55 years or older, from which we selected those who reported having received a diagnosis of cancer. The outcome was the presence of depression (EURO-D score ≥ 4) in the Classification Approach (CA), and the severity of depression at a two-year follow up evaluation (EURO-D sum-score) in the Regression Approach (RA). Multiple RPMs were developed using combinations of sample balancing techniques (no balancing, under- or oversampling), learning methods (Generalized Linear Models (GLM), Decision Trees (DT), Random Forest (RF)) and variable selection methods (none, Backward (BW) and Forward (FW) sequential, Genetic Algorithm (GA)). Results: We identified 90 predictors of depression that were measured in the SHARE dataset and, combining waves 4 to 8, selected a sample of 4057 participants with cancer. Of these, 33.5% were depressed at 2 years follow-up. In the classification approach the RPM based on undersampling, GLM and GA variable selection used 34 predictors and reached satisfying accuracy (74.4%, AUC-ROC: 0.80; PPV: 84.7%; NPV: 60.1%). In the regression approach, the GLM model with Genetic Algorithm reached the best accuracy (75.1%, AUC-ROC: 0.81; PPV: 70.5%; NPV: 76.2%). The calibration curve suggested a satisfactory level of prediction, homogeneous at all levels of risk thresholds. Using a threshold of 50% risk, the model yields a PPV of 80% and an NPV of 75%. Conclusions: The Arturo RPMs can identify older adults with cancer who are at higher risk of developing depression over the following two years. The model is freely available as a web-based calculator for use by individuals, clinicians, and policy makers. Arturo might help to target preventive interventions. Clinical trial number: Not applicable.