A Calibrated Ensemble Algorithm to Address Data Heterogeneity in Machine Learning: An Application to Identify Severe SLE Flares in Lupus Patients

Motivated to address the inconsistency between essential i.i.d. assumption in machine learning theory and data heterogeneity real-world applications, we propose a novel calibrated ensemble (CE) algorithm facilitate with diverse subgroups. Unlike traditional framework which each learner is trained independently using entire dataset, our method exploits strengths of various models by training them simultaneously forming model-ergonomic subgroups as part process. Consequently, unique subset based on their individualized predictive strength. Clinically, can interpret model an expert specializing treating patients particular disease manifestations. We evaluate CE motivating domain identifying lupus severe SLE flares 1541 clinical encounters Mass General Brigham (MGB) Lupus Cohort. Our experimental results demonstrate efficacy across seven performance evaluation metrics compared five individual regular approaches. further utilize ANOVA Tukey HSD post-hoc statistical analysis discover characteristic features clusters for interpretations.