Introduction: Accurate ICU discharge disposition predictions help care teams coordinate post-acute services but often show uneven performance across demographics. These disparities are amplified by hospital-level variation in data distributions and care practices, limiting model fairness and generalizability. We propose a federated learning approach combining global representation learning with hospital-specific personalization and fairness constraints to reduce demographic performance gaps within and across hospitals without data sharing or sacrificing accuracy.
Methods: We used eICU CRD, which includes data from 208 USA hospitals, to predict ICU discharge disposition every 2 hours after admission. The outcome variable has 8 categories, such as home, nursing facility, and rehab. Each hospital was treated as independent client within a federated learning framework, using the FedPer algorithm to train a shared global encoder and hospital-specific personalized classifier heads. Input features included vital signs, lab values, past medical history, gender and ethnicity among 42 others. To address subgroup disparities, we incorporated fairness constraints as regularization terms during local personalization. We optimized for similar predicted class distributions across gender and ethnicity groups, and equalized error to reduce subgroup misclassification differences. This model was compared to a baseline personalized model without fairness constraints.
Results: The fairness aware personalized federated model achieved strong and consistent performance across hospitals with a mean accuracy of 80.1% (±5.4%), computed as a weighted mean of each hospital’s test-set accuracy. The primary advantage of our approach was improved demographic equity. Compared to the baseline, the mean cross-hospital accuracy disparity between gender and ethnicity subgroups decreased by 21%. The mean within-hospital accuracy gap fell by 6.1 percentage points. Subgroup error disparities decreased in more than 75% of hospitals.
Conclusions: This study shows that fairness-aware personalized federated models can reduce prediction disparities across gender and ethnicity while aiding earlier, more equitable ICU discharge planning. Future work should include other fairness constraints like insurance and age to improve clinical utility.
