Fairness-Constrained Risk Modeling for Anastomotic Leak Prediction Under Distribution Shift, Label Noise, and Heterogeneous Clinical Workflows

• Karim Haddad

  Department of Computer and Communications Engineering, Lebanese German University, Rue Fouad Chehab, Sahel Alma, Jounieh, Mount Lebanon, Lebanon

  Author

• Rami Khoury

  Faculty of Computer Science and Information Technology, Arts, Sciences and Technology University in Lebanon (AUL), Cola Intersection, Verdun Street, Beirut, Lebanon

  Author

• Tarek Nassar

  Department of Informatics Engineering, Islamic University of Lebanon, Khaldeh Main Road, Al Wardaniyeh District, Khaldeh, Mount Lebanon, Lebanon

  Author

Anastomotic leak prediction has become an important target for clinical machine learning because the event is uncommon, severe, and strongly time-sensitive. The practical question is not only whether a model can separate patients who will later develop a leak from those who will not, but whether that separation is equitable across patient populations, hospitals, and care pathways. Fairness in this setting is difficult because outcome prevalence is low, labeling is imperfect, surveillance intensity varies, and post-operative management changes the very event that the model is asked to predict. These features make standard fairness language too narrow when used without attention to clinical utility and data generation. This paper develops a framework for fairness in anastomotic leak prediction that integrates statistical risk estimation, calibration, subgroup robustness, causal reasoning, and deployment governance. The discussion treats fairness as a property of the full decision pipeline rather than of a score alone. It examines how imbalance, missingness, treatment selection, procedure mix, and site-specific workflows can create systematic disparities even when sensitive attributes are excluded from the feature set. It then formulates fairness-aware optimization strategies that combine worst-group risk control, calibration preservation, and utility-sensitive threshold design. A complementary audit protocol is described for temporal validation, external transport, intersectional subgroup analysis, and prospective surveillance after deployment. The resulting perspective does not assume a single fairness criterion is sufficient. Instead, it argues that equitable anastomotic leak prediction requires a layered design in which model development, evaluation, and clinical integration are jointly constrained by the need to avoid uneven error burden and uneven access to beneficial intervention.

• PDF

Copyright (c) 2025 authors

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.