Correlated Output Constraints Improve Clinical Code Assignment When Labels Are Missing in Records Hospital Records

Abstract

The automation of clinical code assignment from electronic health records is a critical task in modern healthcare administration, directly impacting billing, epidemiological research, and patient care management. Despite the proliferation of deep learning models for text classification, a persistent challenge in this domain is the pervasive presence of missing labels in training datasets. Clinical coding is inherently subjective and prone to omission due to clinician fatigue, incomplete documentation, or strict coding guidelines that limit the number of recorded diagnoses. Traditional multi-label classification models trained on such incomplete ground truth data suffer from significant performance degradation, as unannotated true positive codes are incorrectly penalized as negative samples during training. To address this limitation, this paper introduces the Correlated Output Constraints Network, a novel architecture designed to robustly assign clinical codes in label-missing scenarios. The proposed methodology leverages the inherent hierarchical and co-occurrence correlations among medical conditions to constrain the model predictions. By integrating medical ontology structures and empirical label correlations into a constraint enforcement layer, the network effectively imputes missing labels during the learning process and prevents the suppression of unannotated but clinically logical codes. Comprehensive evaluations demonstrate that the proposed architecture maintains high predictive accuracy and robustness even when a substantial proportion of ground truth labels are missing, outperforming existing state-of-the-art approaches in automated clinical coding.