DCLA: Deep Cooperative Learning for Advancing Automated Annotation of Electronic Medical Records in Cerebral Palsy.

Abstract

Automated annotation of electronic medical records for patients with cerebral palsy (CP) is crucial for downstream clinical applications. However, most existing methods lack mechanisms to verify model predictions before their acceptance and suffer from labeled data scarcity. To address this challenge, we propose a Deep Cooperative Learning for Automated Annotation (DCLA) framework. DCLA integrates named entity recognition (NER) and relation extraction (RE) models that employ the multi-head attention mechanism and the global pointer to handle complex entities and relations. Building on this foundation, a cooperative learning (CL) mechanism is introduced to evaluate prediction quality through score matrices for sample ranking and selection. Low-quality predictions are verified by annotators, while high-quality predictions are accepted automatically, enabling iterative retraining with cooperatively labeled data. Experiments on a CP-specific corpus demonstrate that DCLA's NER and RE models outperform state-of-the-art methods, while the CL mechanism enhances proofreading efficiency. Overall, DCLA enhances proofreading efficiency, mitigates data scarcity, and supports continuous model refinement.