UArizona at the MADE1.0 NLP Challenge

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Computational Linguistics

Background:

• Electronic Health Records (EHRs) contain valuable clinical information.
• Extracting medication information and adverse drug events (ADEs) from EHRs is crucial for patient safety and pharmacovigilance.
• Public challenges like MADE1.0 drive innovation in this field.

Purpose of the Study:

• To develop and evaluate Named Entity Recognition (NER) and Relation Identification (RI) systems for the MADE1.0 challenge.
• To extract medication and adverse drug event information from clinical text.
• To assess the performance of proposed systems in a competitive setting.

Main Methods:

• A neural network-based Named Entity Recognition (NER) system was developed, incorporating local and contextual word features.
• A Support Vector Machine (SVM) based pairwise relation classification system was employed for identifying relationships between medical entities and attributes.
• Systems were trained and evaluated on datasets provided by the MADE1.0 challenge.

Main Results:

• The NER system achieved an 81.56% F1 score for medical entity recognition.
• The relation classification system achieved an 83.18% F1 score for Task 2 and 59.85% for Task 3.
• The UArizona team ranked among the top three participants for Task 2 and Task 3 of the MADE1.0 challenge.

Conclusions:

• The proposed neural NER and SVM-based RI systems demonstrate strong performance in extracting medication and adverse drug events from EHRs.
• The results highlight the effectiveness of combining local and contextual features for medical NER.
• The systems' competitive performance validates their potential for real-world clinical applications.