A Unified Machine Reading Comprehension Framework for Cohort Selection.

Area of Science:

• Clinical informatics
• Natural Language Processing (NLP)
• Machine Learning

Background:

• Cohort selection is critical for clinical research but often treats criteria independently.
• Existing methods overlook the semantic meaning and interrelations among selection criteria.
• This limitation hinders accurate patient identification for studies.

Purpose of the Study:

• To propose a novel unified machine reading comprehension (MRC) framework for cohort selection.
• To model relationships among cohort selection criteria using a cross-criterion attention mechanism.
• To enhance the accuracy and efficiency of identifying eligible participants for clinical research.

Main Methods:

• Developed an MRC framework using generated questions from criteria and patient record clues as passages.
• Deployed state-of-the-art MRC models (BERT, BioBERT, NCBI-BERT, RoBERTa) for matching criteria to patient data.
• Introduced a cross-criterion attention mechanism to capture interdependencies between selection criteria.

Main Results:

• The NCBI-BERT MRC model with cross-criterion attention achieved a 0.9070 F1-score on the N2C2 dataset and 0.8353 on MIMIC-III.
• Performance was competitive with expert-defined rule-based systems on the N2C2 dataset.
• Integrating rules for mathematical logic criteria further improved performance to a benchmark F1-score of 0.9163.

Conclusions:

• The proposed MRC framework effectively addresses limitations of independent criterion processing in cohort selection.
• NCBI-BERT MRC with cross-criterion attention shows strong performance, especially when combined with rule-based systems for specific criteria.
• This hybrid approach offers a promising direction for improving automated cohort identification in clinical research.