Combination of conditional random field with a rule based method in the extraction of PICO elements

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Research

Background:

• Extracting PICO elements from unstructured medical information is difficult due to information volume and semantic complexity.
• PICO (Patient/Problem, Intervention, Comparison, Outcome) framework is crucial for clinical questions.
• Existing methods struggle with comprehensive PICO element extraction.

Purpose of the Study:

• To investigate methods for extracting PICO elements from medical literature.
• To enhance PICO extraction by combining machine learning methods (MLMs) with rule-based methods (RBMs).
• To facilitate answering clinical questions using the PICO framework.

Main Methods:

• Analyzed MLM aspects influencing PICO extraction quality.
• Developed a hybrid approach combining MLMs and RBMs for improved PICO retrieval.
• Utilized a corpus of 1000 medical abstracts for experimentation.

Main Results:

• Achieved an F-score of 80% for Patient/Problem (P), 64% for Intervention (I), and 92% for Outcome (O) elements.
• Results are competitive, especially considering the low representation of P and I elements in the training corpus (6.5% and 5.8%).
• The combined approach demonstrated superior performance compared to MLMs alone.

Conclusions:

• PICO element extraction from medical texts is a challenging task.
• MLMs alone exhibit good precision but low recall on non-representative corpora.
• RBMs effectively augmented MLMs, increasing recall and overall performance, highlighting the benefit of hybrid methods.