Machine learning-assisted literature screening for a medication-use process-related systematic review

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Evidence Synthesis

Background:

• Literature reviews for annual series like medication-use process (MUP) and ambulatory care MUP (ACMUP) generate extensive datasets.
• Efficiently screening large volumes of literature is crucial for timely evidence synthesis.
• Traditional manual screening can be time-consuming and resource-intensive.

Purpose of the Study:

• To introduce and evaluate a novel methodology using machine learning (ML) for assisting article screening in MUP and ACMUP reviews.
• To demonstrate the efficacy of training ML models with external datasets for predicting article relevance.
• To reduce the manual workload in literature reviews by excluding non-relevant search results.

Main Methods:

• Developed and applied ML algorithms trained on external datasets to predict article relevance.
• Utilized past screening decisions from MUP and ACMUP review series as training data.
• Simulated the ML model's performance against known manual screening decisions to assess accuracy and time savings.

Main Results:

• The ML approach correctly identified 187 out of 192 relevant studies.
• In simulations involving 17,227 unique studies, ML enabled the exclusion of 13,201 studies without manual screening.
• Maintained a recall rate of 95% or greater for relevant articles across simulations.

Conclusions:

• The ML-driven article screening methodology is effective and applicable to systematic reviews and ongoing review series.
• This approach significantly saves time in the evidence synthesis process for MUP and other pharmacy practice disciplines.
• Facilitates more expeditious publication of evidence syntheses, supporting pharmacists in improving care efficiency and reducing medication errors.