Cohort selection for clinical trials using multiple instance learning.

Area of Science:

• Medical Informatics
• Machine Learning
• Clinical Trial Recruitment

Background:

• Identifying eligible patients for clinical trials from electronic health records (EHRs) is complex, often necessitating analysis across multiple patient records.
• Current methods may not fully leverage the longitudinal nature of patient data for accurate cohort selection.

Purpose of the Study:

• To investigate and compare the effectiveness of different computational methods for identifying patients meeting clinical trial eligibility criteria using longitudinal EHR data.
• To evaluate the performance of Multiple Instance Learning (MIL) against rule-based and single-instance learning classifiers for cohort selection.

Main Methods:

• Utilized the n2c2 cohort selection dataset, comprising 2-5 annotated EHRs per patient for thirteen selection criteria.
• Formulated the cohort selection problem as a Multiple Instance Learning (MIL) task.
• Compared MIL classifiers with rule-based and single-instance learning approaches.

Main Results:

• The best MIL-based approach achieved a micro-F score of 0.8765, ranking among the top ten results.
• MIL classifiers outperformed single-instance methods for criteria requiring comprehensive understanding across multiple patient records.
• Rule-based and single-instance methods were more effective for criteria not needing cross-record synthesis.

Conclusions:

• Cohort selection from longitudinal patient records can be effectively modeled as a Multiple Instance Learning (MIL) problem.
• MIL classifiers serve as a valuable supplement to rule-based methods, offering superior performance compared to single-instance learning approaches for complex cohort identification.