Classification of drug use patterns.

Area of Science:

• Pharmacovigilance
• Health Informatics
• Machine Learning in Healthcare

Background:

• Long-term prescription data analysis is complex due to time-varying drug use patterns.
• Traditional methods often oversimplify usage, leading to information loss and biased results.
• Accurate characterization of drug exposure is crucial for understanding treatment effectiveness and safety.

Purpose of the Study:

• To develop and validate an unsupervised machine learning approach for characterizing longitudinal prescription data.
• To identify distinct patient groups based on metformin usage patterns.
• To assess the association between identified user groups and diabetes progression.

Main Methods:

• Utilized administrative healthcare databases covering 1.3 million residents.
• Engineered features capturing average dose, proportion of days covered (PDC), dose change, and dose variability.
• Applied K-means clustering to group patients based on these features.

Main Results:

• Identified 27,786 metformin users and six distinct usage groups: intermittent, decreasing dose, increasing dose, high dose, stable medium dose, and variable medium dose.
• Demonstrated stability and interpretability of feature distributions across different follow-up lengths.
• Found that patients with decreasing or varying dose patterns had an increased likelihood of diabetes progression.

Conclusions:

• The unsupervised machine learning method effectively characterizes longitudinal prescription data into clinically meaningful groups.
• This approach offers a more nuanced understanding of drug use patterns than conventional methods.
• Identifying specific user groups can aid in predicting disease progression and tailoring patient care.