Individualized Prospective Prediction of Opioid Use Disorder.

Area of Science:

• Public Health
• Data Science
• Clinical Informatics

Background:

• Opioid use disorder (OUD) affects millions globally, necessitating early detection and intervention strategies.
• Existing machine learning (ML) models for OUD prediction often lack validation on representative data and prospective testing.
• This limits their real-world applicability for identifying new OUD cases.

Purpose of the Study:

• To develop and prospectively validate a machine learning model for predicting individual opioid use disorder (OUD) cases.
• To utilize representative, large-scale Canadian administrative health data for model development and validation.
• To assess the model's accuracy in predicting future OUD cases.

Main Methods:

• An ensemble machine-learning model was developed using administrative health data from 2014-2018 (n=699,164).
• Model performance was validated on a hold-out sample (2014-2018, n=174,791) and prospectively on a 2019 sample (n=316,039).
• Opioid use disorder (OUD) diagnosis was based on International Classification of Diseases (ICD) codes.

Main Results:

• The model prospectively predicted OUD cases in 2019 with high accuracy (balanced accuracy: 86%, sensitivity: 93%, specificity: 79%).
• Top risk factors identified included indicators of opioid use and a history of other substance use disorders.
• The study identified 6409 OUD cases in the 2019 prospective sample.

Conclusions:

• This study demonstrates the feasibility of individualized, prospective OUD prediction using ML on large administrative health datasets.
• Prospective prediction models are crucial for advancing early detection of OUD in clinical practice.
• The findings support the potential of ML-driven tools for timely OUD intervention.