Identifying Opioid Use Disorder from Longitudinal Healthcare Data using a Multi-stream Transformer

Area of Science:

• Health Informatics
• Artificial Intelligence
• Clinical Data Analysis

Background:

• Opioid Use Disorder (OUD) presents a significant public health challenge, incurring substantial economic costs.
• Longitudinal healthcare data analysis is essential for understanding and mitigating complex health issues like OUD.
• Existing models for OUD identification have limitations in comprehensively analyzing multi-modal patient data.

Purpose of the Study:

• To introduce MUPOD, a novel multi-stream transformer model for improved identification of Opioid Use Disorder.
• To leverage longitudinal healthcare data, including medications and diagnoses, for enhanced OUD detection.
• To evaluate MUPOD's performance against traditional and deep learning models in a large patient cohort.

Main Methods:

• Development of a multi-stream transformer architecture (MUPOD) capable of processing diverse healthcare data.
• Simultaneous analysis of medication and diagnosis data streams within the MUPOD framework.
• Application of MUPOD to a large dataset of 392,492 patients with long-term back pain.

Main Results:

• MUPOD demonstrated significantly superior performance in identifying Opioid Use Disorder compared to existing models.
• The model effectively analyzed and integrated information from multiple healthcare data streams.
• High accuracy in OUD identification was achieved in a large, real-world patient population.

Conclusions:

• The MUPOD model offers a promising advancement in the early identification and management of Opioid Use Disorder.
• Multi-stream transformer models are effective for analyzing complex longitudinal healthcare data.
• Improved OUD detection through advanced AI can aid in addressing the public health crisis.