Predicting Visit Cost of Obstructive Sleep Apnea Using Electronic Healthcare Records With Transformer.

Area of Science:

• Healthcare analytics
• Translational engineering
• Machine learning in healthcare

Background:

• Obstructive sleep apnea (OSA) prevalence is increasing globally, driven by rising obesity rates.
• Effective OSA treatment incurs significant social and financial healthcare costs.
• Accurate prediction of OSA patient healthcare expenditures is essential for fiscal management and resource allocation.

Purpose of the Study:

• To develop a method for predicting annual healthcare visit expenses for obstructive sleep apnea (OSA) patients.
• To address challenges posed by limited high-quality patient data, where only data from patients with over 365 days of follow-up are typically usable for cost prediction models.

Main Methods:

• A translational engineering approach utilizing two Transformer models.
• The first Transformer model augments input data using shorter patient visit histories.
• The second Transformer model predicts costs, incorporating enriched data and cases with over a year of follow-up.

Main Results:

• The two-model solution significantly improved prediction performance (R-squared) from 88.8% to 97.5% compared to a single-model approach using only long-term data.
• Augmenting data with the first Transformer model improved baseline model performance (R-squared) from 61.6% to 81.9%.

Conclusions:

• The proposed method maximizes the utility of available high-quality OSA patient data.
• This approach couples data augmentation with cost prediction, offering a practical solution for healthcare settings with limited data.
• The findings support enhanced patient care and resource allocation in OSA management.