AI Model Estimates Initial Sleep Apnea Therapy Parameters in Multicenter Retrospective Study

A cross-modal teacher–student artificial intelligence framework named FusionPAP can accurately recommend initial positive airway pressure settings for patients with obstructive sleep apnea, according to a retrospective study published in BMC Medical Informatics and Decision Making. Evaluated in a multicenter cohort featuring temporally separated external validation, the deployed student model predicted target therapeutic pressure with a mean absolute error of 1.14 cmH2O while achieving a 0.00 percent boundary violation rate. Investigators designed the multi-task model to help address the persistent clinical challenge of personalizing initial positive airway pressure configurations in routine clinical practice, using reference therapeutic pressures derived directly from attended, in-laboratory manual titration polysomnography examinations.

The targeted parameters include the prescribed therapeutic pressure range, the appropriate mask type, and the required humidification level for each individual patient.

To capture complex clinical decision-making across disparate data types, the investigators constructed a cross-modal teacher–student framework. The teacher model employs a multi-agent system driven by large language models to integrate structured clinical variables with pre-treatment free-text medical reports, directly emulating the multidisciplinary reasoning that clinicians apply when evaluating sleep apnea cases. During practical inference, the deployed student model operates without requiring any free-text inputs, relying exclusively on structured clinical data while retaining the multidisciplinary insights distilled from the multi-agent teacher.

The investigators evaluated the deployed student model in a multicenter retrospective cohort of 415 patients, incorporating temporally separated external validation to verify performance across distinct time periods. In validation testing, the student model achieved a target-pressure mean absolute error of 1.14 cmH2O, an area under the receiver operating characteristic curve of 0.869 for mask recommendation, and a quadratic weighted kappa of 0.807 for humidification settings. The model also recorded a 0.00 percent boundary violation rate, demonstrating strict adherence to clinically permissible parameter bounds. In direct comparisons, FusionPAP outperformed conventional tabular baseline models across these tasks while running efficiently on standard clinical hardware. The study authors concluded that FusionPAP represents a practical and scalable clinical decision-support tool for estimating titration-derived therapeutic pressure alongside associated initial positive airway pressure initialization parameters in obstructive sleep apnea.

Subject of Research: Machine learning for recommending initial positive airway pressure settings in obstructive sleep apnea

Article Title: FusionPAP: a cross-modal teacher-student framework for multi-task recommendation of initial PAP settings in obstructive sleep apnea

Article References: Zeng, Y., Nan, D., Song, A., Lin, H., Fan, J., Liu, D., Hu, J., Melibary, M., Du, X., Sun, W., Gao, Y., & Chen, L. (2026). FusionPAP: a cross-modal teacher-student framework for multi-task recommendation of initial PAP settings in obstructive sleep apnea. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03958-4

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03958-4

Keywords: Obstructive sleep apnea, Positive airway pressure, FusionPAP, Teacher-student framework, Multi-task recommendation, Therapeutic pressure, Polysomnography, Manual titration, Mask recommendation, Humidification settings, Multi-agent systems, Large language models, Knowledge distillation, Multicenter retrospective cohort, Clinical decision support, External validation, Boundary violation rate, Medical informatics