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

Personalizing initial positive airway pressure (PAP) settings for obstructive sleep apnea (OSA) remains challenging in routine practice. We developed FusionPAP, a cross-modal teacher–student framework for multi-task recommendation of initial PAP parameters, including pressure range, mask type, and humidification. The reference therapeutic pressure was derived from attended, in-laboratory manual PAP titration polysomnography (PSG). An LLM-driven multi-agent teacher first integrates structured variables with pre-treatment free-text reports to emulate multidisciplinary reasoning, and this knowledge is then distilled into a lightweight student model that relies only on structured data during inference. In a multicenter retrospective cohort of 415 patients with temporally separated external validation, the deployed student achieved a target-pressure mean absolute error of 1.14 cmH2O, a mask recommendation AUROC of 0.869, a humidification quadratic weighted kappa of 0.807, and a 0.00% boundary violation rate. FusionPAP outperformed conventional tabular baselines while remaining efficient on standard clinical hardware. These findings suggest that FusionPAP may serve as a practical and scalable clinical decision-support tool for estimating titration-derived therapeutic pressure together with associated PAP initialization parameters in OSA.