Using Polysomnography-Derived Parameters to Predict Cardiovascular Outcome

Abstract

Cardiovascular disease (CVD) remains the leading cause of mortality worldwide, and obstructive sleep apnoea (OSA) is recognised as an independent risk factor. Polysomnography (PSG), the gold standard for OSA assessment, captures physiological signals that can be transformed into parameters with prognostic value. The apnoea–hypopnoea index, although widely used for clinical diagnosis, is less informative for predicting CVD outcomes. In contrast, oximetry-derived measures provide a more detailed characterisation of hypoxaemia. However, variations in desaturation criteria, parameter definitions, and computational methods have introduced methodological inconsistencies, limiting comparability across studies. This thesis investigates whether PSG-derived parameters can improve the prediction of CVD mortality and whether explainable machine learning frameworks can deliver robust and clinically meaningful predictions at the individual level.

Two experiments were conducted to address this aim. Experiment 1 systematically compared three major desaturation area–based algorithms within a unified computational framework. The results demonstrated that algorithmic differences influence both parameter values and predictive performance for CVD mortality, explaining inconsistencies reported in prior studies and identifying a robust, best-performing method.

Experiment 2 evaluated whether combining PSG-derived parameters improves prediction beyond single metrics, with a focus on short-term outcomes. The combined approach enhanced predictive performance. Building on these findings, an explainable machine learning framework integrating PSG-derived parameters with demographic, lifestyle, and clinical data was developed. The model achieved strong performance, demonstrated generalisability, and remained interpretable. Its reduced reliance on specialised clinical inputs supports its potential application in large-scale screening, resource-limited settings, and home-based monitoring.