Wrist-worn photoplethysmography (PPG) and tri-axial accelerometry, already standard in many consumer wearables, open the door to home-based assessment of sleep-disordered breathing. Using the public DREAMT corpus with overnight polysomnography ground truth, this study examined whether these two signals alone can separate the five clinically recognized breathing states: normal, hypopnea, obstructive, central and mixed apneas. After basic cleaning and quality checks, statistical, spectral and motion features were extracted from each recording and several lightweight machine-learning models were evaluated. Tree ensembles outperformed other approaches: Random Forest and LightGBM achieved balanced accuracy of 62% and Cohen’s \(\kappa \) of \(\approx \) 0.53 on an independent test set. Both models attained high recall above 70% for central apneas and normal breathing, moderate recall near 60% for obstructive apneas and hypopneas, and lower recall for mixed events, mirroring challenges reported in earlier studies. These findings demonstrate that commodity-grade PPG combined with accelerometry can already support practical five-class screening, laying the groundwork for explainable wrist-based tools capable of bringing full-phenotype sleep breathing monitoring into everyday life.