Obstructive sleep apnoea (OSA) is a prevalent sleep disorder with substantial health and economic consequences, affecting over three billion individuals worldwide. Type III Polysomnography (PSG3), the Home Sleep Apnea Test (HSAT), is a convenient athome screening option. However, its diagnostic accuracy is under scrutiny. This research sets out a machine learningbased OSA severity classification model that uses PSG3 data enriched with sleep position information and multisensor physiological signals. It analysed a dataset of 532 patients from certified sleep centres in Indonesia using six physiological signals: nasal airflow (PFlow), thoracic movement, plethysmography, oxygen saturation (SpO2), heart rate (BPM), and snoring intensity. Data preprocessing included pre-emphasis filtering, skewness correction, standardization, and SMOTE for class balancing. Six ML classifiers were evaluated: Random Forest, Decision Tree, K-Nearest Neighbours, Gradient Boosting, Support Vector Machine, and Logistic Regression. RF achieved the highest classification accuracy at 98.56%, decisively outperforming previous studies that excluded sleeping position and reported up to 91% accuracy. This confirmed the model's robustness using fivefold cross-validation. The results show the effectiveness of integrating multi-sensor fusion and positional context in PSG3-based diagnostics. There is no doubt that this offers promising potential for real-world deployment in telemedicine and home-based OSA screening. © 2025 IEEE.