Introduction: Obstructive apneas (OA) occur despite the use of positive airway pressure (PAP) on obstructive sleep apnea (OSA) patients. We previously developed a machine learning (ML) algorithm that predicts individual respiratory events during sleep, seconds before they would occur. We hypothesized that by using this algorithm, OA events could be prevented without disturbing quality of sleep. Methodology: A preventative protocol was developed and implemented on a commercial PAP machine as an add-on to the APAP algorithm to intervene when an OA event was predicted by the ML algorithm: (i) pressure ramp up of 3 cmH2O over 2 seconds, (ii) pressure hold for 6 minutes, (iii) ramp down over 2 seconds, and (iv) no intervention for 2 minutes. Polysomnography (PSG) was collected, and each participant underwent a control (ML algorithm off) night and an intervention (ML algorithm on) night, assigned at random. Participants and sleep technologists were blinded to randomization. PAP machine?s auto-tagging of sleep events was compared to sleep technologist scoring. Results: Scores from 11 patients showed a statistically significant (p=0.02) reduction in mean apnea-hypopnea index (AHI) from 6.26 (95% CI, 4.59 to 7.94) in the control group to 4.28 (95% CI, 2.97 to 5.59) in the intervention group. Arousal index and questionnaire assessed sleep quality did not differ between groups. A significant (p < 0.001) discrepancy was found between sleep technologist scoring and PAP auto-tagging. Conclusion: We conclude that ML can be used to reduce AHI and speculate that the use of ML on PAP therapy will lead to personalized therapy. Moreover, our findings support the need for universal measures of OSA severity and effectiveness of PAP therapy.