Obstructive sleep apnea (OSA) is common and has major health implications but treatment options are limited. Interestingly, the severity of OSA is profoundly reduced in deep sleep (called "slow wave sleep"), potentially via an increase in the stimulus required to arouse from sleep. Here the investigators test the idea that the medication called "tiagabine" improves slow wave sleep and reduces OSA severity. The investigators will also test whether tiagabine raises the arousal threshold (more negative esophageal pressure), and whether detailed OSA "phenotyping" characteristics can predict the improvement in OSA severity with this intervention.

The current study tests the primary hypothesis that tiagabine improves sleep apnea severity in patients with moderate-to-severe sleep apnea (apnea hypopnea index measured in supine non-REM sleep; hypopneas defined by 3% desaturation or arousal). The investigators test three secondary hypotheses that tiagabine:

1. increases the proportion of total sleep time in slow wave sleep
2. raises the non-REM arousal threshold (more negative esophageal pressure) via (1).
3. is preferentially effective in patients whose OSA phenotype predicts that an increase in the arousal threshold is sufficient to resolve OSA versus those without such favorable physiology. Favorable physiology is defined here as having a low ventilatory drive at which stable breathing is theoretically feasible ("stable Vdrive" is <100% above eupneic ventilatory drive) due to any combination of a "high" upper airway muscle response, "good" passive anatomy (high Vpassive), and "low" steady-state loop gain (see Owens RL et al SLEEP 2014; Wellman A et al J Appl Physiol 2011, 2013; Eckert DJ et al 2013 AJRCCM).

Inclusion Criteria:

• Diagnosed OSA (moderate-to-severe; apnea hypopnea index >15 events/hr)

Exclusion Criteria:

• History of seizures

18 patients enrolled, 2 did not completed the study, One patient was excluded because of predominantly central sleep apnea on both nights, and another was excluded because of low-quality EEG data. Therefore, 14 subjects were analyzed.