<p>One of the most common signs of sleep apnea is breathing pauses during the night, and it is difficult to diagnose and cure. This study proposes an innovative learning methodology utilizing an adaptive neuro-fuzzy inference system with an improved cheetah optimization algorithm (ANFIS-ICOA) for detecting sleep apnea severity. For sleep severity apnea detection, the proposed model considers the electroencephalogram (EEG) as an input signal. The suggested method has three stages: pre-processing and decomposition, feature extraction, and sleep apnea severity detection. First, the EEG input signal is pre-processed using the infinite impulse response Butterworth bandpass filter (IIRBB) and Hilbert-Huang transform (HHT) algorithms to pre-process it for the diagnosis of sleep apnea severity. The pre-processed EEG input signal is then split into five frequency sub-bands using wavelet packet decomposition (WPD). Next, the statistical features are derived from each sub-band. Finally, sleep apnea severity detection is done by ANFIS-ICOA with Levy flight updating. This novel combination of approaches is anticipated to yield major advancements in automated diagnosis of sleep disorders. The suggested model's efficacy is assessed using a variety of indicators, such as specificity, f-measure, kappa, area under curve, recall, precision, sensitivity, and accuracy. The proposed method will be implemented in MATLAB. The proposed model has 98.26% detection accuracy for sleep apnea severity. This methodology efficiently addresses the obstacles and improves the precision of sleep apnea severity diagnosis.</p>