Whale Optimized Temporal Fusion Transformer with Cross-Entropy Ensemble Classifiers for Detecting Sleep Disorders

Sleep disorder diagnosis relies on clinical assessments and Polysomnography (PSG), but complicated by subjective sleep perception. Hence, a Whale Optimized Temporal Fusion Transformer with Cross-Entropy Ensemble Classifiers (WOTFT-CEC) is proposed to classify and detect sleep disorders using EEG, ECG and EMG. The framework integrates Nonlinear Ensemble Recursive Optimization with Adaptive Variance Elimination (NERO-WAVE) algorithm that normalizes data, then selects and ranks the features. Besides, the muscle-twitches arising from proprioceptive feedback collapses multi-regime dynamics of linear features hindering micro-arousals, which is mitigated by a Hierarchical Temporal-Spatial Feature Extraction with Attention and Spectral Fusion (HTSFE-ASF). It validates twitches and micro-arousals, overcoming Surrogate Feature hybrid vigor. Furthermore, lack of Atonia eliminates Tonic-Phasic REM Signature that cluster REM phasic bursts with wake muscle activity, resulting in misdiagnosis. Thus, an Adaptive Multi-Branch Temporal-Spatial Classifier (AM-TSC) uses weight based cross entropy and addresses Sleep State Misperception. Moreover, the comorbid hypertension induces pseudo-transition cues which influence Prediction Orbiting causing oscillatory drifts. To address this, a Predictive Optimization with Temporal Fusion and Bio-Inspired Tuning (POTF-BIT) is incorporated that mitigates pseudo-transition cues hindering Cyclic Predictive Reversal Phenomenon. The model’s robustness is proved by its higher accuracy of 99.50% and specificity of 0.99.