A Comprehensive Assessment of EEG Entropy Measures for Bipolar Disorder Diagnosis Using Machine Learning

Bipolar disorder (BD) is a common psychiatric condition with varying severity, characterized by alternating episodes of mania or hypomania and depression, making accurate and timely diagnosis challenging. Electroencephalography (EEG) nonlinear features may serve as valuable biomarkers for psychological disorders, reflecting the brain’s nonlinear dynamics. Although changes in EEG power spectral density are well-documented in BD patients, entropy measures offer a complementary nonlinear approach that is less explored. This study provides a comprehensive evaluation of various entropy measures, including approximate entropy, sample entropy, permutation entropy (PermEn), fuzzy entropy (FuzzyEn), and dispersion entropy (DispEn). Additionally, to capture the complexity of signal fluctuations at different scales, four multiscale entropies are evaluated to assess their effectiveness in distinguishing individuals with BD from healthy controls (HCs). We used resting-state EEG data from both eyes-closed and eyes-open (EO) conditions, collected from 60 individuals: 30 patients with BD (15 females, mean age =37.33, standard deviation (SD) =8.83) and 30 HC participants (13 females, mean age =35.73, SD =9.01). After preprocessing the raw EEG data, five basic and four multiscale entropy measures were calculated separately for the five cortical brain regions. Statistical analyses and machine learning approaches were then applied to assess the differences in entropy measures between the two groups. The results showed that BD patients exhibited higher irregularity in the frontal and central brain regions compared to HC participants across all basic entropy measures. The most significant changes were observed for DispEn and FuzzyEn measures. When considering multiscale entropies, the most significant results were found in the beta frequency band (approximately scale factor 4) in anterior brain regions, particularly for multiscale dispersion entropy (MDE). Furthermore, the best classification performance was achieved during the EO condition for the MDE measure, with an accuracy of 95%, specificity of 93.33%, and sensitivity of 96.66%. The study highlights alterations in entropy measures, particularly MDE in the frontal brain region, suggesting its potential as a sensitive biomarker for bipolar and related disorders.