A holo-spectral EEG analysis provides an early detection of cognitive decline and predicts the progression to Alzheimer’s disease

AIMS: Our aim was to differentiate patients with mild cognitive impairment (MCI) and Alzheimer’s disease (AD) from cognitively normal (CN) individuals and predict the progression from MCI to AD within a 3-year longitudinal follow-up. A newly developed Holo-Hilbert Spectral Analysis (HHSA) was applied to resting state EEG (rsEEG), and features were extracted and subjected to machine learning algorithms. METHODS: A total of 205 participants were recruited from three hospitals, with CN (n = 51, MMSE > 26), MCI (n = 42, CDR = 0.5, MMSE ≥ 25), AD1 (n = 61, CDR = 1, MMSE < 25), AD2 (n = 35, CDR = 2, MMSE < 16), and AD3 (n = 16, CDR = 3, MMSE < 16). rsEEG was also acquired from all subjects. Seventy-two MCI patients (CDR = 0.5) were longitudinally followed up with two rsEEG recordings within 3 years and further subdivided into an MCI-stable group (MCI-S, n = 36) and an MCI-converted group (MCI-C, n = 36). The HHSA was then applied to the rsEEG data, and features were extracted and subjected to machine-learning algorithms. RESULTS: (a) At the group level analysis, the HHSA contrast of MCI and different stages of AD showed augmented amplitude modulation (AM) power of lower-frequency oscillations (LFO; delta and theta bands) with attenuated AM power of higher-frequency oscillations (HFO; beta and gamma bands) compared with cognitively normal elderly controls. The alpha frequency oscillation showed augmented AM power across MCI to AD1 with a reverse trend at AD2. (b) At the individual level of cross-sectional analysis, implementation of machine learning algorithms discriminated between groups with good sensitivity (Sen) and specificity (Spec) as follows: CN elderly vs. MCI: 0.82 (Sen)/0.80 (Spec), CN vs. AD1: 0.94 (Sen)/0.80 (Spec), CN vs. AD2: 0.93 (Sen)/0.90 (Spec), and CN vs. AD3: 0.75 (Sen)/1.00 (Spec). (c) In the longitudinal MCI follow-up, the initial contrasted HHSA between MCI-S and MCI-C groups showed significantly attenuated AM power of alpha and beta band oscillations. (d) At the individual level analysis of longitudinal MCI groups, deploying machine learning algorithms with the best seven features resulted in a sensitivity of 0.9 by the support vector machine (SVM) classifier, with a specificity of 0.8 yielded by the decision tree classifier. CONCLUSION: Integrating HHSA into EEG signals and machine learning algorithms can differentiate between CN and MCI as well as also predict AD progression at the MCI stage.