Electroencephalographic complexity and cognitive decline: Insights from approximate entropy across aging and dementia

Abstract

INTRODUCTION

Physiological aging involves gradual modifications of brain activity patterns sometimes leading to pathological outcomes. Understanding changes in brain signal complexity from healthy aging to mild cognitive impairment (MCI) and Alzheimer's disease (AD) may provide insights into cognitive decline and support early diagnosis.

METHODS

Electroencephalography (EEG) data were collected from 269 participants divided into three groups: healthy elderly (HE, n = 109), MCI (n = 75), and AD (n = 85). Approximate entropy (ApEn) was computed for the total EEG spectrum and frequency bands. A random forest classifier with a one‐versus‐one strategy and recursive feature elimination (RFE) was used for group discrimination. Additionally, group comparisons and correlations with neuropsychological scores were performed to further investigate the functional relevance of these entropy changes by assessing their association with cognitive and functional performance.

RESULTS

Significant group differences in total ApEn were found, particularly in central brain regions, with progressive decreases from HE to MCI and AD. Frequency‐specific patterns emerged: (1) delta ApEn was increased in those with AD across the scalp; (2) alpha 1 ApEn showed frontal decreases and occipital increases in the AD group, with an intermediate pattern in MCI; (3) gamma ApEn was markedly reduced in those with AD compared to both HE and MCI groups, with central and frontal localization, and extended occipital differences in those with MCI versus AD. The classifier achieved 66.55% accuracy, with sensitivities of 78.90% (HE), 49.33% (MCI), and 65.88% (AD). Positive correlations were found between ApEn and cognitive scores in both the total EEG spectrum and gamma band; conversely, delta and alpha 1 ApEn showed negative correlations with cognitive and functional scores.

DISCUSSION

The distinct entropy patterns observed across frequency bands and scalp/brain regions may help characterize electrophysiological changes associated with cognitive decline and AD.

Highlights

• Total approximate entropy (ApEn) progressively decreases from healthy elderly (HE) to individuals with mild cognitive impairment (MCI) and Alzheimer's disease (AD).

• Patients with AD show increased delta, decreased gamma, and region‐specific alpha 1 changes.

• Electroencephalographic entropy correlates with cognitive and functional performance scores.

• A random forest classifier achieved 66.9% accuracy across HE, MCI, and AD groups.

• Frequency‐specific ApEn may aid early diagnosis and clinical stratification.

1. BACKGROUND

Aging is a multifactorial biological process that profoundly affects brain functional dynamics. ¹ Electroencephalography (EEG) offers a window into brain function, combining non‐invasiveness, safety, and portability. ² , ³ Changes in EEG‐derived metrics provide insights into the neural mechanisms and have been proposed as biomarkers of brain dysfunction. ⁴ , ⁵ , ⁶ Entropy‐based measures—particularly approximate entropy (ApEn)—capture alterations in brain signal complexity, reflecting both physiological aging processes and pathological deviations. ⁷

EEG represents a valuable tool for monitoring disease progression, especially in Alzheimer's disease (AD) and its prodromal stage, mild cognitive impairment (MCI). In this framework, EEG has proven sensitive in capturing disease‐related neural alterations in individuals with MCI and AD. ⁸ , ⁹ Importantly, EEG alterations in aging and dementia are captured by a range of complementary indices, including spectral measures among others. Compared to these approaches, entropy‐based metrics quantify signal irregularity and complexity in the time domain, thereby providing additional and complementary information on brain dynamics in pathological aging. Several studies have reported reduced ApEn in resting‐state EEG of patients with AD, indicating diminished brain signal complexity. ¹⁰ , ¹¹ Most previous studies have focused on AD versus healthy controls, with limited attention to the MCI stage and often relying on global rather than frequency‐specific analyses, typically in relatively small samples. Additionally, complexity measures may offer particular sensitivity to the prodromal phase. Jeong ¹² proposed that reduced non‐linear EEG complexity and impaired information transmission among cortical areas are hallmarks of early AD and MCI. ¹² Neuropsychological assessment also plays a fundamental role, providing essential context for interpreting EEG‐derived measures and helping to correlate electrophysiological alterations with cognitive and functional decline.

In the present study, we recorded resting state EEG from three well age/sex/education‐matched cohorts—healthy elderly (HE), individuals with MCI, and patients with clinically diagnosed mild to moderate AD—and computed ApEn over the total EEG spectrum as well as within delta (2–4 Hz), theta (4–8 Hz), alpha 1 (8–11 Hz), alpha 2 (11–13 Hz), beta 1 (13–20 Hz), beta 2 (20–30 Hz), and gamma (30–45 Hz) bands. All participants completed a neuropsychological assessment, allowing for the integration of cognitive profiles.

The primary objective of this study was to examine entropy modulations in pathological aging. A key feature of our approach was the assessment of ApEn within specific EEG frequency bands. A second aim was to evaluate the functional significance of these modulations by training a multiclass classifier with channel‐level ApEn features and correlating regionally aggregated ApEn values with neuropsychological measures. This strategy enabled us to examine associations between reduced entropy and cognitive impairment, highlighting EEG‐derived entropy as a measure of electrophysiological alterations in pathological aging.

2. METHODS

2.1. Participants

Two hundred sixty‐nine volunteers participated in the study. Specifically, 109 HE (58 females; mean age  =  72.5 ± 1.6 years; mean years of education  =  13.5 ± 0.4), 75 subjects with MCI (51 females; mean age  = 74.4 ± 1.1 years; mean years of education  =  12.0 ± 0.9), and 85 patients with AD (59 females; mean age  =  74.3 ± 1.5 years; mean years of education  =  12.9 ± 0.8) were recruited. The three groups were matched for age, sex, and education. Further information regarding the recruitment criteria is provided in the Supplementary Materials in supporting information.

RESEARCH IN CONTEXT

1. Systematic review: We reviewed electroencephalography (EEG) studies on signal complexity in healthy aging, mild cognitive impairment (MCI), and Alzheimer's disease (AD) dementia, focusing on entropy‐based measures and their relationships with cognitive decline. Prior work generally reports reduced neural complexity in dementia, with heterogeneous findings across frequency bands and scalp regions, and growing use of machine‐learning approaches.

2. Interpretation: Our results demonstrate stage‐dependent, frequency‐specific alterations of EEG entropy across the aging–MCI–dementia continuum. Reduced total and gamma‐band entropy and changes in delta and alpha 1 entropy reflect progressive disruption of cortical network dynamics. They are significantly associated with cognitive and functional impairment, supporting EEG entropy as a clinically relevant biomarker.

3. Future directions: Longitudinal and multimodal studies are needed to determine prognostic utility, clarify underlying mechanisms, and improve early diagnosis and patient stratification.

The study protocol adhered to the Declaration of Helsinki and the World Medical Association's ethical guidelines (1997) and it was approved by the local ethical committee for the IRCCS San Raffaele Roma (Ethics Committee of the San Raffaele Scientific Institut n. RP 22/25); and of the University of Sassari (Sardinia Ethics Committee n. PG/2023/5172, 06/04/2023). All individuals provided written informed consent prior to participation.

2.2. Neuropsychological assessment

All participants underwent a neuropsychological assessment for a characterization of cognitive profile. Further details on the administered tests are provided in the Supplementary Materials.

Scores were adjusted for participants’ age and education levels. The mean (± standard error) scores across the three groups are reported in Table 1 along with demographic features.

TABLE 1.

Mean (± standard error) scores for each demographic, cognitive and functional assessment across the three groups: healthy elderly (HE), mild cognitive impairment (MCI), and Alzheimer's disease (AD).

	Age	Education	MMSE	ACE_R	ADL	A‐IADL‐Q	DSST	ACER_AO	ACER_ME	ACER_FL	ACER_LIN	ACER_VS
HE	72.5±1.6	13.5±0.4	29.4±0.1	92.9±0.5	5.9±0.0	8.0±0.0	42.8±1.0	17.9±0.0	22.4±0.3	11.5±0.2	25.8±0.1	15.1±0.1
MCI	74.4±1.1	12.0±0.9	27.7±0.4	83.1±1.6	5.8±0.2	7.6±0.3	29.3±1.7	17.1±0.2	17.8±0.9	9.9±0.4	24.1±0.5	14.2±0.4
AD	74.3±1.5	12.9±0.8	19.4±1.0	59.5±2.4	4.6±0.3	1.9±0.5	11.9±1.9	13.8±0.7	9.6±0.9	5.3±0.6	20.2±1.2	10.4±0.6
HE vs. MCI	0.589794278	0.074883773	6.550E–44	1.455E–47	1.269E–15	2.181E–60	2.313E–20	5.094E–29	1.623E–30	5.364E–23	1.818E–18	1.334E–20
HE vs. AD	0.153893093	0.57164904	2.336E–11	9.321E–15	3.773E–02	6.546E–05	5.392E–11	1.974E–10	2.434E–09	4.106E–05	3.424E–08	2.519E–03
AD vs. MCI	0.363511039	0.569965212	1.820E–11	4.427E–11	1.701E–03	5.453E–14	1.772E–06	1.295E–06	1.495E–07	3.708E–08	8.202E–04	4.299E–07

Note: MMSE refers to the Mini‐Mental State Examination, a global measure of cognitive functioning. ADL indicates Activities of Daily Living, assessing basic functional abilities, while A‐IADL‐Q is the Amsterdam Instrumental Activities of Daily Living Questionnaire, which evaluates complex daily functioning. DSST corresponds to the Digit Symbol Substitution Test, measuring processing speed and attention. ACER is the Addenbrooke's Cognitive Examination–Revised, a multidomain cognitive screening test, with the following subtests: AO (Attention and Orientation), ME (Memory), FL (Fluency), LIN (Language), and VS (Visuospatial). Pairwise statistical comparisons (HE vs. MCI, HE vs. AD, and MCI vs. AD) are reported in the last three rows of the table for each variable. Bolded p values indicate statistically significant differences.

2.3. EEG acquisition and preprocessing

Continuous EEG was recorded using a 61‐electrode cap (EasyCap GmbH, Brain Products) configured according to the international 10‐20 system, with Fpz as the reference. Further methodological details are provided in the Supplementary Materials.

2.4. Entropy analysis

ApEn was computed for each participant, electrode, and frequency band using a bespoke MATLAB routine. EEG data were band‐pass filtered into seven canonical bands (delta, theta, alpha 1, alpha 2, beta 1, beta 2, and gamma). A detailed description of the methodology used for the computation of ApEn is provided in the Supplementary Materials.

Scalp topographies were generated by interpolating ApEn values to visualize spatial distributions and pathology‐related shifts in EEG complexity.

2.5. Statistical evaluation

Pairwise comparisons were carried out among the three groups for each frequency band. For each electrode and frequency band, two‐tailed unpaired Student t tests were used to evaluate statistical differences among groups. A preliminary threshold for statistical significance was set at p < 0.05. To account for multiple comparisons across electrodes and bands, a false discovery rate (FDR) correction was applied. Only results with p values below the adjusted FDR threshold were considered statistically significant. ¹³

2.6. Machine learning

To classify subjects into the three distinct categories, we used a random forest (RF) classifier. ¹⁴ Given the multiclass nature of the classification task, a one‐versus‐one (OvO) strategy was adopted.

For the classification task, we applied feature selection using recursive feature elimination (RFE) with a RF estimator. ¹⁵

To ensure the robustness of the classification results, we adopted a fully nested 5‐fold cross‐validation strategy. Specifically, a stratified 5‐fold cross‐validation was used as the outer loop to estimate model performance. Within each outer training set, a second stratified 5‐fold cross‐validation (inner loop) was performed, and RFE was applied exclusively within the training data of each inner fold. The final performance of the classifier was then obtained by averaging the results across all folds, thus reducing the risk of overfitting and providing a more reliable estimate of the model's ability to generalize to new data. To further increase the robustness of the feature selection process, we evaluated the frequency with which each feature was selected across folds. Only features selected in at least 40% of the folds (i.e., in at least two out of five folds) were retained. Among the features satisfying this criterion, we focused on those that were also consistent with our previous analyses. Further methodological details are provided in the Supplementary Materials.

2.7. Correlation analysis

A correlation analysis was performed to investigate the relationship between ApEn and neuropsychological test scores across 10 regions of interest (ROIs; frontal, central, parietal, temporal, and occipital in both the left and right hemispheres), considering both the total EEG spectrum and the principal canonical frequency bands. To control for multiple comparisons, p values were corrected using the FDR method. We used a matrix‐based representation: for each frequency band, a correlation matrix was generated using a color scale ranging from blue (–1) to red (1), indicating the direction and strength of the correlation. Only statistically significant correlations were displayed, and the intensity of the color reflected the level of statistical significance. The primary aim was to identify potential patterns of association between EEG complexity and cognitive performance across regions. The analysis was conducted at the ROI level thus substantially reducing the number of statistical tests compared to a channel‐by‐channel approach. Further methodological details are provided in the Supplementary Materials.

3. Results

3.1. ApEn topographical representation

ApEn has been evaluated in three distinct groups through EEG total spectrum and typical frequency bands. Significant differences between groups were calculated and are presented in Figure 1. The new significance thresholds identified by the FDR correction for each comparison and frequency band are reported in Table 2.

FIGURE 1.

Topographical maps display the distribution of approximate entropy (ApEn) values across the scalp in the three groups (healthy elderly [HE], mild cognitive impairment [MCI], and Alzheimer's disease [AD]) for the entire electroencephalography (EEG) spectrum (0.2‐47 Hz); (A) and for each frequency band: delta (2–4 Hz); (B), alpha 1 (8–11 Hz); (C), and Gamma (30–45 Hz); (D). Entropy values are color coded from blue (low entropy/complexity) to red (high entropy/complexity), indicating increasing signal irregularity and complexity. In the central panels of each frequency band section, pairwise group comparisons are shown using false discovery rate (FDR)‐corrected p values (HE vs. MCI, HE vs. AD, MCI vs. AD). The p value maps are color coded from dark red to white, with the blue–orange scale representing statistically significant comparisons after FDR correction (thresholded at the corrected significance level). For each frequency band, the arrangement of the maps follows a matrix structure in which the top row and the lateral column represent the distribution of ApEn values across the scalp in the three groups while each column indicates a different pairwise comparison.

TABLE 2.

Significance thresholds identified using the false discovery rate (FDR) correction for the statistical comparisons among groups: healthy elderly (HE), mild cognitive impairment (MCI), and Alzheimer's disease, (AD) across the different electroencephalography (EEG) frequency bands.

	HE vs. MCI	HE vs. AD	MCI vs. AD
Total EEG Spectrum	0	0.01209	0
Delta	0	0.04518	0.04590
Theta	0	0	0
Alpha 1	0	0	0.00126
Alpha 2	0	0	0
Beta 1	0	0	0
Beta 2	0	0	0
Gamma	0	0.01204	0.02497

Note: Threshold values are expressed as adjusted p values after FDR correction. A value of 0 indicates that no significant threshold was identified for that specific comparison and frequency band.

Statistically significant differences in ApEn values among the three groups were identified in the total EEG spectrum and were visualized using a symmetric matrix representation (Figure 1A). In the comparison between HE and AD groups, the differences predominantly affected the central region and additionally extended to the parietal area. Conversely, no statistically significant differences were found between the MCI and AD groups across any region.

When examining the distinct EEG frequency bands, statistically significant differences emerged specifically within the delta, alpha 1, and gamma bands.

A highly statistically significant result was observed in the delta frequency band (Figure 1B). The difference in entropy distribution in patients with AD was evident across the entire scalp, where an increase in ApEn values was observed compared to the other two groups. However, this increase was slightly less pronounced comparing AD specifically to the MCI group. In contrast, this group exhibited a slight decrease in ApEn in the frontal regions compared to HE, accompanied by a modest increase in the occipital areas. Consequently, statistical significance was reached only at the two electrodes corresponding to these respective areas.

In the alpha 1 band (Figure 1C), differences in ApEn values culminated in a statistically significant effect localized in the centro‐frontal electrodes in the comparison between participants with MCI and those with AD, with higher ApEn values observed in the former group. By contrast, no statistically significant differences were found between HE and those with MCI in this frequency band.

The final observed difference was identified in the gamma frequency band (Figure 1D). Statistically significant differences within this band were evident in the comparisons between HE and AD groups, as well as between MCI and AD groups. Patients with AD demonstrated a reduction in entropy across the entire scalp. These statistically significant differences were primarily localized to electrodes overlying the central and frontal cortices. Conversely, in the MCI versus AD comparison, significant alterations extended beyond the central and frontal regions to include occipital areas.

3.2. Machine learning

For the classification process, a RF model was used, trained using a OvO strategy combined with RFE‐based feature selection to address a three‐class problem. The selected features (n = 18) were: P3_Total, TP10_Total, AF7_Total, Fp1_Delta, Fz_Delta, FC2_Delta, CP1_Delta, FC5_Delta, AF7_Delta, AF8_Delta, Fpz_Delta, P7_Alpha1, CP2_Alpha1, O2_Beta2, Cz_Beta2, FC6_Beta2, FC2_Gamma, FC6_Gamma. The classes were defined as follows: Class 0 comprised HE (n = 109), Class 1 included subjects with MCI (n = 75), and Class 2 consisted of patients with AD (n = 85). The final results showed an accuracy of 66.55% (95% confidence interval [CI]: 63.02–70.08), a sensitivity of 64.71% (95% CI: 60.54–68.88), and a specificity of 82.62% (95% CI: 80.84–84.41). These confidence intervals were computed across the outer folds, allowing us to quantify the variability of the model performance. Specifically, HE (Class 0) were classified with relatively high accuracy, with 78.90% (95% CI: 65.99–91.85) of subjects correctly identified. Misclassifications in this group were limited, with 10.09% (95% CI: 0.73–19.27) incorrectly labeled as MCI and 11.01% (95% CI: 2.23–19.93) as AD. In contrast, the classification performance for subjects with MCI (Class 1) was notably lower. Only 49.33% (95% CI: 36.78–61.89) of individuals with MCI were correctly classified, while a substantial proportion was misclassified as healthy controls (34.67%, 95% CI: 17.70–51.63). A smaller fraction (16.00%, 95% CI: 6.56–25.44) was classified as AD. This pattern indicates a significant overlap between MCI and both adjacent classes, particularly with HE. Patients with AD (Class 2) were identified with moderate‐to‐high accuracy, with 65.88% (95% CI: 53.88–77.88) correctly classified. Misclassifications primarily occurred toward the healthy class (23.53%, 95% CI: 13.20–33.86), while a smaller proportion was labeled as MCI (10.59%, 95% CI: 7.32–13.85). In addition, the confusion matrix aggregated across the outer folds was computed and represented in Figure 2.

FIGURE 2.

Classification performance of a random forest model trained using an One‐vs‐Rest strategy and recursive feature elimination–based feature selection on a three‐class problem: Class 0 = healthy elderly (N = 109), Class 1 = mild cognitive impairment (N = 75), Class 2 = Alzheimer's disease (N = 85). The values are presented as percentages of the total number of subjects in each true class (i.e., rows sum to 100%). Diagonal elements indicate correct classifications (true positives for each class), while off‐diagonal values indicate misclassifications. Color intensity reflects the proportion of correctly or incorrectly classified instances, with darker shades representing higher percentages. CI, confidence interval; CV, coefficient of variation.

3.3. Correlation analysis

Correlational analyses between ApEn values and neuropsychological test scores were conducted within the predefined ROIs for those EEG frequency bands that previously demonstrated statistically significant group differences. The results of the correlation analysis are reported in Figure 3. For the total EEG spectrum ApEn (Figure 3A), significant positive correlations emerged especially between ApEn values in central, parietal, and occipital ROIs and both global cognitive measures (Mini‐Mental State Examination) and functional abilities (Instrumental Activities of Daily Living [IADL], IADL Questionnaire). Conversely, delta ApEn exhibited negative correlations across most ROIs with global cognitive measures (Figure 3B), functional abilities, and several Addenbrooke's Cognitive Examination subdomains, indicating that higher delta entropy is associated with poorer performance. In the alpha 1 band (Figure 3C), negative correlations were observed in more localized area. Gamma ApEn showed in the frontal areas positive associations with global cognition and functional abilities (Figure 3D).

FIGURE 3.

Representation of statistically significant correlations between ApEn and neuropsychological test scores across all participants. Results are shown for the entire EEG spectrum (A), the delta band (B), the alpha 1 band (C), and the gamma band (D). For each statistically significant EEG frequency band and cognitive domain, bar plots display the correlation coefficients observed in predefined bilateral regions of interest, grouped by hemisphere (left and right). The representations of the regions of interest are reported below, in the following order: frontal, central, parietal, temporal, and occipital, first for the left hemisphere and then for the right hemisphere. Positive correlations are shown in red, and negative correlations in blue, with color intensity reflecting the strength of the correlation. All correlations were considered statistically significant after controlling for multiple comparisons using the FDR correction. ACER, Addenbrooke's Cognitive Examination—Revised; ADL, Activities of Daily Living; A‐IADL‐Q, Amsterdam Instrumental Activities of Daily Living Questionnaire; AO, attention and orientation; ApEn, approximate entropy; DSST, Digit Symbol Substitution Test; EEG, electroencephalogram; FDR, false discovery rate; FL, fluency; LIN, language; ME, memory; MMSE, Mini‐Mental State Examination; VS, visuospatial.

Overall, these findings suggest that increased total and gamma and decreased delta and alpha EEG complexity is associated with better cognitive and functional status.

4. Discussion

Physiological aging entails progressive brain modifications that may progress into pathological trajectories, such as AD. The present study examined AD‐related changes in brain complexity. Significant ApEn differences emerged among HE, MCI, and AD groups both across the full spectrum and within delta, alpha 1, and gamma bands. The AD group showed markedly lower total ApEn than the other groups. Region‐specific patterns also emerged: delta ApEn increased diffusely across the scalp in AD, whereas alpha 1 and gamma ApEn were reduced in frontal regions compared to HE and MCI.

These findings align with our previous work ¹⁶ in which we showed that physiological aging is marked by increased ApEn in low‐frequency bands and decreased ApEn in higher frequencies in older compared to younger adults. Several studies ¹⁰ , ¹⁷ , ¹⁸ reported significantly reduced ApEn values in patients with AD across the entire EEG spectrum, particularly in the parietal and occipital areas. These reductions have been linked to neuronal degeneration, deficits in acetylcholine neurotransmission, and disrupted local network connectivity. Additionally, because EEG reflects cortical processing, reduced complexity suggests compromised cortical function, possibly due to network deactivation or diminished adaptability to external stimuli. ¹² Particularly interesting is the absence of significant differences in the comparison between AD and MCI. One possible explanation is that subjects with MCI already exhibit an intermediate level of EEG complexity. This pattern is consistent with the notion that MCI represents a transitional stage along the continuum from normal aging to dementia.

In the delta band, our analysis revealed an increase in ApEn across the scalp in patients with AD compared to both MCI and HE groups. A less pronounced increase in ApEn was observed in subjects with MCI, suggesting that alterations in slow‐wave activity begin early in the disease process and intensify with progression of AD. The topographical distribution of delta changes in those with MCI seems to show a more nuanced pattern: while frontal delta activity tends to be reduced compared to HE, a slight increase can be observed over occipital regions. This spatial dissociation might reflect region‐specific compensatory mechanisms during the early stages of neurodegeneration. The frontal reduction may be associated with early disruption of fronto‐subcortical circuits, which are particularly sensitive to neurodegenerative processes. Therefore, the less pronounced and spatially heterogeneous increase in delta activity in MCI could be interpreted as an intermediate stage between normal aging and the more globally altered slow‐wave dynamics observed in AD. Consistently with our results, Abazid et al., ¹⁹ using spatiotemporal entropy analysis of spontaneous EEG activity, also reported significantly higher entropy values in the lower frequency band of patients with AD compared to healthy controls, supporting the view that alterations in slow oscillations are a robust feature of AD. ¹⁹ In fact, this rise in ApEn suggests a higher degree of irregularity and unpredictability in slow cortical oscillations. Such irregularity may represent a pathological compensatory mechanism, in which the brain—faced with declining efficiency in its neural circuits—generates more variable and less predictable slow‐wave activity in an attempt to maintain function. Alternatively, it could serve as an early marker of disease progression, reflecting instability in low‐frequency network dynamics. ¹⁹

In contrast, within the Alpha 1 band, we observed a significant decrease in ApEn values in the AD group compared to the MCI one. This reduction in complexity in the frontal regions aligns with the progressive loss of neural variability and functional adaptability typically seen in the disease. One possible explanation is that these frontal alterations may be related from dysfunction of the brainstem noradrenergic circuits, particularly the locus coeruleus, observed in subjects with MCI and those with AD, which may contribute to a reduction of thalamocortical modulation, and cortical arousal, thereby affecting neural variability and cognitive functions. ²⁰ Given the role of the alpha rhythm in attention, working memory, and information processing, reduced entropy in this band likely reflects impaired functional integration across brain regions; however, these interpretations require validation with multimodal neuroimaging and neurophysiological evidence. Our results are consistent with those reported by Cao et al., ²¹ who found lower entropy values in the alpha band of patients with AD compared to healthy controls. Such convergence of evidence reinforces the interpretation that alpha slowing and reduced complexity are hallmark indicators of cortical network disintegration in AD. ²¹

We found a reduction in gamma ApEn values, particularly over the frontal regions, in AD participants. Similar findings have been reported by Jungrungrueang et al., ²² who observed reduced Shannon entropy of connectivity in the gamma band of patients with AD, reflecting a persistent pattern of diminished neural complexity in high‐frequency oscillations. These findings show that AD alters brain dynamics in a frequency‐specific way, with distinct changes across slow, intermediate, and fast oscillations, each linked to cognitive decline.

We observed that MCI ApEn values occupy an intermediate position between HE individuals and patients with AD across all examined frequency bands, providing novel evidence for their transitional status.

Our findings across the EEG frequency bands are further supported by the work of Maturana‐Candelas et al. ²³ They found that carriers of protective apolipoprotein E alleles exhibited higher entropy in alpha and gamma bands and lower entropy in delta band, whereas individuals with risk alleles showed the opposite pattern. This entropy profile closely mirrors that observed in HE individuals, who tend to present with lower delta ApEn and higher alpha 1 and gamma ApEn compared to patients with AD. Even though we did not have access to genetic data in our cohort, our findings suggest that entropy alterations may reflect disease‐related network changes and also capture modulatory effects of genetic risk factors on brain dynamics.

Furthermore, we explored the potential discriminative value of ApEn using a machine learning approach. An RF classifier was used to distinguish HE, MCI, and AD subjects, using RFE for feature selection. The RFE‐selected features aligned with regions and frequency bands identified in our ApEn analyses, suggesting that the classification‐relevant features reflect electrophysiological patterns associated with group differences, including delta increases (central/frontal), alpha 1 reductions (parietal), and gamma decreases (frontal). Concerning the model performance, it achieves relatively good sensitivity and specificity for HE and subjects with AD. On the contrary, subjects with MCI were more challenging to classify, with a sensitivity of 49.33%, indicating that many cases were misclassified as either HE or AD. This lower performance likely reflects the inherent heterogeneity and transitional nature of the MCI stage, during which electrophysiological alterations are more subtle across individuals. In fact, rather than representing a clear‐cut category, MCI likely encompasses a continuum of intermediate profiles, ranging from near‐healthy aging to early AD‐like features.

Although entropy‐based classification highlights strengths and limitations in distinguishing disease stages, their functional meaning remains unclear. To assess whether ApEn reflects not only disease staging but also cognitive impairment, we correlated ApEn values from 10 ROIs with neuropsychological scores, focusing on frequency bands showing significant group differences. In bilateral occipital regions, ApEn correlated positively with cognition in the total spectrum and gamma band, and negatively in delta and alpha 1. These patterns align with multi scale entropy findings ²⁴ showing that higher complexity in slow oscillations relates to poorer cognitive performance, whereas greater complexity in fast oscillations reflects preserved or compensatory processing. Delta showed the strongest and most widespread negative correlations, consistent with its association with cortical slowing and cognitive dysfunction, ²⁵ and was also the only band differentiating AD from both MCI and HE across nearly the entire scalp.

While previous studies have reported reduced EEG complexity in AD, the present work provides a more detailed characterization of neural dynamics. Specifically, we examined band‐specific entropy changes and used machine‐learning classification to assess how entropy features distinguish HE, MCI, and AD subjects. We further explored how these electrophysiological changes correspond to cognitive performance. Importantly, these analyses capture aspects of neural dynamics and signal complexity that are not evident from conventional EEG measures. At the same time, both the intrinsic spatial and spectral dependencies of EEG signals and the sensitivity of entropy‐based measures to preprocessing choices, epoch length, and parameter settings should be considered when interpreting the present findings, particularly in the absence of formal robustness analyses. Moreover, the lack of direct comparison to conventional spectral slowing markers prevents assessment of the incremental predictive value of entropy features beyond standard EEG metrics. Future studies should address these limitations through dedicated robustness analyses and comparative frameworks. Overall, these findings suggest that EEG complexity measures may capture relevant aspects of the electrophysiological alterations associated with cognitive decline and AD. However, the current results do not support robust predictive performance to justify claims. The limited sensitivity observed for MCI indicates that cross‐sectional entropy‐based measures alone may not be sufficient for reliable classification at this stage. Larger and longitudinal studies, ideally combining ApEn with multimodal biomarkers, are needed to better clarify its potential contribution.

AUTHOR CONTRIBUTIONS

Alessia Cacciotti: Original idea; writing—original draft; data curation; investigation. Anna Fabbrocino: Writing—original draft; data curation; investigation. Chiara Pappalettera: Writing—original draft; data curation; investigation. Francesca Ginatempo: Writing—reviewing and editing. Nicola Loi: Writing—reviewing and editing. Damiano Sottana: Writing—reviewing and editing. Franca Deriu: Data curation; writing—reviewing and editing. Paolo Maria Rossini: Writing—reviewing and editing. Fabrizio Vecchio: Original idea; investigation; writing—reviewing and editing.

CONFLICT OF INTEREST STATEMENT

None of the authors have potential conflicts of interest to be disclosed. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All individuals provided written informed consent prior to participation.

Supporting information

Supporting Information: trc270284‐Sup‐0001‐SuppMat.docx

Supporting Information: trc270284‐Sup‐0002‐ICMJE.pdf

DATA AVAILABILITY STATEMENT

Data will be available under request to the corresponding author.