Capturing silent oxidative stress in early Alzheimer’s disease: prediction of CSF biomarkers from sleep qEEG data

Abstract

Oxidative stress is a central pathogenic process in the earliest stages of Alzheimer’s disease (AD), promoting non-enzymatic protein modifications that accumulate in cerebrospinal fluid (CSF) before measurable neurodegeneration. These alterations impair proteostasis and disrupt sleep-regulating neural circuits, producing characteristic changes in sleep electroencephalographic patterns. Because CSF sampling is invasive, quantitative electroencephalography (qEEG) has emerged as a promising non-invasive proxy for early oxidative processes. Here, we investigated whether nonlinear and time-domain sleep qEEG features can estimate CSF oxidative stress biomarkers in early AD using machine learning (ML) models. Forty-two mild-to-moderate AD patients underwent overnight polysomnography, from which sleep qEEG features were extracted. CSF protein oxidation biomarkers—glutamic semialdehyde, aminoadipic semialdehyde, N-carboxyethyl-lysine, N-carboxymethyl-lysine, and N-malondialdehyde-lysine—were quantified by gas chromatography/mass spectrometry, and ML models were trained to predict CSF biomarker levels from qEEG features. The best-performing model was a random forest trained on the first principal component, achieving an $R^2$ of 0.625 and a mean absolute error (MAE) of 467.1 pg/mL. Features derived from frontal and central electrodes during slow-wave sleep and rapid eye movement sleep contributed most strongly to predictive performance. Predictions for healthy controls displayed distributions distinct from those of AD patients, supporting the biological specificity of the qEEG-based estimates. These exploratory analyses suggest that sleep qEEG combined with ML can noninvasively capture silent oxidative processes involved early in AD pathological cascades, with potential for risk stratification, disease monitoring, and non-invasive upstream biomarkers development.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s11357-026-02133-x.

Introduction

Alzheimer’s disease (AD) is the primary cause of progressive neurodegeneration, accounting for the majority of dementia cases. Its projected rise due to population aging poses a critical public health challenge, highlighting the need for early diagnostic strategies and disease-modifying therapies [1, 2]. AD is marked by progressive neuronal loss in cortical and hippocampal regions, associated with extracellular amyloid-beta (A$\beta$) deposits and intracellular Tau aggregates forming neurofibrillary tangles [1].

In the preclinical stages of AD, pathological changes progress silently before cognitive symptoms emerge, creating a neurotoxic environment that disrupts neuronal network integrity [1, 3–8]. Identifying biomarkers that capture early disease alterations is crucial as timely intervention may delay or prevent progression [9, 10].

It is known that oxidative stress—stemming from an imbalance between pro- and antioxidants—is a major contributor to early mechanisms preceding overt AD pathology [4–8]. Oxidative stress triggers excess reactive oxygen species (ROS) production that damages key macromolecules, disrupting cellular homeostasis and fostering amyloid-$\beta$ (A$\beta$) production and tau hyperphosphorylation [7].

Brain proteins are particularly vulnerable to oxidative damage, which results in irreversible structural and functional alterations [11, 12]. This process promotes the accumulation of oxidized species, such as advanced glycation end-products (AGEs), advanced lipoxidation end-products (ALEs), and protein carbonyls—recognized biomarkers of oxidative protein damage in biological samples [12–15].

Quantitative analyses of specific carbonyls, AGEs, and ALEs in the cortical brain tissue of AD patients revealed a notable, though variable, elevation in the levels of amino-adipic semialdehyde (AASA), glutamic semialdehyde (GSA), N$\epsilon$-malondialdehyde-lysine (MDAL), N$\epsilon$-car-boxymethyl-lysine (CML), and N$\epsilon$-carboxyethyl-lysine (CEL) when compared to age-matched controls [16, 17]. Besides, CML levels in the cerebrospinal fluid (CSF) of AD patients are significantly higher compared to those of cognitively healthy individuals [18, 19].

These findings reflect elevated oxidative damage and impaired proteostasis in AD [16, 17, 20], mechanistic pathways underlying synaptic failure and neurodegeneration, which ultimately compromise the integrity of brain regions and neural networks involved in memory and sleep–wake regulation [7, 21–24].

Such pathological changes manifest as altered oscillatory patterns in sleep electroencephalogram (EEG), particularly during slow-wave sleep (SWS) and rapid eye movement (REM) sleep, which are crucial for synaptic homeostasis and cognition [22–25].

Although CSF oxidative stress biomarkers are not yet validated for clinical use, they provide valuable insight into early molecular damage and may serve as complementary indicators of biological vulnerability in individuals at increased risk for AD [26]. Their broader implementation, however, is constrained by the invasiveness of CSF collection and the limited robustness of oxidized protein markers in peripheral biofluids [27–29].

In this context, sleep EEG emerges as a promising, non-invasive, and cost-effective tool for assessing early oxidative stress-related processes. Quantitative EEG (qEEG) captures neurophysiological alterations associated with oxidative damage and excitotoxic cascades [30, 31]. By doing so, it circumvents the limitations of brain tissue sampling [16] and the impracticality of routine CSF analysis, while preserving high specificity for early AD-related changes [19, 26, 29, 32]. When combined with computational approaches, including machine learning (ML), qEEG offers a powerful framework for indirectly inferring physiological profiles and supporting early detection, risk stratification, and monitoring of AD progression.

Quantitative EEG enables automated, objective analysis of brain connectivity and network dynamics. Traditional sleep qEEG analyses, based on spectral power and sleep stage scoring, provide a limited perspective on neurophysiological changes [33]. In contrast, nonlinear dynamics and time-domain qEEG measures offer a more comprehensive characterization of brain network complexity and functional organization [33–35]. These advanced metrics exhibit greater sensitivity, capturing subtle preclinical neural alterations that conventional methods often overlook [34].

We hypothesize that specific nonlinear and time-domain features of sleep EEG act as quantitative markers of neural activity, allowing us to approximate individual profiles of oxidative stress-related biomarkers. These features may capture network vulnerabilities even before the onset of cognitive impairment or overt neuropathology. We propose that EEG-derived metrics can uncover underlying biochemical and physiological patterns characteristic of AD-associated oxidative dysregulation.

Machine learning (ML) techniques are well suited for EEG analysis due to their ability to handle high-dimensional data and capture complex, subtle neural patterns. In the context of AD, ML has shown strong potential to improve the diagnostic accuracy and efficiency of qEEG-based biomarkers [36, 37]. Recent studies have successfully applied ML to sleep qEEG for AD-related tasks [38–41], including the classification of AD patients versus cognitively healthy individuals from sleep EEG recordings [41]. However, ML approaches have rarely been used to estimate CSF biomarker profiles associated with non-enzymatic oxidative damage, particularly during sleep. Addressing this gap may provide critical insights into early oxidative stress-driven neurodegenerative processes.

This study aims to address the limitations of invasive and costly CSF biomarker collection by investigating the feasibility of non-invasive, EEG-based approaches for early oxidative risk stratification (Fig. 1). By identifying electrophysiological correlates of oxidative stress, we seek to support the development of personalized, sleep-based prevention strategies for individuals at risk of AD.

Fig. 1.

Overview of the analytical workflow. Sleep EEG features were extracted from overnight PSG recordings at the subject level, while CSF oxidative stress biomarkers were independently quantified. Exploratory statistical analyses were used to characterize association patterns between EEG features and biomarkers. Supervized ML models were then trained to predict CSF biomarker profiles from EEG-derived features. Dimensionality reduction and model fitting were performed on the training data and applied to held-out test data. Keys: AASA, aminoadipic semialdehyde; CEL, N$\epsilon$-carboxyethyl-lysine; CML, N$\epsilon$-carboxymethyl-lysine; CSF, cerebrospinal fluid; EEG, electroencephalogram; GSA, glutamic semialdehyde; MDAL, N$\epsilon$-malondialdehyde-lysine; ML, machine learning; PSG, polisomnography

To this end, we evaluate the predictive capacity of diverse quantitative sleep EEG features in approximating individual oxidative stress biomarker profiles, including GSA, AASA, CEL, CML, and MDAL, using ML-driven modeling to uncover relevant AD pathophysiological signatures.

Materials and methods

Sample and study design

This is an ancillary study from NCT02814045, a prospective trial designed to evaluate the evolution of AD patients with and without obstructive sleep apnea (OSA) after 1 year of follow-up. The study was conducted in line with the Declaration of Helsinki and was approved by the Human Research Ethics Committee of Hospital Arnau de Villanova (CE-1218). General recruitment and screening procedures are previously described [42]. In short, eligible patients were recruited from the Cognitive Disorders Unit at the Hospital Universitari Santa Maria (Lleida, Spain), between 2014 and 2018.

The subject involved in the analysis met the following criteria: (1) drug naïve; (2) aged over 60 years; (3) diagnosed with AD according to the NIA-AA criteria [43] and (4) mild-to-moderate cognitive impairment, defined by Mini-Mental State Examination (MMSE) scores ranging from 20 to 26, consistent with established clinical staging conventions, with scores of 21–26 indicating mild AD and a score of 20 representing the upper boundary of the moderate range  [44, 45]; (5) having signed informed consent along with the responsible caregiver (and/or legal representative if different from the responsible caregiver). Exclusion criteria for this study were: (1) history of sleep disorders; (2) comorbidities such as cancer, severe depression, severe renal or hepatic insufficiency, and severe cardiac or respiratory failure; (3) history of mental disorders according to DSM-V-TR™criteria; (4) history of untreated (or treated by less than 3 months before the screening visit) vitamin B12 or folate deficiency; (5) history of untreated thyroid disease; (6) current use of medication under investigation or beta-blockers, antidepressants, neuroleptics, and hypnotics use for less than 15 days before the conduction of polysomnography (PSG); (7) Magnetic resonance image (MRI) evidence of hydrocephalus, stroke, space-occupying lesion, or any clinical relevant central nervous system disease besides AD; (8) excessive alcohol intake (>280 gr/week); and (9) visual and/or communication disturbances affecting the compliance with the study procedures.

Eligible patients undergoing overnight PSG, blood, and CSF samples were collected to determine the Apolipoprotein E genotypes (APOE) and the levels of the biomarkers, respectively. Sociodemographic and clinical variables were collected, including age, sex, ethnicity, anthropometric data, clinical history, and sleep-related symptoms.

A total of 144 patients with a recent diagnosis of mild-to-moderate AD were initially recruited, of whom 128 completed overnight PSG assessment [42]. Due to archival and technical limitations affecting legacy sleep recordings, as well as missing CSF biomarker data, a subset of participants was excluded. The final analytical sample comprises 42 participants, whose data were used for all subsequent EEG, biomarker, and ML analyses.

Sample collection

Fasting blood and CSF samples were obtained between 8:00 and 10:00 a.m. to minimize circadian-related fluctuations. CSF was collected in polypropylene tubes and centrifuged at 2000$\times$g for 10 min at 4 $_{}^{\circ }$C to remove cells and other insoluble components. Blood samples were drawn into EDTA tubes, and plasma was separated by centrifuging the blood at 1500 rpm for 20 min. The buffy coat was isolated for DNA extraction and subsequent APOE genotyping. All samples were stored at −80 $_{}^{\circ }$C until further analysis. The samples were provided with the assistance of the IRBLleida Biobank (B.0000682) and Plataforma Biobancos PT17/0015/0027 and were subsequently processed for biomarker analysis.

APOE genotyping

APOE genotype was performed with a Maxwell^® RCS blood DNA kit (Promega, USA) from 2 mL of a peripheral blood sample, using 20 $\mathrm{\mu }$L of DNA. Participants were dichotomized according to APOE genotype as a homozygous or heterozygous carrier (APOE$\epsilon$4+) or non-carrier (APOE$\epsilon$4 -) [46].

AD core biomarker measurement

Commercial kits (Innotest^® $\beta$-Amyloid1-42; Innotest^® hTAU Ag; and Innotest^® Phospho-TAU181P, Fujirebio-Europe, Gent, Belgium) were used to evaluate the main AD biomarkers (A$\beta$42, t-tau, and p-tau). Analyses were performed by board-certified laboratory technicians masked to clinical data in one round of experiments using one batch of reagents. Intra-assay coefficients of variation were lower than 10%for internal quality control samples (two per plate). The amounts of product were expressed as pg/ml. Concentrations of A$\beta$42 below 600 pg/ml, t-tau exceeding 425 pg/ml, and p-tau above 65 pg/ml were classified as positive or indicative of abnormality [47].

Oxidative damage markers measurement

A total of five biomarkers of non-enzymatic protein modifications indicative of oxidative stress were detected and quantified, as previously described [16, 47]. These included markers of protein oxidation—GSA and AASA; CEL and CML; and MDAL. All were identified as trifluoroacetic acid methyl ester (TFAME) derivatives in acid-hydrolyzed, delipidated, and reduced CSF protein samples using gas chromatography-mass spectrometry (GC/MS).

Briefly, protein samples (500 $\mu$g) were delipidated twice by adding methanol:chloroform (2:1, v/v). Following centrifugation, the methanol phase containing proteins was precipitated with cold trichloroacetic acid (TCA) and reduced with 500 mM sodium borohydride in 0.2 M borate buffer (pH 9.2) after overnight incubation at room temperature. The samples were then reprecipitated with TCA, and a mixture of deuterated internal standards (IS) for each marker was added, including [2H8]Lys (d8-Lys), [2H5]5-hydroxy-2-aminovaleric acid (d5-HAVA) as a stable derivative for GSA, [2H4]6-hydroxy-2-aminocaproic acid (d4-HACA) as a stable derivative for AASA, [2H4]CEL (d4-CEL), [2H4]CML (d4-CML), and [2H8]MDAL (d8-MDAL). Proteins were then hydrolyzed with hydrochloric acid (HCl) at 155 $_{}^{\circ }$C and dried under vacuum.

TFAME derivatives were prepared by adding metha-nolic-HCl followed by trifluoroacetic anhydride incubation. The derivatized protein hydrolysates were resuspended in dichloromethane and transferred to autosampler vials for SIM-GC/MS analysis. Reference materials from NIST and quality control samples were included every five samples. Four microliters were injected in splitless mode into an Intuvo 9000 GC system equipped with an HP-5MS UI capillary column and coupled to a 5977 MSD (Agilent Technologies, Barcelona, Spain). The injector temperature was maintained at 250 °C.

The oven temperature program was as follows: 110 $_{}^{\circ }$C for 5 min, ramped to 150 $_{}^{\circ }$C and held for 2 min, then to 240 $_{}^{\circ }$C for 5 min, followed by 300 $_{}^{\circ }$C for 15 min, and held at 300 $_{}^{\circ }$C for an additional 5 min. A 3-min post-run at 110 $_{}^{\circ }$C completed the cycle, yielding a total runtime of 52 min. Data acquisition was performed in selected ion monitoring (SIM) mode across an m/z range of 30–650. The MSD source temperature was set at 250 $_{}^{\circ }$C, and the transfer line temperature at 310 $_{}^{\circ }$C. Helium served as the carrier gas. Detected ion pairs were as follows: Lys and d8-Lys, m/z 180 and 187, respectively; HAVA (GSA) and d5–HAVA (d5-GSA), m/z 280 and 285, respectively; HACA (AASA) and d4–HACA (d4-AASA), m/z 294 and 298, respectively; CEL and d4-CEL, m/z 379 and 383, respectively; CML and d4-CML, m/z 392 and 396, respectively; and MDAL and d8-MDAL, m/z 474 and 482, respectively. The concentration of each marker was expressed as micromoles of GSA, AASA, CEL, CML, or MDAL per mole of lysine.

Polysomnography

Polysomnography (PSG) was performed overnight using the Philips Respironics Alice 6 LDx system. Each recording included two bilateral EOG signals, six EEG channels from the bilateral central (C3-A2, C4-A1), frontal (F3-A2, F4-A1), and occipital (O1-A2, O2-A1) regions, as well as chin EMG, respiratory effort from chest and abdomen, airflow, pulse oximetry, ECG, body position, snoring, and leg movements.

Manual scoring and diagnosis were conducted by expert physicians in accordance with the American Academy of Sleep Medicine (AASM) guidelines, which included annotating artifact segments due to movements or poor sensor contact that hindered physiological signal interpretation [48].

Sleep quality measures included total sleep time (TST), sleep efficiency (SE), and arousal index (AI). TST was calculated by subtracting sleep latency and wakefulness after sleep onset from total recording time. SE was the percentage of TST relative to time in bed (TIB), defined as the period from lights off to lights on. AI indicated the number of arousals per hour of sleep [48].

Additional PSG measures included time spent awake (W), which is defined as the duration and percentage of TIB spent awake. Latency to stage 1 (N1) of non-rapid eye movement (NREM) sleep was measured from lights off until the first N1 episode. Latency to REM sleep was similarly defined. N1, N2, and N3 sleep durations were quantified in minutes and as percentages of TST, with N3 also recognized as SWS. REM sleep duration was similarly measured [48].

For respiratory parameters, apnea was defined as a reduction in breathing amplitude to below 10% of baseline for over 10 s, while hypopnea involved a decrease in airflow by at least 30% for more than 10 s, leading to arousal or oxygen desaturation (a decrease in oxygen saturation greater than 3%). The apnea-hypopnea index (AHI) was calculated as the number of apnea and hypopnea events per hour of sleep. Cumulative sleep time percentage with oxygen saturation (SpO2) < 90% was defined as CT90, and average SpO2 was the mean oxygen saturation during the study. Obstructive sleep apnea (OSA) was defined as an AHI $\ge$ 15 events/hour [48].

Healthy controls database

To validate model predictions and compare AD patient data with healthy individuals, we compiled a database of sleep EEG recordings from three publicly available datasets. All included subjects were free from cardiac, neurological, and psychiatric disorders and had no clinical diagnosis of sleep disorders.

• ISRUC-SLEEP Dataset (Subgroup III): This includes recordings from 10 healthy individuals (9 males, 1 female; mean age $\sim$40 years). EEG signals were recorded from four channels (C3-A2, C4-A1, F4-A1, O2-A1), filtered between 0.3 and 35 Hz, and sampled at 200 Hz [49].

• DREAMS Open Dataset - Healthy (DOD-H): This contains data from 25 participants with no history of recent severe illness. EEG was recorded using a Siesta 802 device with six channels sampled at 256 Hz and filtered from 0.03–35 Hz. For this study, we used four consistent channels: C3-A2, C4-A1, F4-A1, and O2-A1 [50].

• Sleep Disorders Research Center (SDRC) Dataset: This comprises recordings from 11 healthy individuals from Kermanshah, Iran. EEG was recorded at 256 Hz using a SOMNOscreen^TM plus PSG system. From the 14 recorded EEG channels, we selected C3-A2, C4-A1, F4-A1, and O2-A1 for consistency [51].

All EEG recordings used in this study were monopolar, ensuring uniform signal referencing across datasets. Each subject’s sleep stages were scored by clinical experts, and hypnograms were available for all recordings.

Given the slight age differences across datasets, particularly between healthy controls and AD patients, we selected EEG channels from multiple brain regions, including frontal, central, and occipital sites, to minimize bias related to regional age-associated EEG variability. Prior studies have shown that parietal regions may be especially sensitive to age effects in sleep EEG [52], which further supports this electrode selection strategy.

Signal preprocessing

A signal preprocessing phase is needed to homogenize the overall signal information retrieved from the different EEG recordings from AD and healthy patients and mitigate noise effects during the analysis phase, as previously performed [53]. To preprocess and analyze signal data, Python (version 3.10.12) was employed.

The initial dataset consisted of full overnight PSG recordings with sleep stage annotations in accordance with the AASM guidelines, as described above [48]. EEG was identified as the primary signal of interest among PSG recordings. All EEG signals were acquired in a monopolar configuration, referenced to the earlobe or mastoid electrodes, and digitally sampled at an initial frequency ($f_{s0}$) ranging from 200 to 500 Hz. To ensure methodological consistency and comparability with the electrode configurations across cohorts, four EEG derivations were selected for subsequent analyses: C3-A2, C4-A1, F4-A1, and O2-A1.

These centrally located electrodes provide a balanced representation of sleep-related brain activity while limiting muscle artifacts and inter-patient variability, in line with reported age-related EEG changes and inter-hemispheric alterations in AD [54].

The preprocessing phase was broken down into five primary steps:

• Resampling: The EEG signals were resampled for consistency across channels and subjects, addressing variations in sampling rates to an output rate ($f{s}_f$) of 500 Hz following the AASM guidelines [48].

• Filtering: Bandpass filtering was applied to all signals, with cutoff frequencies set according to AASM guidelines [48]. This ensured consistent spectral content within 0.3–35 Hz while preserving essential EEG frequencies. Specifically, frequency cuts removed the DC signal (0 Hz) while retaining Delta (0.5–4 Hz), Theta (4–8 Hz), Alpha (8–13 Hz), and Beta (13–32 Hz) bands. A fifth-order digital Butterworth filter was used for this process, chosen for its “maximally flat” response in the passband, which minimizes ripple effects despite a slower transition band [55].

• Normalization and segmentation: All the signals were normalized to a zero mean and a standard deviation of one, reducing variability across databases and enhancing numerical stability. After this, signals were segmented based on sleep stages annotated by clinicians from hypnograms, resulting in four separate databases representing each sleep stage: N1, N2, N3, and REM.

• Artifact removal: Segments containing artifacts were removed based on annotations provided by experts for each session and their respective timestamps.

• Handling missing values: EEG data preprocessing included handling missing values using Multiple Imputation by Chained Equations (MICE), which iteratively imputes missing entries based on multivariate relationships among variables [56, 57].

EEG feature extraction

EEG signals were analyzed to extract a predefined set of seven features encompassing descriptors and a nonlinear measure, using established signal-processing techniques widely applied in sleep EEG analysis [33, 58–61]. These features capture complementary aspects of EEG signal that are known to reflect sleep stage—specific neurophysiological dynamics and pathological alterations [33, 58–62].

Feature extraction was performed on EEG segments annotated by sleep stage according to AASM criteria [48]. For each stage, features were computed over 2-min windows with 50% overlap and subsequently averaged across all windows, yielding one summary value per feature, per sleep stage, and per EEG channel for the entire overnight recording. This aggregation strategy reduces sensitivity to transient fluctuations while preserving stable, stage-specific signal characteristics relevant for subject-level analysis.

The extracted features comprised the following:

• Time-domain features included mean amplitude, peak-to-peak amplitude, standard deviation, variance, skewness, and maximum value. These descriptors respectively capture signal amplitude, variability, dispersion, and distributional asymmetry, providing interpretable measures of neural synchronization, cortical excitability, and signal irregularity during sleep  [33, 59–61].

• Nonlinear feature, namely Shannon entropy, is used to characterize EEG signal complexity and non-stationarity [33, 61].

The feature set was selected to balance descriptive richness with robustness and interpretability. By restricting the analysis to low-order, largely parameter-free features, dimensionality and redundancy were limited, thereby reducing the risk of overfitting in a limited sample setting while preserving sensitivity to sleep-related EEG dynamics across electrodes and sleep stages [33, 58, 62, 63].

With four sleep stages (N1, N2, N3, and REM), four EEG channels, and seven features per channel and stage, each participant was represented by a single 112-dimensional feature vector. Detailed definitions and mathematical formulations of all extracted EEG features are provided in Appendix 1 and a complete list of all computed EEG features, in Appendix 2 (Table 4). The notation used throughout the manuscript when referring to specific features follows the format: EEG_‘Channel’_‘EEG Feature’_‘Sleep Stage’.

Table 4.

EEG features computed for the EEG F4_A1, C3_A2, C4_A1, O2_A1 channels, segmented by sleep stages. Keys: EEG, electroencephalogram; mean, mean value; max, maximum value; std, standard deviation; ShanEnt, Shannon entropy. The notation used throughout the manuscript when referring to specific features follows the format: EEG_’Channel’_’EEG Feature’_’Sleep Stage’

Sleep stage	Channel	Feature type	EEG features
N1	F4_A1	Time domain	mean, peak-to-peak, std, variance, skewness, max
	O2_A1	Nonlinear	ShanEnt
	C3_A2
	C4_A1
N2	F4_A1	Time domain	mean, peak-to-peak, std, variance, skewness, max
	O2_A1	Nonlinear	ShanEntv
	C3_A2
	C4_A1
N3	F4_A1	Time domain	mean, peak-to-peak, std, variance, skewness, max
	O2_A1	Nonlinear	ShanEnt
	C3_A2
	C4_A1
REM	F4_A1	Time domain	mean, peak-to-peak, std, variance, skewness, max
	O2_A1	Nonlinear	ShanEnt
	C3_A2
	C4_A1

Statistical analysis

Descriptive statistics were computed to characterize the demographic and clinical features of the patient cohort. Depending on the nature of each variable, means, medians, standard deviations, or frequencies were reported. Missing data were handled using listwise deletion. All descriptive statistical analyses were carried out using Python (version 3.10.12).

Machine learning methods for the prediction of CSF oxidative damage biomarkers

Before describing the modeling steps in detail, we summarize the model evaluation procedure. Although leave-one-subject-out and other small-fold cross-validation strategies were considered, an 80/20 train-test split was ultimately adopted to ensure more stable performance estimation given the limited sample size and feature dimensionality.

To prevent information leakage, all steps influencing model training—including exploratory feature screening, dimensionality reduction, and model fitting—were performed exclusively on the training data. All analyses were implemented in Python (version 3.10.12).

Dimensionality reduction: CSF biomarkers

As an initial exploratory step, interrelationships among CSF biomarkers were examined using correlation matrices to assess the degree of shared variance among them. This analysis was particularly relevant given the limited sample size and informed subsequent modeling decisions. To address the strong interdependencies (Fig. 2), we integrated the CSF biomarkers using principal component analysis (PCA), a dimensionality reduction technique that captures shared variance across correlated variables while preserving most of the original information [64]. For this study, a single principal component (1PC) was retained, with each CSF biomarker contributing with a distinct weight, thereby providing a compact representation of the common biomarker signal.

Fig. 2.

Correlation matrix of CSF oxidative damage biomarkers. Heatmap showing pairwise Spearman correlation coefficients between GSA, AASA, CEL, CML, and MDAL across the study cohort. Color intensity reflects the strength of the correlation, with warmer colors indicating stronger positive correlations, whereas cooler colors indicate weaker or negative correlations. The observed intercorrelations among CSF biomarkers suggest shared pathological mechanisms and justify the integration of these biomarkers using PCA to reduce dimensionality and enhance predictive modeling accuracy. Keys: AASA, aminoadipic semialdehyde; CEL, N$\epsilon$-carboxyethyl-lysine; CML, N$\epsilon$-Carboxymethyl-lysine; CSF, cerebrospinal fluid; EEG, electroencephalogram; GSA, glutamic semialdehyde; MDAL, N$\epsilon$-malondialdehyde-lysine; ML, machine learning; PCA, principal component analysis; PSG, polisomnography

Dimensionality reduction: EEG-derived features

Associations between CSF biomarkers (GSA, AASA, CEL, CML, MDAL) and EEG-derived features from C3-A2, C4-A1, F4-A1, and O1-A2 electrodes across N1, N2, N3, and REM sleep stages were assessed using Spearman’s rank correlation. In addition to unadjusted correlations, partial correlations controlling for age, sex, and ApoE4 status were computed. All tests were two-tailed. Nominal p-value thresholds ($p<0.05$) were used as a heuristic criterion to retain EEG features showing systematic associations with multiple biomarkers. These analyses were intended to support dimensionality reduction and qualitative assessment of signal relevance and were not used for statistical inference. This resulting reduced EEG feature set was subsequently used as input to the machine learning models, and the same feature subset was applied unchanged to the held-out test set.

In parallel, the full set of 112 EEG features (i.e., without prior statistical screening) was subjected to PCA to reduce dimensionality and mitigate noise. The PCA transformation was learned exclusively from the training data and applied unchanged to the test data.

The resulting CSF and EEG representations from each processing pipeline were then used as inputs to the predictive models. Model outputs obtained in the reduced space (1PC) were finally mapped back to the original CSF biomarker scale using the inverse PCA transformation, yielding predicted values for each biomarker, as described in the next subsection.

Predictive models

Using the reduced EEG and CSF representations described above, supervized machine learning models were trained to predict CSF oxidative damage biomarkers from sleep EEG features. Two complementary modeling approaches were considered: random forest (RF) and Gaussian process (GP) models. These algorithms were selected for their ability to capture nonlinear relationships and their suitability for small-sample, high-dimensional settings.

RF models are supervized ensemble learning methods that combine multiple decision trees trained independently to create a robust predictive framework [65]. RF models were used for both regression and classification tasks. In the regression setting, RF was trained to predict either the 1PC summarizing the CSF biomarkers or the individual biomarkers directly. Model performance was evaluated using alternative EEG input representations, as summarized in Table 1, enabling direct comparison between feature-selection-based and PCA-based pipelines without introducing additional modeling complexity.

Table 1.

Summary of machine learning models, tasks, input representations, and evaluation metrics. Keys: Class., classification; Reg., regression

Model	Task	Input	Target	Algorithm	Evaluation metrics
M1	Reg.	Selected EEG features	CSF 1PC	RF	$R^2$, RMSE, MAE, MedAE, MAPE
M2	Reg.	All EEG features (PCA$_X^{}$)	CSF 1PC	RF	$R^2$, RMSE, MAE, MedAE, MAPE
M3	Reg.	Selected EEG features	Individual bioms	RF	$R^2$, RMSE, MAE, MedAE, MAPE
M4	Reg.	Selected EEG features	CSF 1PC	GP (RBF + ARD)	$R^2$, RMSE, MAE, MedAE, MAPE + uncert.
M5	Reg.	Selected EEG features	Individual bioms	GP (RBF + ARD)	$R^2$, RMSE, MAE, MedAE, MAPE + uncert.
M6	Class.	Selected EEG features	AD vs. HC	RF	Accuracy, ROC–AUC

GP regression models a probability distribution over functions that fit the observed data, incorporating predictive uncertainty through a covariance (kernel) function [66]. GP regression models were trained having the EEG selected features as input to predict the same CSF targets as the RF models, allowing direct performance comparison across algorithms. In addition to modeling nonlinear relationships, GP provides probabilistic predictions with associated uncertainty estimates, offering complementary information regarding prediction confidence.

To support post hoc interpretation of the trained models, we computed feature importance measures using algorithm-specific approaches. For RF models, feature importance was estimated using the mean decrease in impurity (MDI) method [67]. At each decision tree split, the reduction in impurity (measured as the decrease in residual variance) is attributed to the splitting feature. The final importance score of each feature is computed as the sum of these reductions across all trees, with higher scores indicating features more frequently used in informative splits [67]. For GP models, automatic relevance determination (ARD) was implemented within the radial basis function (RBF) kernel [68, 69]. This method assigns an individual length-scale parameter to each input feature. Feature relevance was inferred as the inverse of the learned length scale, with shorter length scales indicating greater predictive influence [68, 69]. All feature importance values were normalized to facilitate comparison across models.

In addition to regression, RF models were also applied to a binary classification task. By incorporating EEG data from healthy individuals into the dataset, we assessed whether the same EEG features used for biomarker prediction could reliably distinguish between AD patients and healthy controls. As discussed in “Strengths and limitations,” this analysis was intended as an exploratory sanity check of disease-related signal differences rather than as a validation of biomarker prediction accuracy in the absence of ground-truth CSF measurements in healthy controls.

Table 1 summarizes all machine learning methods explored in this study. This overview clarifies the distinction between the primary feature-screening-based pipeline and the complementary validation analyses.

Model evaluation and performance metrics

Model performance was evaluated using five complementary metrics: R-squared (R²), root mean squared error (RMSE, expressed in pg/ml), median absolute error (MedAE, pg/ml), mean absolute error (MAE, pg/ml), and mean absolute percentage error (MAPE). All metrics were computed exclusively on the held-out test set, corresponding to 20% of the data, ensuring unbiased performance assessment.

Together, these metrics provide a comprehensive evaluation of model fit and predictive accuracy, capturing both absolute deviations in biomarker concentration (pg/ml) and relative percentage errors. These combinations allow performance to be assessed across the full range of CSF biomarker values.

R² quantifies how well the regression model explains the variance in the dependent variable, indicating the proportion of variability accounted for by the predictor variables [70]. Its values range from 0 to 1, with higher values indicating better model performance. RMSE measures the square root of the average squared difference between predicted and actual values, providing a sensitive indicator of large errors. MedAE reflects the median of absolute errors, offering robustness to outliers. MAPE expresses prediction error as a percentage, making it interpretable across scales, while MAE represents the mean of the absolute errors, giving a linear view of prediction deviations. Lower values of RMSE, MedAE, MAPE, and MAE indicate more accurate predictions [71].

Results

Characteristics of the cohorts

Table 2 summarizes the sample characteristics, including socio-demographic information, comorbidities, cognitive assessments, CFS biomarker levels, APOE genotyping, self-reported drowsiness, and key polysomnographic parameters.

Table 2.

Characteristics of patients with mild-to-moderate Alzheimer’s disease (AD). Keys: AHI, apnea-hypopnea index; AI, arousal index; $A\beta 42$, amyloid-beta protein; AASA, aminoadipic semialdehyde; CEL, N$\epsilon$-carboxyethyl-lysine; CML, N$\epsilon$-Carboxymethyl-lysine; BMI, body mass index; ESS: Epworth sleepiness scale; IQR, interquartile range; $p-tau$, phospho-tau protein; SD, standard deviation of mean; $t-tau$, total-tau protein; GSA, glutamic semialdehyde; MDAL, N$\epsilon$-malondialdehyde-lysine; MMSE, Mini-Mental State Examination; TIB, time in bed; $t-tau$, total-tau protein; TST, total sleep time; T90, cumulative sleep time percentage with oxyhemoglobin saturation (SpO2) < 90%

Data	Number (%) or mean [SD]
Sociodemographic data
Women	21 (50%)
Age, years	74.71 [69.32; 80.11]
BMI, kg/m2	27.96 [22.61; 33.30]
Comorbidities
Hypertension	24 (57.14%)
Diabetes mellitus	7 (16.67%)
Dyslipidemia	21 (50%)
Heart diseases	5 (11.90%)
Stroke	2 (4.76%)
Cognition
MMSE	23.31 [20.96; 25.66]
Sleep-related symptoms
ESS	5.98 [2.27; 9.68]
Snoring	1.36 [0.15; 2.56]
Not restorative sleep	0.17 [−0.32; 0.66]
Witnessed apneas	0.19 [−0.32; 0.70]
Insomnia	0.33 [−0.24; 0.91]
Polysomnographic parameters
TIB, minutes	412.98 [379.02; 446.94]
TST, minutes	262.28 [170.2; 354.36]
Sleep efficiency, %	63.67 [41.53; 85.8]
N1 sleep stage, %	13.27 [5.40; 21.13]
N2 sleep stage, %	25.96 [10.63; 41.29]
N3 sleep stage, %	16.30 [4.32; 28.28]
REM sleep, %	8.05 [1.05; 15.05]
Latency to N1, minutes	36.95 [−12.45; 86.35]
Latency to REM sleep, minutes	156.34 [58.04; 254.65]
AHI events/hour	21.90 [5.30; 38.49]
AI events/hour	38.91 [19.38; 58.44]
Average SpO2, %	92.74 [90.39; 95.09]
T90%	7.95 [−5.56; 21.46]
ODI events/hour	28.89 [5.96; 51.83]
AD core biomarkers
A$\beta$42, pg/ml	510.64 [376.79; 644.50]
t-Tau, pg/ml	550.55 [281.51; 819.59]
p-Tau, pg/ml	81.27 [48.57; 113.96]
Oxidative damage biomarkers
GSA, $\mu$mol/mol of lysine	6556.15 [3884.08; 9228.23]
AASA, $\mu$mol/mol of lysine	88.58 [50.61; 126.54]
CEL, $\mu$mol/mol of lysine	116.07 [62.44; 169.71]
CML, $\mu$mol/mol of lysine	1126.60 [813.04; 1440.17]
MDAL, $\mu$mol/mol of lysine	210.68 [−21.64; 443.00]

Correlations between EEG features and CSF biomarkers

Prior to examining EEG-CSF associations, descriptive statistics of the extracted EEG features are provided in the Supplementary Material (Table S1), which reports the average values of the 112 EEG features across EEG channels and sleep stages.

We examined the relationships between qEEG-derived features and CSF biomarkers using Spearman correlation analysis, as described in “Machine learning methods for the prediction of CSF oxidative damage biomarkers”. Figure 3 summarizes the statistically significant associations before and after adjustment for potential confounding factors (age, sex, and ApoE4 status).

Fig. 3.

Heatmap of Spearman correlations between sleep EEG features and cerebrospinal fluid biomarkers according to sleep stages for each one of the electrodes used. The represented values are unadjusted (Rho) or adjusted for age, sex, and ApoE4 status (Rho adjusted). Given the large number of feature–biomarker pairs examined, this analysis is exploratory and intended to highlight overall association patterns that informed subsequent machine learning modeling. Warmer colors indicate stronger positive correlations, while cooler colors represent weaker or negative correlations. Asterisks (*) highlight statistically significant correlations ($p<$0.05). Keys: AASA, aminoadipic semialdehyde; CEL, N$\epsilon$-carboxyethyl-lysine; CML, N$\epsilon$-carboxymethyl-lysine; GSA, glutamic semialdehyde; MDAL, N$\epsilon$-malondialdehyde-lysine; LempelZiv, Lempel-Ziv; max, maximum value; SampEnt, sample entropy; ShanEnt, Shannon entropy

Correlation analysis among CSF biomarkers revealed strong interdependencies (Fig. 2), indicating substantial shared variance across markers. Consistent with this structure, a single principal component (1PC) captured 54.89% of the total variance (Figure S1, Supplementary Material), providing a compact representation of the common biomarker signal.

Exploratory correlations between EEG features and CSF biomarkers revealed structured association patterns across sleep stages and electrode locations. In particular, EEG features derived from NREM sleep—especially N2 and N3—and from frontal and central electrodes consistently showed associations with multiple oxidative stress biomarkers. These patterns suggest that sleep-stage- and region-specific EEG features capture physiologically meaningful variability related to CSF biomarker levels.

For transparency and interpretability, we report below the EEG features that exhibited the most consistent associations across biomarkers and conditions:

• Central region (C3, C4): maximum amplitude and peak-to-peak amplitude during N2 and N3 sleep stages; variance, standard deviation, and skewness during N1 and N2; and Shannon entropy during REM sleep

• Frontal region (F4): peak-to-peak amplitude during N2 and N3; standard deviation during N1 and N3; variance during N3; and Shannon entropy during REM sleep

• Occipital region (O2): peak-to-peak amplitude, standard deviation, and variance during N3

These descriptive correlations complement the predictive analyses presented in the following sections by providing an initial qualitative assessment of EEG–biomarker relationships. From a ML perspective, the presence of structured association patterns suggests that EEG features encode relevant information related to oxidative damage and motivated their use for feature selection in supervized modeling. Importantly, these analyses were exploratory and were not used for statistical inference or hypothesis testing.

CSF prediction

Model performance results are summarized in Table 3. Among the evaluated approaches, the random forest (RF) model trained to predict a single principal component (1PC) of CSF biomarkers achieved the best overall performance, with a mean test-set $R^2$ of 0.625. This indicates the strongest model fit and consistently superior results across all error-based metrics.

Table 3.

Comparison of performance metrics for CSF biomarker prediction across different models. This reports the mean test set values for five metrics: $R^2$ score, RMSE, MedAE, MAPE, and MAE. Results are shown for four models: RF with separate predictions for each biomarker, RF with 1PC, GP with a separate prediction, and GP with 1PC. Higher $R^2$ values indicate better model fit, while lower RMSE, MedAE, MAPE, and MAE values indicate greater accuracy. In bold, the best-performing results are highlighted. Keys: RMSE, root mean squared error; $R^2$, R-squared; MeadAE, median absolute error; MAPE, mean absolute percentage error; MAE, mean absolute error

	RF (Bioms)	RF (1PC)	GP (Bioms)	GP (1PC)
	M1	M3	M4	M5
$R^2$	0.616	0.625	0.529	0.561
RMSE (pg/ml)	960.29	896.94	1141.37	1100.62
MedAE (pg/ml)	94.48	92.70	110.06	110.37
MAPE (%)	43.47	43.35	53.87	54.39
MAE (pg/ml)	496.34	467.15	586.32	568.94

The RF model trained to predict individual biomarkers separately showed comparable performance, particularly for RMSE (960.29 pg/ml) and MAE (496.34 pg/ml), but exhibited slightly reduced consistency across biomarkers compared to the 1PC-based approach.

Performance metrics are reported as averages across test-set patients. Individual prediction plots (Fig. 4) reveal variability in accuracy across subjects, reflecting inter-patient differences. While the models do not provide exact biomarker point estimates, they capture stable and reproducible trends in CSF biomarker variability.

Fig. 4.

$R^2$ scores for test predictions across all biomarkers obtained with the random forest model trained to predict the CSF principal component (1PC). Results are shown per patient on a logarithmic scale to enhance visualization. Predicted versus true values are displayed for cerebrospinal fluid (CSF) oxidative stress biomarkers. Keys: AASA, aminoadipic semialdehyde; CEL, N$\epsilon$-carboxyethyl-lysine; CML, N$\epsilon$-carboxymethyl-lysine; GSA, glutamic semialdehyde; MDAL, N$\epsilon$-malondialdehyde-lysine; N3, N3 sleep stage; max, maximum value; $R^2$, R-squared; ShanEnt, Shannon entropy

A complementary pipeline was implemented in which the full set of EEG features was projected onto 12 principal components, capturing 87.26% of the total EEG variance, and used to train an alternative RF model together with the 1PC biomarker component. This PCA-based model achieved comparable overall performance (Table S2, Supplementary Material), but exhibited slightly reduced robustness across biomarkers, suggesting that models trained on EEG features retained after exploratory screening provide more stable predictions in this dataset.

Post hoc analysis of feature importance revealed EEG features that consistently contributed to prediction across both RF and GP models. Figure 6 highlights the eight most informative features from the top fifteen ranked predictors, identified based on a clear drop in importance values across models. The most relevant features include the following:

• $O2\_A1\_ptp\_N3$ (peak-to-peak amplitude, occipital region, N3 sleep)—important in all models

• $C4\_A1\_ptp\_N2$ (peak-to-peak amplitude, central region, N2 sleep)—top-ranked across all models

• $F4\_A1\_shannon\_entropy\_REM$ (Shannon en-tropy, frontal region, REM)—consistently informative all models

• $C4\_A1\_skewness\_N2/N1$ (skewness, central re-gion, N1–N2)—significant in both RF and GP 1PC settings

• $C3\_A2\_var\_N1$ and $C3\_A2\_std\_N1$ (variance and standard deviation, central region, N1 sleep)—dominant in GP models

Across individual models, a large fraction of top-ranked predictors originated from slow-wave sleep (SWS), highlighting this stage as the most informative and consistently contributing the greatest number of relevant EEG features.

Fig. 6.

Feature importance scores for the prediction of cerebrospinal fluid oxidative stress biomarkers using sleep qEEG-derived features. a and b Results from random forest models (RF) trained to predict individual biomarkers (a) or their first principal component (1PC) (b). c and d Corresponding results for Gaussian process (GP) models. The eight most relevant features for each model are highlighted. The most predictive features were predominantly derived from non-REM sleep stages (N1–N3), particularly from central and occipital EEG channels. Key features included peak-to-peak amplitude, variance, skewness, and Shannon entropy, underscoring the contribution of time-domain and nonlinear EEG characteristics to the modeling of oxidative stress-related alterations in early Alzheimer’s disease

Finally, GP models provided uncertainty-aware predictions; however, prediction uncertainty increased at higher biomarker values, resulting in large error margins, as seen in Fig. 5. This behavior likely reflects the limited sample size and the scarcity of patients with extreme biomarker levels, particularly in more advanced stages of Alzheimer’s disease.

Fig. 5.

$R^2$ scores for test predictions across all biomarkers obtained with the Gaussian process model trained to predict the CSF principal component (1PC). Results are shown per patient on a logarithmic scale to enhance visualization. Predicted versus true values are displayed for cerebrospinal fluid (CSF) oxidative stress biomarkers, together with uncertainty estimates derived from the probabilistic model. Keys: AASA, aminoadipic semialdehyde; CEL, N$\epsilon$-carboxyethyl-lysine; CML, N$\epsilon$-carboxymethyl-lysine; GSA, glutamic semialdehyde; MDAL, N$\epsilon$-malondialdehyde-lysine; N3, N3 sleep stage; max, maximum value; $R^2$, R-squared; ShanEnt, Shannon entropy

Validation with healthy controls

To further assess the robustness and disease-specificity of the proposed regression models, we conducted additional analyses using a control cohort of cognitively healthy controls (HCs). Importantly, CSF biomarker measurements were not available for this group due to ethical constraints, and therefore, HC data were not used during model training. Accordingly, analyses involving healthy controls are presented as exploratory and qualitative, and are not interpreted as direct estimates of CSF biomarker levels in the absence of ground-truth measurements, as further discussed in “Strengths and limitations”.

Nevertheless, EEG features extracted from the HC sleep recordings were provided as input to the models trained exclusively on AD data in order to evaluate generalization behavior and disease specificity. The predicted biomarker values for healthy controls were consistently distinct from both the true and predicted values observed in AD patients, typically falling outside the AD biomarker distribution. While the model does not fully capture the extremes of the AD biomarker spectrum—likely due to limited size and range of the training data (Figs. 4, 5)—they nonetheless produce structured and biologically plausible predictions that differentiate between groups.

To further assess whether EEG features encode disease-related information independently of CSF bio-marker availability, we trained a random forest classifier to distinguish AD patients from healthy controls using the same EEG inputs. As reported in Figure S2 (Supplementary Material), the classifier achieved strong performance ($AUC=0.99$, $Accuracy=0.91$), reinforcing the idea that the selected EEG features reflect pathophysiologically meaningful alteration rather than spurious dataset-specific effects.

Discussion

In this study, we developed advanced ML models to evaluate the predictive capacity of nonlinear and time-domain features derived from sleep EEG for estimating CSF protein oxidation markers in early-stage AD. Our results show that peak-to-peak amplitude, variance, skewness, and standard deviation during NREM sleep, as well as Shannon entropy during REM sleep, particularly in central and frontal brain regions, are strong predictors of non-enzymatic oxidative stress biomarkers. These biomarkers, including protein oxidation, AGEs, and ALEs, are well-established indicators of irreversible oxidative damage in the AD brain.

Integration of functional and molecular markers

Our results introduce a novel, clinically relevant framework that integrates functional neurophysiological data (sleep EEG) with molecular biomarkers, bridging two fundamental aspects of AD pathology: network dysfunction and oxidative molecular damage [7, 24]. By leveraging non-invasive sleep EEG alongside CSF biomarker analysis, we enhance the biological specificity of qEEG as a tool for early detection.

Recent studies indicate that nonlinear EEG metrics detect AD-related brain alterations that traditional frequency-based analyses often miss [35, 72]. Features such as Shannon entropy capture declining brain signal complexity, while variance and skewness reflect regional connectivity disruptions, characteristics of AD [73–75]. However, limited research has explored these markers specifically in sleep EEG in AD, a gap this study addresses.

Bidirectional relationship between sleep and oxidative stress

Our findings, indicating that SWS and REM sleep yield the most informative qEEG signals, reinforce prior evidence of a bidirectional interplay between sleep disturbances and oxidative stress [24]. Disruptions in SWS and REM architecture not only reflect the impact of oxidative damage on thalamocortical and corticocortical circuits involved in sleep regulation, but also drive further reactive oxygen species production [76–78].

SWS plays a key role in cellular repair, mitochondrial function, and antioxidant defence [76]. Reduced SWS duration or fragmentation is associated with elevated CSF oxidative stress markers and accelerated AD progression [77]. Similarly, the glymphatic system, which clears neurotoxic waste such as A$\beta$, operates most effectively during NREM sleep, particularly SWS [79, 80]. Disruptions to A$\beta$ clearance from SWS impairment or hypoxia (e.g., sleep apnea) may promote A$\beta$ accumulation and neuronal vulnerability [81].

REM sleep also plays a critical role in neurotoxin clearance and antioxidant regulation. Its deprivation increases oxidative stress, impairs cognitive function, and depletes antioxidants, linking reduced REM sleep to lower CSF A$\beta$ levels and cholinergic neurons degeneration in AD [76, 78, 82, 83].

Our findings align with region-specific oxidative damage in AD, showing that qEEG features predictive of oxidative stress markers are most prominent in the frontal brain during SWS and REM sleep [25, 84, 85].

Methodological considerations in the use of supervized learning models

The application of supervized learning models such as RF and GP offers complementary advantages and methodological challenges in the context of biomarker prediction from EEG data. One key strength of GP modeling is its ability to estimate predictive uncertainty, which allows the model to express confidence levels in its output [86]. However, the observed increase in error margins, particularly at higher biomarker levels, raises concerns regarding the model’s reliability in extreme cases. This limitation is likely influenced by the relatively small sample size and the underrepresentation of patients with advanced AD in the training data.

Although RF models do not natively provide uncertainty estimates [87], the parallel observation of increa-sed prediction error at higher biomarker values suggests that both approaches are affected by a common challenge: accurately modeling the distributional extremes of CSF biomarkers. This pattern has important implications for clinical translation, where extreme values may be the most diagnostically informative yet also the most difficult to predict.

In this context, the magnitude of the prediction deviations observed indicates that, although the models capture meaningful associations between qEEG features and CSF biomarker variability, their predictive precision remains insufficient for clinical decision-making. This highlights the importance of expanding the dataset—particularly with individuals spanning the higher biomarker ranges—and refining the modeling strategies to improve accuracy across the full distribution.

To assess whether the learned EEG representations nevertheless capture disease-relevant information beyond pointwise biomarker prediction, we further explored model generalization and disease specificity using a control cohort of cognitively healthy individuals. These analyses, while not contributing to training, demonstrated the capacity of the EEG-based models to differentiate between pathological and non-pathological signal patterns, even in the absence of true biomarker values. This suggests that the EEG features extracted retain disease-related information and that the supervized models are able to exploit these signals meaningfully.

Strengths and limitations

Our study has several key strengths. First, we analyzed a rigorously defined cohort of mild-to-moderate AD patients, ensuring a well-characterized population. Second, we utilized a comprehensive assessment battery, including overnight PSG and CSF biomarker analysis, a combination rarely used in AD research. Third, we included only drug-free participants to eliminate potential medication-induced distortions in REM sleep EEG features [88].

A key strength of our approach is the use of machine learning models to explore complex, potentially nonlinear relationships between qEEG features and CSF oxidative stress biomarkers [89, 90]; however, this strength must be interpreted in light of the methodological constraints of the study. Unlike prior AD studies that primarily relied on linear EEG metrics [76, 91], our work incorporates nonlinear and time-domain EEG features, offering a deeper understanding of oxidative stress-related neurophysiological alterations.

Although additional physiological signals (e.g., electrooculography, respiration) and clinical or demographic variables were available in the dataset, these were not included as regressors in the present analysis. This choice was motivated by the aim of establishing a clear proof of concept based on a single, widely available, and non-invasive modality, as well as by the need to limit model complexity given the relatively small sample size. Future studies integrating complementary physiological modalities and additional covariates may further improve predictive performance and enable a more comprehensive characterization of early AD-related alterations.

Despite these strengths, some limitations must be acknowledged. The primary constraint is the small dataset from a single center, which may limit generalizability. However, our use of advanced ML models helped mitigate this issue by enhancing predictive performance. While a low number of patients with high CSF biomarker levels contributed to increased prediction error, especially at extreme biomarker values, both RF and GP models reliably captured trends in the data, consistently identifying key EEG features (see Fig. 6). Furthermore, GP modeling provided uncertainty estimates that confirm stronger confidence in lower-to-mid range predictions, which could improve with larger datasets.

Other limitations include the use of only four EEG channels, which restricts detailed topographic analysis and prevents a fine-grained mapping of regional oxidative stress–related changes. At the same time, this low-density montage can also be viewed as a pragmatic advantage from a translational perspective. Converging evidence from wearable and reduced-montage EEG studies indicates that two-channel frontal/prefrontal recordings, combined with machine learning models, are sufficient to detect dementia-related slowing, peak-frequency shifts, and connectivity alterations and to classify patients with dementia or MCI versus cognitively normal individuals with high accuracy, sensitivity, and specificity [92, 93].

In line with these findings, our results suggest that REM sleep qEEG markers of CSF oxidative stress can be captured even at a coarse spatial resolution. This supports the idea that early oxidative stress–driven network alterations, which are thought to precede clinically evident amyloid deposition and neurodegeneration, may be accessible using simplified, minimally obtrusive EEG configurations. Thus, while higher-density EEG will be important in future work to refine regional interpretation and potentially improve predictive performance, the present low-channel approach is consistent with the development of scalable, low-burden devices for early risk stratification and monitoring along the AD continuum. In addition, the reduced channel configuration lowers model complexity, which is particularly advantageous given the limited sample size.

At the same time, the application of the trained models to healthy control data should be interpreted with caution, as ground-truth CSF biomarker measurements are not available for this group due to ethical constraints. Consequently, these analyses do not constitute a validation of biomarker prediction accuracy in healthy individuals.

To mitigate potential confounding effects related to data origin and acquisition differences, we applied a comprehensive and standardized preprocessing pipeline across all participants and datasets. This approach was intended to ensure that observed differences between AD patients and healthy controls more likely reflect underlying physiological variability rather than database- or acquisition-specific biases [41]. Nevertheless, residual cross-dataset heterogeneity cannot be fully excluded.

Within these constraints, the divergence observed in predicted biomarker distributions for healthy controls and the results of the classification analysis are intended as a qualitative sanity check of the learned EEG–biomarker relationships, evaluating whether the models capture systematic and non-degenerate EEG-derived variations when applied beyond the AD cohort. The presence of structured differences between AD patients and healthy controls supports sensitivity to disease-related EEG alterations, whereas their absence would have raised concerns regarding model validity.

Finally, while non-enzymatic protein biomarkers show promise for early AD risk stratification and monitoring of disease progression, their clinical validation remains ongoing [47].

Clinical and research implications

This integrative qEEG-CSF biomarker strategy could pioneer a paradigm shift in AD risk stratification and early diagnosis. By demonstrating that non-invasive sleep EEG features could estimate key oxidative stress biomarkers, we propose a cost-effective alternative to invasive CSF sampling [94].

Moving away from invasive molecular markers toward a non-invasive, sleep-based monitoring system could enable clinicians to identify individuals at risk for AD—where early intervention is critical [9, 10]—as well as track disease progression in AD patients with heightened oxidative stress-related network damage.

This approach could be particularly valuable for large-scale screening of asymptomatic individuals, allowing for early intervention before clinical symptoms emerge. It can be seamlessly integrated into longitudinal monitoring protocols, complementing traditional biomarkers and neuroimaging methods [10].

Additionally, sleep is a modifiable risk factor for oxidative stress, making this strategy highly relevant for personalized prevention. Targeted lifestyle and sleep interventions—including cognitive behavioral therapy for insomnia (CBT-I), transcranial stimulation, or melatonin therapy—could improve SWS and REM sleep, reduce oxidative damage, and potentially delay or prevent symptom onset [6, 25, 95].

Furthermore, EEG-derived features could serve as biomarkers to assess treatment efficacy, offering real-time insights into the effects of oxidative stress-targeting therapies on neural activity [6]. Lastly, our methodology may propel research efforts toward the validation of CSF oxidatively modified protein biomarkers for clinical use without requiring invasive sampling procedures.

Future directions

Future research should validate these findings in larger, diverse cohorts, including preclinical AD cases and healthy older adults, to distinguish AD pathology from normal aging [10]. Longitudinal studies are crucial for evaluating the stability and predictive value of qEEG biomarkers over time, as well as their role in disease progression and treatment response. Factors such as OSA and the ApoE genotype, which can exacerbate oxidative damage and potentially affect the accuracy of qEEG-based biomarkers, should be considered [23, 26, 96].

Future studies should explore how qEEG-based oxidative stress markers correlate with established AD biomarkers ($A\beta$, p-tau, t-tau) to refine risk models. High-density EEG should also be incorporated to improve spatial resolution, enabling a more detailed topographic analysis of AD-related network disruptions [97, 98]. Lastly, intervention-based studies should evaluate whether sleep-targeted therapies [6] can modulate qEEG features, reduce oxidative stress, and delay cognitive decline.

On the other hand, from a methodological perspective, adequate representation of participants with extreme biomarker values—who remain underrepresented but are critical for improving model robustness and generalization—should be ensured. Furthermore, approaches that integrate uncertainty estimation (either inherently, as in GP models, or through post hoc calibration in tree-based methods) could enhance the interpretability and clinical trustworthiness of EEG-based predictive models. AI-driven approaches could further optimize automated qEEG feature selection and classification pipelines, with deep learning techniques supporting improved model generalization [99, 100].

Conclusions

Our study demonstrates that advanced nonlinear and time-domain sleep EEG features, combined with an ML, reliably estimate CSF protein oxidation-derived markers in early AD. This approach outperforms traditional methods and offers a non-invasive, cost-effective alternative to lumbar puncture sampling. Future research will focus on developing customized algorithms for different patient populations, refining predictive models, and integrating qEEG biomarkers into early AD detection frameworks. If validated, this strategy could facilitate personalized risk assessment and targeted interventions aimed at reducing oxidative stress and slowing neurodegeneration in aging populations.

Supplementary Information

Below is the link to the electronic supplementary material.

Appendix 1: Definitions and mathematical formulas for quantitative EEG features

This appendix provides detailed definitions and corresponding mathematical formulas for the time-domain and nonlinear features extracted from EEG signals.

A.1 Time-domain statistics

• Mean Mean amplitude of the EEG signal [101].

  $$\begin{array}{c}\hfill \text{Mean}(X)=\frac{1}{N}\sum _{i=1}^{N}x(i)\end{array}$$ A1
• Peak-to-peak amplitude (PPAF) PPAF quantifies fluctuations in EEG by measuring successive peak-to-peak amplitude differences, capturing regions of heightened electrical activity often associated with seizures [102].

  $$\begin{array}{c}\hfill \text{PPAF}(X)=\frac{1}{N}\sum _{i=1}^N(Pea{k}_i-Troug{h}_i)\end{array}$$ A2

  where $Pea{k}_i$ is the amplitude of the i-th peak in the EEG signal, $Troug{h}_i$ represents the amplitude of the i-th trough following the i-th peak and N the total number of peak-to-trough pairs.
• Standard deviation (SD) Standard deviation quantifies the variability or dispersion of the EEG signals [101].

  $$\begin{array}{c}\hfill \text{SD}(X)=\sqrt{\frac{1}{N-1}\sum _{i=1}^N(x_i-\mu )}\end{array}$$ A3

  where $\mu$ is the mean.
• $Variance(Var)$ Variance measures the dispersion of the EEG signal values around the mean.

  $$\begin{array}{c}\hfill \text{Var}(X)=\frac{1}{N}\sum _{i=1}^N{(X_i-\mu )}^2\end{array}$$ A4

  where $\mu$ is the mean of the signal X, and N is the number of samples [59].
• Skewness (Skew) Skewness measures the asymmetry of the probability distribution of the EEG signal.

  $$\begin{array}{c}\hfill \text{Skew}(X)=\frac{1}{N}\sum _{i=1}^N{\left(\frac{X_i-\mu }{\sigma }\right)}^3\end{array}$$ A5

  where $\sigma$ is the standard deviation of the signal X [59].
• Maximum value (max) The maximum value is the highest amplitude observed in the EEG signal [60].

  $$\begin{array}{c}\hfill \text{max}(X)=max(X)\end{array}$$ A6

A.2 Non-linear feature

• Shannon entropy (ShanEnt) Shannon entropy quantifies the uncertainty or randomness in the EEG signal.

  $$\begin{array}{c}\hfill \text{ShanEnt}(X)=-\sum _{i=1}^{N}p(X_i)logp(X_i)\end{array}$$ A7

  where $p(X_i)$ is the probability of the i-th signal value [58].

Author contribution

Anna Michela Gaeta: conceptualization, writing—original draft, investigation, methodology, visualization. Lorena Gallego Viñarás: formal analysis, writing—original draft, methodology, investigation, data curation, visualization. Ferran Barbé: project administration, writing—review and editing, investigation, methodology. Pablo Martínez Olmos: writing—review and editing, formal analysis, methodology, investigation, visualization. Reinald Pamplona: methodology, data curation, writing—review and editing. Farida Dakterzada: methodology, data curation, writing—review and editing. Arrate Muñoz-Barrutia: project administration, investigation, formal analysis, methodology, writing—review and editing. Gerard Piñol-Ripoll: project administration, conceptualization, investigation, methodology, writing—review and editing.

Funding

This research project was funded by the Generalitat of Catalonia: Department of Health (PERIS 2019 SLT008/18/00050) and Agency for Management of University and Research Grants (2021SGR00761), the “Fundació La Marató TV3” (464/C/2014), the Institute of Health Carlos III (PI22/01687), and the Diputació de Lleida (PIRS 2021 and PIRS2024) to GP-R. The study has also been partially supported under grants PID2023-152233OB-I00 (to RP), PID2021-123182OB-I00 (to PMO), and PID2023-152631OB-I00 (to AMB), by Ministerio de Ciencia, Innovación y Universidades, Agencia Estatal de Investigación, MCIN/AEI/10.13039/501100011033, co-financed by European Regional Development Fund (ERDF), ‘A way of making Europe. This research was also funded by the Diputació de Lleida (PIRS2021 and PIRS2024), and the Generalitat of Catalonia: Department of Health (PERIS ref. SLT002/16/00250) and Agency for Management of University and Research Grants (2021SGR00990) to R.P. PMO also acknowledges the support by BBVA Foundation (Beca Leonardo a Investigadores y Creadores Culturales 2024) and by the Comunidad de Madrid under grants IND2022/TIC-23550 and ELLIS Unit Madrid. IRBLleida is a CERCA Programme/Generalitat of Catalonia. FB is supported by the ICREA program, Generalitat of Catalonia.

Data availability

The data supporting this study are available upon reasonable request to the corresponding author.

Declarations

Conflict of interest

The authors declare no competing interests.