Alzheimer disease (AD) represents a complex clinical challenge characterized by gradual and heterogeneous neurodegenerative progression over years to decades. While AD pathology involves amyloid and tau accumulation, its clinical manifestations span a continuum from individuals who were cognitively normal (CN) with AD pathology to mild cognitive impairment (MCI) and dementia. The subtle and variable trajectory of cognitive decline limits the sensitivity of traditional, episodic clinical assessments such as positron emission tomography (PET) imaging or cerebrospinal fluid analysis [1,2]. Continuous and home-based monitoring technologies offer a transformative approach to early disease detection by enabling real-time tracking of subclinical functional changes in naturalistic settings [3-6]. Quantitative physiological and behavioral biomarkers, such as gait velocity, sleep architecture, and circadian activity, can reveal gradual deterioration over weekly or monthly timescales that are often missed during infrequent clinical evaluations. Integrating these continuous monitoring systems into clinical workflows expands the therapeutic window for both pharmacological and nonpharmacological interventions, including targeted cognitive rehabilitation and personalized lifestyle management [7].

Olfactory dysfunction is one of the earliest symptoms of AD. It is recognized both as an early indicator and a potential clinical marker of the progression and severity of the disorder [8-11]. In addition to cognitive impairment, approximately 90% of patients with dementia of the Alzheimer type (DAT) experience a loss of olfactory abilities, such as difficulties with olfactory recognition and identification [12-16]. These deficits were previously attributed to pathological changes in the olfactory bulb (OB) and associated regions, including the entorhinal cortex, hippocampal cornu ammonis-1 region, and subiculum [17-19]. However, the timing of olfactory dysfunction during the progression of AD and its relationship with cognitive function remain unclear. Sniffing kits, such as the University of Pennsylvania Smell Identification Test (UPSIT), are frequently used to examine olfactory dysfunction in patients suspected of having MCI or DAT [19-22]. Although behavioral olfactory tests such as UPSIT and the Brief Smell Identification Test (B-SIT) have demonstrated feasibility in cognitively impaired populations, they fundamentally depend on participants’ subjective reports of odor identification. This subjective nature introduces variability from individual interpretation, decision-making processes, and verbal or written communication abilities that may confound assessment of actual olfactory function [23-25]. Even with full participant cooperation, responses reflect a complex integration of sensory perception, cognitive processing, and motor output rather than isolated olfactory capability. In contrast, electrophysiological recordings from the OB capture objective neural signals during odor processing, providing quantifiable biomarkers in terms of latency, amplitude, and frequency components. Consequently, several studies have proposed using electroencephalography recordings during odor tests to reduce nonolfactory-related performance interference [26-30]. Structural magnetic resonance imaging (MRI) studies have also sought to differentiate MCI from DAT by evaluating volumetric decreases in the OB and primary olfactory cortex [31-34]. Therefore, olfactory evaluation using sniffing kits is likely to be undervalued in clinical assessment of AD despite its potential as an early indicator.

To address these challenges, recent studies have proposed noninvasive and temporally precise recordings of the human OB, validated across a range of experiments, as a reliable method for assessing signals from the OB [35,36]. This approach has demonstrated olfactory dysfunction in patients with Parkinson disease (PD) compared with participants who were CN [37]. In this study, we conducted noninvasive digital recordings of olfactory responses among individuals who were CN, patients with MCI, and patients with DAT, using advanced digital signal processing and machine learning algorithms to transform raw electrophysiological data into quantifiable biomarkers. This computational approach allows the extraction of subtle features from neural signals that can objectively differentiate between cognitive states. We compared the results of various clinical examinations, including diffusion tensor imaging (DTI), MRI, and neuropsychological assessments. This study aimed to (1) identify differences in olfactory responses between CN and patient groups to extract olfactory characteristics critical for AD diagnosis, (2) explore the association between olfactory response characteristics and cognitive impairment in clinically diagnosed MCI and DAT using a hospital-based cohort, and (3) establish how the dynamics of olfactory responses across frequency bands relate to structural brain changes observed through medical imaging. Digital recording of odorant-mediated OB responses addresses the subjectivity inherent in behavioral olfactory assessments by providing objective, reproducible neural measurements. Unlike self-reported odor identification, electrophysiological signals directly reflect the functional integrity of olfactory neural circuits and can reveal temporal dynamics of olfactory processing not accessible through behavioral testing alone. The rapid advancement of mobile health (mHealth) and ubiquitous health technologies has created new opportunities for continuous, home-based assessment of neurological function. Our electrophysiological olfactory testing approach embodies these paradigms by using a lightweight, wearable 6-electrode electroencephalography device that can be quickly applied by users or caregivers. The entire testing procedure requires only 5‐10 minutes, enabling regular self-assessment or caregiver-assisted monitoring in home settings. Unlike conventional clinic-based evaluations, this system allows individuals to track their olfactory and cognitive function through brief, repeatable sessions. Its portability and ease of use facilitate longitudinal monitoring, which is critical for detecting subtle, early changes in olfactory processing associated with neurodegeneration. We hypothesized that digital recordings of OB responses could provide a more comprehensive understanding of cognitive decline in AD, serving as a potential digital biomarker.

This study used a cohort design across 3 cohorts: CN, MCI, and DAT. The prospective registry study of olfactory response heterogeneity in AD progression enrolled participants from 2 tertiary medical centers in the Republic of Korea (Hanyang University Hospital and Myongji Hospital) between September 2023 and September 2024. The community-dwelling cohort was recruited from 3 geographically diverse provinces: Seoul, Gyeonggi, and North Chungcheong. A total of 129 individuals were initially screened for this study, of whom 71 were included in the final analysis, including 18 CN controls, 39 individuals with MCI, and 14 patients with DAT. All participants with DAT satisfied the diagnostic criteria of the National Institute on Aging and the Alzheimer’s Association (NIA-AA) [25]. Comprehensive baseline assessments, including neuroimaging and neuropsychological evaluations, were conducted according to standardized protocols under physician supervision.

The study defined eligibility criteria for participant inclusion, restricting the age range for enrollment to 50‐85 years. Inclusion criteria required participants to have sufficient cognitive capacity to complete both neuropsychological and olfactory assessments. Participants with contraindications to neuroimaging were excluded. The diagnostic status was determined following the updated NIA-AA workgroup criteria [23]. Notably, MCI was diagnosed independently of the Seoul Neuropsychological Screening Battery II (SNSB-II) score using a comprehensive clinical evaluation. Diagnosis was based on documented cognitive decline corroborated by caregivers, objective episodic memory impairment on the delayed recall component of the Seoul Verbal Learning Test (performance below −1.5 SD, adjusted for age and education), a Clinical Dementia Rating score of 0.5, and fulfillment of the NIA-AA core clinical criteria for MCI due to AD. After the diagnosis of MCI was established, participants were further subdivided into early and late stages according to cognitive severity. The SNSB-II provides age-, sex-, and education-adjusted z scores across 5 cognitive domains (attention, language, visuospatial function, memory, and frontal and executive function), from which a composite z score was calculated by averaging domain scores. This composite score was used solely for severity stratification within the MCI group and was not involved in diagnostic determination. Participants with a composite z score ≥−1.0 were classified as early mild cognitive impairment (EMCI), whereas those with scores <−1.0 were classified as late MCI, in line with prior studies linking lower composite scores to broader multidomain impairment and increased risk of progression to dementia [38,39]. Using this criterion, 30 participants were classified as EMCI and 9 as late MCI. The relatively preserved composite scores observed in the EMCI group (mean 0.15, SD 0.58) reflect focal memory impairment that becomes attenuated when averaged across largely preserved cognitive domains, a pattern commonly observed in early-stage MCI. Clinical evaluation consisted of a comprehensive medical history review, neuroimaging analysis, and neuropsychological examination conducted by a qualified physician. Detailed diagnostic criteria are provided in Multimedia Appendix 1.

Neuropsychological assessments were performed using a standardized neuropsychological battery called the SNSB-II [38]. This battery includes examinations of attention, language, calculation ability, visuospatial skills, memory, frontal-executive function, and general cognition, such as the Mini-Mental State Examination (MMSE) [40] and Clinical Dementia Rating. The SNSB-II comprises the digit span test (forward and backward), the Korean version of the Boston Naming Test, Rey-Osterrieth Complex Figure Test, Seoul Verbal Learning Test, Phonemic Controlled Oral Word Association Test, and Korean Color Word Stroop Test. All SNSB-II variables were converted to z scores and standardized for age, sex, and years of education.

All participants underwent an 18F-florbetaben amyloid PET and structural MRI scan. Brain DTI was performed using a 3.0-T Achieva MRI system (Philips Medical Systems). DTI images were acquired in the transverse orientation using a pulse sequence with a single-shot spin-echo diffusion-sensitized echo-planar imaging sequence (15 gradient directions plus a B0 image with b value=800 s/mm², repetition time/echo time=13898/55 ms, field of view=224×224, matrix=112×112, number of averages=2, slice thickness=2.0 mm, and flip angle=90°). Amyloid positron emission tomography–computed tomography imaging was conducted following intravenous injection of 8 mCi of the amyloid radiotracer F-18 Florbetaben (Neuraceq; Piramal). Following a 90-minute uptake period, imaging was conducted for approximately 20 minutes using a dedicated PET/computed tomography (CT) scanner (Biograph 6; Siemens Medical Systems). The PET scans were reviewed by 2 expert PET readers and 1 neurologist based on predefined regional cortical tracer uptake and brain amyloid plaque load scoring systems. Four regions of interest, including the frontal, temporal, and parietal cortices along with the posterior cingulate and precuneus, were assessed in the visual analysis of the 18F-florbetaben PET scans. The regional cortical tracer uptake scores were condensed into a single 3-grade scoring system for each PET scan (brain amyloid plaque load score) as follows: 1 for no Amyloid-beta (Aβ) load, 2 for minor Aβ load, and 3 for significant Aβ load. Tract-based spatial statistics (TBSS) analysis was conducted to examine differences in DTI between patients with AD and healthy controls using the Functional Magnetic Resonance Imaging of the Brain Software Library (University of Oxford). The preprocessing pipeline included correction for head motion and eddy current distortions, followed by the computation of diffusion tensor metrics, specifically fractional anisotropy (FA), axial diffusivity, and radial diffusivity, to characterize the white matter fiber architecture. Subsequently, FA images were spatially normalized to a standardized brain template using the Functional Magnetic Resonance Imaging of the Brain nonlinear imaging registration tool. A mean FA skeleton was generated using projection techniques to represent the central white matter pathways common to all participants, with an applied threshold criterion of 0.2. Individual FA maps were then projected onto this skeleton for group-level analyses. Regional white matter tract identification was facilitated using the Johns Hopkins University ICBM-DTI-81 white matter atlas [41]. Statistical analyses used permutation-based nonparametric inference (5000 permutations) to assess between-group differences [42]. The statistical threshold was established at P<.05, with correction for multiple comparisons implemented using threshold-free cluster enhancement. Volumetric quantification of the OB was performed by manual delineation of sequential T1-weighted images using ImageJ (version 1.53h; National Institutes of Health). Bilateral OB annotations were independently conducted by 2 experienced raters using FreeView (Massachusetts General Hospital), a FreeSurfer visualization tool, and ambiguous cases were resolved by consultation with a qualified neurosurgeon [43]. To minimize bias, the raters were blinded to participant group assignment, age, and sex. Individual OB volumes were calculated by summing the product of the surface area measurements and slice thickness, with the total OB volume derived from the mean of the bilateral measurements.

Olfactory function was assessed using a modified 6-item version adapted from the B-SIT, which itself is a 12-item subset of the 40-item UPSIT [44]. Participants were presented with a 4-category multiple-choice questionnaire and asked to identify which smells corresponded to the scent strip. As familiarity and responsiveness to certain odors may vary among ethnic groups, we selected 6 types of odors (cinnamon, smoke, chocolate, gasoline, soap, and onion) with a correct answer rate of 70% or more among healthy individuals from the 12 smells in B-SIT for this experiment. This culturally adapted version showed comparable diagnostic performance to the full B-SIT while avoiding misclassification due to odor unfamiliarity rather than true olfactory dysfunction.

The olfactory assessment protocol integrated the 6-item modified B-SIT, a validated and standardized instrument for evaluating olfactory function, with concurrent electrophysiological measurements to characterize the temporal dynamics of olfactory processing. Olfactory responses were recorded using a 6-channel active electrode electroencephalography system (Cyton; OpenBCI), and data were collected and digitized at a sampling rate of 250 Hz. In this study, the configuration of the OpenBCI electroencephalography cap was modified to optimize the detection of signals from the OB, following the approach proposed by Iravani et al [35]. Six electrodes were symmetrically placed on a mount along the centerline of the forehead, with each electrode spaced 5 cm apart to optimize the detection of OB signals. The helmet mount was designed to be adjustable, allowing the examiner to loosen it for easier electrode placement and subsequently tighten it to maintain firm and consistent contact with the participant’s skin. This configuration ensured reliable signal detection from the OB, as illustrated in Figures 1A and 1B. The electroencephalography signals were recorded continuously throughout the test session.

During olfactory testing, a respiratory sensor monitored participants’ breathing patterns to trigger olfactory stimulation. Recordings were segmented into epochs ranging from 1200 ms prestimulus to 1800 ms post stimulus, time locked to respiratory triggers, and rereferenced to the average of the left and right earlobe electrodes. Recorded signals were notch filtered at 60 Hz to eliminate power line noise. Additionally, muscle artifacts were detected by applying a bandpass filter (Butterworth filter, order 8) to the raw data, followed by a Hilbert transformation to extract the amplitude values. These values were standardized using z scores, with trials exceeding a z score of 6 assumed to be contaminated by muscle artifacts and excluded from further analysis. Considering the impact of eye movements and blinking on olfactory response signals, a lower z score threshold of 4 was used to improve the detection sensitivity, leading to the exclusion of trials with z scores exceeding this threshold. To further reduce blinking-related artifacts, participants were instructed to keep their eyes closed throughout the test. For signal processing, preprocessed and artifact-corrected olfactory signals were decomposed into time-frequency maps using a multitaper sliding window approach. This method was applied across a frequency range of 0.1‐100 Hz, with a step size of 0.1 Hz, and across a time range of –1 to 1 second, with a step size of 5 ms. Power in each time-frequency bin was estimated using 2 tapers derived from discrete prolate spheroidal sequences. To capture at least 2 cycles of each frequency bin, the taper length was selected to ensure accurate power estimation across the spectrum. Wavelet transformation for each time bin was conducted using 2 sets of wavelet functions derived from discrete prolate spheroidal sequence tapers. This transformation involved convolving the wavelet functions with the olfactory response signal in the frequency domain by multiplying the fast Fourier coefficients of the signal by those of the wavelets. The frequency-smoothing parameter was set to 80% of the value in each frequency bin. To account for motor responses to sniffing, baseline-corrected odor trials were compared with baseline-corrected sniff trials to analyze phase-locked responses, which were subsequently converted into decibels. For regression and classification, features were extracted by identifying time-frequency regions with maximal group differences via permutation testing and calculating each participant’s mean power within these regions. This process yielded 5 electrophysiological features in the beta and gamma bands for subsequent analyses. All preprocessing and time-frequency decompositions were performed using the open-source Fieldtrip toolbox in MATLAB (version R2023b; The MathWorks, Inc).

All statistical analyses were conducted using MATLAB (version R2023b) with the Signal Processing and Fieldtrip toolboxes [45]. Spectral densities of the 6 electrodes were averaged across the channels. Initially, the average power comparison was used to assess the intensity of gamma synchronization (30‐100 Hz) in the olfactory response one second after olfactory stimulation. Time-frequency maps of the olfactory responses between the 2 comparable groups were assessed using a nonparametric Monte Carlo permutation test. A total of 1000 permutation tests were conducted to identify significant clusters that differed between the 2 groups within the time-frequency maps, with significance defined as P<.05 and a cluster size exceeding 300 bins. This cluster-based permutation approach inherently controls for multiple comparisons across all time-frequency bins by evaluating cluster-level statistics, thereby maintaining the family-wise error rate without requiring additional correction methods. The identified clusters were labeled according to their latencies (for example, components 1 and 2; Figure 2). Machine learning classification was performed using a support vector machine with a radial basis function kernel. Model development involved leave-one-out cross-validation to maximize sample usage given the limited dataset. The 5 electrophysiological components were used as input features. Classification performance was evaluated using accuracy, precision, recall, and F₁-score metrics. Regression and classification analyses used only the 5 electrophysiological components extracted from the time-frequency analysis as input features, without inclusion of demographic covariates. All t tests were 2-tailed. Statistical significance was indicated by asterisks, with * representing P<.05, ** denoting P<.01, and *** signifying P<.001.

This study was approved by the Institutional Review Board of Hanyang University Medical Center (2023-06-019) and conducted in accordance with the Declaration of Helsinki. All participants or their legally authorized representatives provided written informed consent and were informed of their right to withdraw at any time without penalty. All participant data were deidentified and assigned unique study codes, with the linking key stored separately in a password-protected file accessible only to the principal investigator. Neuroimaging data were anonymized using defacing algorithms to remove identifiable facial features. No financial compensation was provided to participants. However, participants received their complete diagnostic test results and medical consultations free of charge, which were disclosed as benefits of participation during the informed consent process.

Digital olfactory testing was conducted alongside several traditional methods for AD diagnosis, including brain imaging techniques such as MRI and DTI, as well as neurobehavioral assessments such as the SNSB-II and MMSE. To evaluate olfactory ability, we conducted a 6-item modified B-SIT while simultaneously recording the electrical activity of the OB across the CN, MCI, andDAT groups. Demographic characteristics, cognitive status, and medication profiles of the participants are summarized in Table 1. Six electrodes were attached to positions capable of measuring OB responses for digital olfactory testing (Figure 3). The obtained signals were subjected to preprocessing and spectral analysis, resulting in a spectrogram. The spectral density of the signal was time-locked to stimulus onset and averaged across the 6 electrodes and trials to optimize the signal-to-noise ratio. The time-frequency map for the odor trials was determined within the specified time and frequency bands. To exclude contamination from sniffing and other motor-related artifacts, we compared the air and odor conditions. Odor event-related synchronization was observed in the beta (13‐30 Hz) and gamma (30‐100 Hz) bands at approximately 200‐550 ms post stimulus. Subsequently, specific areas showing significant differences between the CN and DAT groups were identified using nonparametric tests with 1000 permutations, resulting in the isolation of 5 distinct components. The 5 extracted components were analyzed for correlations with existing medical imaging results, such as MRI and DTI findings, as well as neurobehavioral test outcomes (MMSE and SNSB-II), to validate their effectiveness. This process was investigated to determine whether these olfactory characteristics could serve as early diagnostic markers of AD. An illustration of this process is shown in Figure 1.

Patients with DAT reported olfactory impairment when assessed using the B-SIT [20,22,25], which included 12 odors selected from the UPSIT [44]. As familiarity and responsiveness to certain odors may vary among ethnic groups, we first selected 6 types of odors (cinnamon, smoke, chocolate, gasoline, soap, and onion) with a correct answer rate of 70% or more among healthy individuals from the 12 smells in the B-SIT (Figure 4). Behavioral olfactory testing failed to distinguish between the CN and MCI groups. Discrimination rates were comparable across all 6 odors tested, with both groups showing relatively preserved identification performance (mean correct: CN 81.5%, SD 16.5% vs MCI 74.2%, SD 25.1%; P=.14). Only the clinical dementia group showed significant impairment (mean 45.3%, SD 31.8%; P<.001 vs both CN and MCI). Figure 2A shows the spectrum of olfactory responses to the 6 selected odors in healthy individuals and patients with DAT. In the CN group, odor-responsive signals were observed within 0.6 seconds after odor exposure, specifically in the beta and gamma frequency bands across 6 electrodes. Conversely, patients with DAT showed no detectable gamma synchronization at these temporal and frequency intervals, thereby confirming the ability of the control group to identify robust olfactory signals. As shown in Figure 2B, the average power recorded in the response of patients with DAT to the selected odors was significantly lower than that in the CN group for cinnamon (P=.02), smoke (P=.04), chocolate (P=.01), gasoline (P=.03), soap (P=.01), and onions (P=.02). This uniform impairment across chemically diverse odorants indicates a central olfactory processing deficit rather than peripheral receptor-specific dysfunction. Figure 2C shows the statistically significant differences in odor responses observed between the CN and DAT groups. To identify specific regions between the 2 groups, we performed nonparametric tests with 1000 permutations. The results are indicated by threshold t-maps, highlighting areas with P<.05. The upper panel demonstrates that the CN group showed enhanced power in the beta and gamma bands in response to the selected odors compared to the DAT group (CN > DAT). Notably, no significant difference in power was observed in patients with DAT when compared to the CN group (CN < DAT). This analysis revealed significant increases in the beta and gamma band power in the CN group compared to the DAT group at approximately 250 ms postodor onset (t₃₀=2.1192; P=.02). Conversely, a decline in gamma frequency power was observed at approximately 950 ms and 980 ms in the CN group compared with the DAT group (t₃₀=−1.8412; P=.04). Significant clusters (P<.05; cluster size > 300) were identified and sequentially labeled based on their temporal order as component 1, component 2, and so on. As shown in Figure 2D, 5 distinct components were identified, each of which was visually represented by a different color. The statistical parameters for each component were as follows: component 1 (t₃₀=1.9934; P=.03), component 2 (t₃₀=2.1823; P=.03), component 3 (t₃₀=1.9938; P=.03), component 4 (t₃₀=1.9182; P=.04), and component 5 (t₃₀=2.1393; P=.02). Additionally, Figure 5 provides the differences in the time-frequency maps between the CN and DAT groups in response to each of the 6 odorants.

For the loss of smell to be used as an early diagnostic method for AD, the detection of early signs is essential during the MCI stage, which precedes dementia. However, existing approaches do not clearly differentiate between individuals who were CN and those with MCI. Therefore, we conducted an experiment to determine whether digitalized measurements of olfactory function could effectively classify MCI severity. To examine the progression of olfactory dysfunction across the cognitive spectrum, the MCI group was stratified according to cognitive severity. Using the SNSB-II composite z score with a cutoff of –1.0 (adjusted for age and education), 30 participants were identified as EMCI (mean z score 0.15, SD 0.58) and 9 as late MCI (mean z score –1.65, SD 0.42) [38,39]. This score reflects multiple cognitive domains, including language, attention, visuospatial skills, memory, and frontal and executive function. Neuropsychological evaluations and relationships between the groups are summarized in Table 1 and Figure 6. First, we generated a differential time-frequency spectrum comparing EMCI to late MCI (EMCI > late MCI) and vice versa (EMCI < late LMCI). No significant differences in the odor-induced gamma band power were observed among the groups. Notably, the EMCI group exhibited faster responses within 490 ms of odor onset across the 6 electrodes. In contrast, the late MCI group showed delayed gamma synchronization at approximately 850 ms post stimulus after controlling for sniffing and evoked responses (Figure 7A). However, no difference in the average gamma-band power was found between the groups (Figure 7B). Furthermore, we analyzed significant increases in the average power spectrum across the CN, EMCI, late MCI, and DAT groups. As shown in Figure 8A, the gamma band power spectrum showed a significant reduction and delayed response time starting from the late MCI stage compared with individuals who were CN. Additionally, the beta band power spectrum progressively decreased as the condition advanced from CN to EMCI, LMCI, and DAT stages. These results were consistent for each of the 6 selected odors, as shown in Figures 9A and 9B. In this regard, the structural similarity index matrix (SSIM) analysis of the olfactory response spectrograms revealed high similarity between the control and EMCI groups (SSIM=0.43) and between the late MCI and DAT groups (SSIM=0.61), as provided in Figure 10. Olfactory characteristics were derived by computing the mean values across the temporal stages for each component. Components 1, 3, and 5 were associated with the beta band, whereas components 2 and 4 were associated with the gamma band (Figure 2D). The spectral analysis in Figure 8B revealed significant decreases in the beta band power from EMCI (P<.001) to late mild cognitive impairment (LMCI; P=.03), culminating in DAT (P=.009). The gamma band power demonstrated a distinct temporal progression, with significant reductions emerging at LMCI (P=.01) and progressing through DAT (P=.001), while showing marginal alterations at EMCI (P=.08).

To evaluate clinical applicability, the extracted spectral characteristics, categorized into beta and gamma frequency bands, were used to examine their association with cognitive performance. As shown in Figure 8C, multiple linear regression analysis demonstrated that electrophysiological features explained 91.69% of the variance in SNSB-II z scores (P<.001). This remarkably strong association between olfactory biomarkers and cognitive function was consistent across the entire cognitive spectrum from CN to DAT. The extremely low P value indicates that this relationship is highly unlikely to have occurred by chance. Additionally, MMSE scores showed similarly strong associations with olfactory responses (R^²=0.94; Figure 10). These results demonstrate that the magnitude of electrophysiological responses in beta and gamma bands closely tracks with the severity of cognitive impairment, supporting the potential of these biomarkers for objective cognitive assessment.

Model selection was conducted using the Akaike information criterion to assess the accuracy of fit for each olfactory component based on the in-sample prediction error. In classifying the CN, MCI, and DAT groups, component 2 showed the optimal performance, with the lowest Akaike information criterion value of 122.82, followed by component 3 (124.69), component 4 (125.61), component 1 (126.57), and component 5 (126.62). Subsequently, we used a stepwise component addition approach within the regression framework to optimize the prediction accuracy. The analysis revealed a peak performance with a 3-component model that included components 5, 1, and 4, as shown in Figure 8D. We implemented a support vector machine regressor that integrates component-specific olfactory characteristics to assess the diagnostic utility of digitalized olfactory response recordings. Model performance was evaluated using leave-one-out cross-validation, where each participant was held out as a test sample. The resulting classification model demonstrated exceptional discriminative capability across the 3 groups, achieving 83.09% accuracy, 88.64% precision, perfect recall (100%), and an F₁-score of 93.98% (Figure 8E). These metrics underscore the robust diagnostic potential of the proposed multicomponent approach. Therefore, the strong predictive relationship between odorant-mediated olfactory responses and cognitive performance underscores the potential clinical utility of this method for early diagnosis of AD. Further analyses were performed to evaluate the component-specific contributions in distinguishing between adjacent cognitive stages (CN-EMCI, EMCI-LMCI, and LMCI-DAT) to identify the key components for the early detection of AD. The discrimination between CN and EMCI states was achieved with remarkable precision using components 1, 3, and 5, all representing beta band activity, resulting in 93.75% accuracy, 90.91% precision, and 100% recall (Figure 8F). Notably, the predominant use of beta band components rather than gamma band components in the pre-DAT stage classification aligned with the findings presented in Figure 8B. The gamma band components were critical in differentiating patients with DAT from the LMCI group, marking a distinct transition point in disease progression. Table 2 summarizes the classification performance across different diagnostic comparisons.

To elucidate the brain areas associated with the beta and gamma components within the olfactory system, we used a dual analytical approach that combined TBSS and volumetric MRI quantification. TBSS analysis revealed significantly reduced FA in the hippocampus cingulum of the DAT group compared with that in the CN group (P<.05), as shown in Figure 11A. The inverse correlation between FA values and cognitive impairment suggests that reduced white matter integrity in the hippocampal cingulum region is associated with more severe cognitive deficits [46-48]. Further analysis revealed a positive correlation approaching statistical significance between the beta-band components (mean of components 1, 3, and 5) in olfactory responses and FA values within the hippocampal-cingulum tracts in patients with DAT (R^²=0.28; P=.06), consistent with beta-band oscillations representing top-down cortical feedback signals. In contrast, gamma-band components showed minimal correlation with the integrity of the hippocampal-cingulum tracts (R^²=0.013; P=.53; Figure 11B).

In T1-weighted MR images, volumetric analysis of the OB revealed progressive atrophy across the disease spectrum, with a notable reduction in volume as cognitive status declined from CN to EMCI, LMCI, and DAT stages. Figure 11C provides representative T1-weighted MR images from each stage, with the OB regions delineated by red dashed circles, illustrating sequential volume reduction corresponding to cognitive decline. Quantitative analysis revealed significantly reduced OB volumes in patients with LMCI compared with those in the CN group (P=.005), whereas patients with DAT exhibiting a pronounced volume reduction (P<.001), as shown in Figure 11D. In contrast, the EMCI group did not show a statistically significant volume reduction relative to the control group (P=.20), indicating that substantial OB atrophy may become more evident during later stages of cognitive decline [32,49]. This temporal pattern aligns with the delayed reduction in gamma-band olfactory components (components 2 and 4), suggesting parallel trajectories between structural and functional changes during later stages of the disease. Demographic analyses confirmed minimal influence on our primary outcomes. For OB volume, weak correlations with age (r=–0.23; P=.05) and education (r=0.17; P=.15) cannot explain the stage-specific volume reductions observed between groups (P<.01). Further analysis of OB volume and frequency-band power revealed a positive correlation between average frequency-band power and OB volume (Figure 11E). Specifically, for gamma bands, the coefficient of determination was 0.2108, with P<.001, indicating a strong and statistically significant relationship. In contrast, the correlation observed for the beta bands showed minimal optimization in linear fitting, yielding an R^² value of 0.0613 (P=.05), suggesting a weaker association compared with the gamma bands, although still statistically significant.

Objective measurement of odorant-mediated olfactory responses has the potential to enable the early diagnosis of AD by accurately predicting clinical neurobehavioral examination scores. In this study, we aimed to determine whether the olfactory response can serve as an early biomarker for AD, enabling us to distinguish between CN, EMCI, LMCI, and DAT groups. Consequently, a significant difference in the average power was observed between individuals who were CN and patients with DAT, which is consistent with the results of the 6-item modified B-SIT. These findings highlight a potential advantage of electrophysiological over behavioral olfactory assessments. While behavioral testing revealed no significant differences between CN and MCI groups, OB recordings showed distinct neural signatures that may reflect subtle disease-related changes. However, given the modest sample size and partial overlap in performance, these results should be considered preliminary. Larger validation studies are warranted to confirm whether OB electrophysiological responses can reliably detect early olfactory dysfunction not captured by behavioral tests. When measuring the electrical signals from the OB across all 4 groups, we observed a decrease in the beta- and gamma-band power and a delay in responses to odor exposure, which correlated with the progression of cognitive dysfunction.

In addition, signals obtained through digitalized olfactory measurements enabled the identification of olfactory response components distributed across various frequencies and time intervals, which allowed for discrimination between the DAT and CN groups at the individual level. The model exhibited outstanding performance in distinguishing among groups: CN (72.22% accuracy), MCI (100% accuracy), and DAT (50% accuracy, 100% precision). These results correlated strongly with the SNSB-II cognitive dysfunction levels and structural characteristics of the brain. Post hoc analyses demonstrated that potential demographic confounders had minimal influence on our findings, with negligible correlations observed between age and cognitive scores (r=–0.16; P>.20) and between education and cognitive scores (r=0.24; P>.05). These findings were further confirmed by formal demographic adjustment, in which the primary electrophysiological predictor retained independent significance for both SNSB-II (P=.04) and MMSE (P=.009) after controlling for age, sex, and education. While age and education shared some explanatory variance with secondary spectral components, this overlap is expected given the known influence of these factors on cognitive performance and the fact that SNSB-II scores are already standardized for age, sex, and education. Importantly, classification performance did not improve when demographic covariates were included, supporting the independent contribution of electrophysiological features. Future studies with larger, demographically balanced cohorts will be essential to further refine the generalizability and demographic robustness of these electrophysiological biomarkers.

Furthermore, we recorded signals from 6 electrodes using a slightly modified version of a previously reported electrobulbogram to measure the electrical response of the OB. This method has proven to reliably assess OB function in healthy participants, demonstrating high test-retest reliability in previous studies [35]. Notably, patients with PD show a high prevalence of olfactory dysfunction, with approximately 90% of patients experiencing some degree of smell impairment [50]. The high rate of olfactory dysfunction, which is often evident years before the onset of motor symptoms, highlights its potential role as an early indicator of disease. Recently, a noteworthy study presented the electrobulbogram method as an effective tool for detecting PD by distinguishing patients diagnosed with the condition from control participants [37].

Olfactory dysfunction is one of the earliest clinical symptoms of AD. It is recognized as both an early indicator and a potential clinical marker of the progression and severity of the disorder [8-10]. In addition to cognitive impairment, approximately 90% of patients with DAT experience loss of olfactory abilities, such as difficulties with olfactory recognition and identification [12]. Sniffing kits, such as UPSIT, are frequently used to examine olfactory dysfunction in patients suspected of having MCI or AD [20-22]. In this study, although the sniffing kit clearly distinguished between healthy individuals and patients with DAT, the differences between healthy and MCI participants were not significant. When electrical signals from the OB were compared across all 4 groups, a decrease in beta- and gamma-band power and delayed responses to odor exposure were observed. Given that the gamma band is primarily associated with processes within the OB [51-55], these observed differences indicated impaired OB processing in DAT due to neuronal loss [10]. Moreover, activity in the beta band is likely generated by top-down projections, which is further supported by the increased latencies and high correlation with cingulum hippocampus diffusivity observed in our data [54,56]. These findings suggest an impairment in the bidirectional interaction between the OB and higher olfactory cortex regions, such as the amygdala and orbitofrontal cortex, which are involved in the processing of various aspects of odor perception. Although a significant reduction in total power was noted between the CN and DAT participants, no significant decrease in total power was found between the CN and MCI groups (early and late) or between the early and late MCI groups. These results suggest a gradual delay in gamma-band synchronization, ultimately leading to a complete loss of olfactory response. This observation is consistent with prior research, indicating that the recognition ability in patients with DAT declines early, whereas the detection threshold is preserved for a relatively longer period [22]. The combined use of TBSS and volumetric analyses offers complementary insights; TBSS enables a detailed investigation of white matter tracts, whereas volumetric quantification assesses regional brain volume changes. This comprehensive structural analysis framework helps map the relationships between functional olfactory components and their underlying neuroanatomical correlations. The observed differential patterns of association suggest distinct neuroanatomical substrates for the beta- and gamma-frequency components of olfactory processing. Our cross-sectional findings show that functional olfactory changes can be present even when volumetric measures remain relatively intact. However, the cross-sectional nature of our study precludes any conclusions about temporal relationships or causality. While we observed functional changes in the presence of preserved structure in some groups, we cannot determine whether electrophysiological changes truly precede volumetric alterations during disease progression. Longitudinal studies are essential to establish these temporal relationships and to confirm whether OB dysfunction represents an early biomarker that emerges before structural brain changes.

In summary, we attempted to digitize the measurement of OB function to differentiate individuals who were CN from patients with EMCI, LMCI, or DAT in a noninvasive manner. Accordingly, we isolated specific olfactory characteristics that significantly differentiate CN and DAT groups, which were successfully regressed against clinical AD diagnostic examinations, demonstrating high levels of sensitivity, specificity, and accuracy. Notably, these olfactory responses are specifically associated with disease characteristics and affected brain regions according to their frequency bands. This finding suggests that the OB, which is responsible for olfactory responses, is pathologically damaged in patients with DAT, resulting in olfactory dysfunction caused not only by the affected brain regions but also by direct biological damage to the OB. Therefore, our results support the potential of this methodology for further development as a tool to facilitate the early diagnosis of AD and provide a robust and objective biomarker for the disease.

Our study demonstrates that noninvasive digital recordings of OB function can serve as a novel biomarker for continuous and home-based monitoring of cognitive states across the continuum from normal aging to Alzheimer dementia, supporting early detection and timely intervention within an mHealth framework. The electrophysiological biomarkers we identified—particularly the progressive decline in beta-band power and later-stage reduction in gamma-band power—correlate strongly with established clinical and neuroimaging markers of AD progression. These electrophysiological changes may precede structural brain alterations, highlighting their potential value for early detection and longitudinal tracking. As a digital health tool, this technology has potential applications beyond diagnosis, including monitoring disease progression and potentially assessing therapeutic efficacy in clinical trials. Beyond objective signal capture, our wearable olfactory recording system offers distinct practical advantages for mHealth deployment. The device can be fitted in less than 2 minutes, and the entire assessment takes approximately 5‐10 minutes, allowing frequent, home-based measurements with minimal burden on users or caregivers. This approach embodies the mHealth paradigm, where continuous, real-world monitoring of physiological biomarkers enables early detection of cognitive decline and supports timely, personalized interventions. These findings support the development of accessible mHealth technologies for early AD detection and ongoing disease surveillance, addressing the growing need for noninvasive, cost-effective screening tools that can be deployed more widely than expensive neuroimaging techniques or invasive procedures.