A new paradigm for stratified cognitive assessment: fusing serum Raman spectroscopy and ensemble learning

Abstract

Background

In China, the growing need for stratified cognitive assessment demands portable solutions. However, in community-based screening, neuropsychological assessments suffer from low acceptance and low efficiency. Thus, we aimed to develop a serum Raman spectroscopy-based ensemble learning approach for graded cognitive screening, and explore the Raman spectral features of cognitive impairment and their mapping relationship to brain function.

Methods

We recruited 220 subjects for modeling and 40 subjects for validation. Ensemble learning model was built using serum Raman spectra. High-weight features were analyzed for inter-group differences, graph theory properties, and associations with brain networks.

Results

We developed a serum Raman spectroscopy-based ensemble learning approach for graded cognitive screening. The model distinguishing normal cognition, mild cognitive impairment, and dementia achieved area under curve of 0.92 (testing) and 0.89 (validation). Raman shifts at 1602, 1002, and 1666 cm⁻¹ showed intensity and nodal changes, linking their modulation to both cognitive decline and altered posterior default mode network (pDMN) interactions.

Conclusions

The Raman spectroscopy-based ensemble learning model was powerful for cognitive screening. The alteration at the 1602, 1002, and 1666 cm⁻¹ represented key Raman signatures of cognitive impairment, reflecting impaired pDMN-related inter-network interactions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13195-026-01975-z.

Background

By 2025, China will have entered a phase of profound population aging. Dementia, a major age-related condition, has already become the fifth leading cause of death nationwide [1]. Currently, few medical interventions can reverse cognitive decline, making early screening and diagnosis critically important.

Yet cognitive screening within Chinese communities remains exceptionally challenging. Current approaches rely heavily on neuropsychological assessments such as the Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA). These tests must be administered one-on-one by trained clinicians, a process that is time-intensive (≥ 40 min per individual) and difficult to scale. Furthermore, linguistic diversity and low adherence among older adults limit the effectiveness of such tools. Scaling these methods to meet the target set by China’s National Action Plan for Addressing Dementia in the Elderly (2024–2030), which aims for ≥ 80% screening coverage among older adults under health management, appears impractical.

In contrast, peripheral blood testing is more readily accepted in community populations. Raman spectroscopy, in particular, is well-suited for community screening due to its operational simplicity, low cost, and high acceptability. The technique detects molecular vibrations via inelastic scattering of light, capturing real-time, sensitive information on biomolecular changes associated with disease, such as protein conformational shifts or metabolic alterations [2–4]. Raman spectroscopy reflected amyloid deposition and neurofibrillary tangles in the brain, and its integration with machine learning enabled evaluation of Alzheimer’s disease (AD) progression, which shows promise in diagnosing central nervous system disorders. However, existing models remain limited to binary classification (such as AD vs. healthy controls), which lack universality and detailed classification of cognitive impairment [5, 6]. Machine learning, which can process complex spectral data rapidly, has advanced the standardized application of Raman spectroscopy. The development of artificial intelligence has brought us machine learning models with stronger performance and generalizability, such as ensemble learning. Ensemble learning is a machine learning methodology that trains and combines multiple base learners to accomplish learning tasks. Compared to single models, ensemble methods integrate the strengths of diverse traditional machine learning and deep learning architectures, mitigate overfitting risks, and enhance predictive accuracy and robustness [7]. Combining Raman spectroscopy with ensemble learning may thus enable scalable, precise, and widely applicable cognitive screening.

Cross-modal joint analysis may open new horizons for understanding disease mechanisms. Neuroimaging features are critical clinical manifestations of cognitive impairment and currently serve as a key basis for neuromodulation. Spectral dynamic causal modeling (spDCM) is a brain network analysis method that can infer direct (causal) connections between neural systems [8]. For instance, integrating Raman spectroscopy with neuroimaging may be helpful for identifying key alterations of brain regions in cognitive impairment, thereby providing references for future neuromodulation therapies. In summary, this study will employ Raman spectroscopy, ensemble learning, and neuroimaging to: (1) develop a graded cognitive impairment classification model using serum Raman spectra and ensemble learning, (2) explore core Raman shift in cognitive impairment, and (3) uncover the neuroimaging correlates of spectral changes to advance our understanding of disease mechanisms and therapeutic targeting.

Methods

Participants

A total of 260 subjects were enrolled in this study by the department of neurology in Nanjing Drum Tower Hospital (Nanjing, China) from 2022 to 2023. There were two cohorts in this study: Gulou cohort (220 subjects) and the community cohort (40 subjects). Subjects of Gulou cohort, used for model training and testing, were consecutively recruited from Nanjing Drum Tower Hospital. Subjects of the community cohort, used for validation, were consecutively recruited from communities in Nanjing. The study was approved by the Ethics Committee of Nanjing Drum Tower Hospital (Ethical Approval Code: 2022-472-02) in accordance with the Declaration of Helsinki and registered in the National Medical Research Registration Information System (MR-32-25-006642). Written informed consent was obtained from each participant or legal guardian. All participants had their demographic information collected, blood sample collection, neuropsychological evaluation, and magnetic resonance imaging (MRI) scanning.

The inclusion criteria were: (1) age between 40 and 90 years; (2) right-handed; (3) compliance with the definition of normal cognition (NC), mild cognitive impairment (MCI), or dementia in this study. The Research flowchart was shown in Fig. 1.

Fig. 1.

Research flowchart. Abbreviations: MMSE, Minimum Mental State Examination; MoCA, Montreal Cognitive Assessment; ADL, Activities of Daily Living; HAMD, Hamilton Depression Rating Scale; HAMA, Hamilton Anxiety Rating Scale; MRI, magnetic resonance imaging; SNR, signal-to-noise ratio; BOLD: blood oxygen on level depending; 3D-T1, the high-resolution T1-weighted turbo gradient echo sequence.

The exclusion criteria were as follows: (1) some mental disorder, such as bipolar disorder, schizophrenia, and so on; (2) some acute states, such as acute cerebral infarction, acute cerebral hemorrhage, acute traumatic brain injury, migraine attack, onset of the vertigo, fever, and so on; (3) limited ability to cooperate with the study protocol, such as severe visual impairment, dysaudia, lalopathy, physical disability, and so on.

Demographic information collection and neuropsychological evaluation

The demographic data collection included sex, age, education, hypertension, diabetes, hyperlipidemia, smoking, and drinking. The neuropsychological evaluation included general cognitive function, and emotion. The assessment of general cognitive function contained MMSE, MoCA [9] (Beijing version 26 August 2006 translated by Wei Wang & Hengge Xie, www.mocatest.org), and Activities of Daily Living (ADL). The emotional evaluation contained the Hamilton Depression Rating Scale (HAMD) [10] and Hamilton Anxiety Rating Scale (HAMA) [11].

The education-adjusted cut-off values for the MMSE were > 17 for illiteracy, > 20 for 1‒6 years of education, and > 24 for > 6 years of education. The education-adjusted cut-off values for the MoCA were > 13 for illiteracy, > 19 for 1‒6 years of education, > 24 for 7‒12 years of education, and ≥ 26 for > 12 years of education.

The NC was defined as (1) both the MMSE and MoCA scores above the cut-off values adjusted for education; and (2) ADL = 14. The MCI was defined as (1) cognitive impairment detected by patients, informants, or clinical physicians; (2) having an MMSE score above the cut-off value adjusted for education and a MoCA score below its cut-off value adjusted for education; and (3) ADL = 14. The dementia was defined as (1) cognitive impairment detected by patients, informants, or clinical physicians; (2) having an MMSE score below the cut-off value adjusted for education; and (3) impaired performance in daily work and general activities (reported by the subjects or insiders).

Serum Raman spectroscopy

Raman spectral collection

Serum samples were collected on the same day as the neuropsychological evaluation. Blood samples in serum separator tube were centrifuged at 1200 g for 10 min after allowing blood to clot at room temperature for 10 min. Separated serum was collected into polypropylene tubes and frozen at −80℃ within 2 h after collection until serum Raman spectral collection was performed. Serum were applied to a low-background-noise chip and air-dried for approximately 10–15 min prior to testing. All spectra were acquired using a WITec alpha 300R + Raman confocal microspectrometer (Ulm, Germany) equipped with a piezoelectric stage (UHTS 300, WITec, GmbH.), 100X air objective lens (Zeiss EC EPIPLAN, N.A. = 0.9), green solid-state excitation laser (λ = 532 nm, 20 mW, WITec, GmbH.) and an imaging spectrograph (Newton, Andor Technology Ltd. UK) equipped with a 600 groove/mm grating and a thermoelectrically cooled (60℃) charged-coupled detector optically connected to the objective through a 10 μm diameter single mode silica fiber-optic cable. This setup enabled acquisition of spectral data across a wavenumber range from 400 to 3400 cm^− 1. In all cases, the excitation laser intensity was kept constant between sample scans, as well as integration time of 5s.

In Gulou cohort, our objective was to ultimately retain 50 valid spectra per subject. During spectral acquisition, Raman spectra preprocessing was performed concurrently (details in preprocessing were shown in Spectral preprocessing and Supplementary Material). Prior to the signal-to-noise ratio (SNR) calculation and filters, the target was to obtain 56 spectra processed through anomaly filters, spike removal, and baseline correction, so that after the SNR calculation and filters, 50 spectra would remain. For all samples, acquisition automatically stopped once 56 spectra processed through anomaly filters, spike removal, and baseline correction were collected; the maximum raw acquisition number was set to 70.

In the community cohort (the validation set), to improve time efficiency in practical applications, the target number of valid spectra was reduced to 20. Here, the goal was to collect 23 spectra before SNR calculation and filters, after which acquisition automatically stopped, with a maximum raw acquisition number of 40.

Spectral preprocessing

Raman spectra were analyzed by Python 3.9. Wavenumber and intensity calibration were conducted before testing real samples. In order to reduce unnecessary interference, the spectra was data-cleaned before statistical analyses by preprocessing, which included anomaly filters [12], spike removal [13], baseline correction, SNR calculation and filters, smoothing [14, 15], spectral normalization, and Raman silent region (1800 ~ 2700 cm^{− 1}) removal. Detailed information of Raman spectra preprocessing was shown in supplementary material.

Ensemble learning model construction.

The ensemble learning model was constructed by Python 3.9. During the modeling process, the libraries involved included sklearn, tensorflow, pytorch, scipy, matplotlib, seaborn, and so on. The objective of the ensemble learning model was to recognize NC, MCI, and dementia individuals in the subjects. In cases of ties (equal votes for multiple classes), the class with the highest priority order (NC > MCI > dementia) was selected as the definitive spectral label. Forty subjects recruited from communities were served as validation set. Synthetic minority over-sampling technique (SMOTE) was applied to the training set to balance the sample size among different cognitive gradings. SMOTE was implemented with default parameters (k_neighbors = 5) using the imbalanced-learn library, generating synthetic samples for minority classes until class balance was achieved. In image classification models and sequence models, all spectral data are included in the features. Otherwise, spectral data were dimensionality reduced through principal component analysis (50 dimensions), and the independence of variables was ensured.

In ensemble learning, we first trained several traditional machine learning models and deep learning models using 5-fold cross-validation on the training set for model development, parameter tuning and validation. Subsequently, all models were evaluated on the held-out testing set for performance assessment. The traditional machine learning models included support vector machine, linear discriminant analysis (LDA), random forest, gradient boosting, k-nearest neighbor classification, artificial neural network (ANN) (single layer), and so on. The 5-fold cross validation method was used for hyperparameter optimization to establish the optimal performance model. Deep learning models including multi-layer perceptron (MLP), gated recurrent units, visual geometry group network (VGGnet), and so on. Deep learning models were fine-tuned based on commonly used architectures. LDA, ANN, and MLP, which were the three most efficient in these models, were chosen for stacking in the ensemble learning model. During the modeling process, we strove to reduce the computational burden while ensuring the effectiveness of the model. For example, the parameters were reduced by stacking multiple small convolution blocks in the VGGnet.

After stacking, the predicting label of each spectrum were promoted by the majority vote. In cases of ties (equal votes for multiple classes), the class with the highest priority order (NC > MCI > dementia) was selected as the definitive spectral label. All Raman spectral labels of each subject were collected, and the label of the subject was obtained by the mode. Finally, independent external validation was conducted using Community cohort to assess the generalizability of the established ensemble learning model.

Graph theory analysis for Raman spectra

Raman spectroscopic data are characterized by high dimensionality and rich information content. Graph theory analysis facilitates the acquisition of novel spectral fingerprints, enabling the extraction of deep-level key information from complex spectral data while allowing focused investigation on specific nodes of interest. Preprocessed Raman spectral data were subjected to graph theory analysis. Each Raman shift was defined as a node, while the correlation of spectral intensities (Pearson correlation coefficient) between nodes was defined as edges. Using correlation coefficients as edge weights, a weighted network was constructed. During network construction, strong edges and nodes were filtered based on individual thresholds (P < 0.05, absolute correlation coefficient ≥ 0.8). Since this study focuses on nodal attributes, the following nodal metrics were calculated: clustering coefficient, degree centrality, closeness centrality, and betweenness centrality. The detailed information was shown in supplementary material.

MRI analysis

MRI sequences acquisition

The MRI scanning was conducted in the Nanjing Drum Tower Hospital with a Philips 3.0-T scanner (Philips Medical Systems, Netherlands). The multimodal MRI scans included (1) the high-resolution T1-weighted turbo gradient echo sequence, with repetition time (TR) = 8.2 ms, flip angle (FA) = 8◦, echo time (TE) = 3.8 ms, field of view (FOV) = 250 mm × 250 mm, number of slices = 192, acquisition matrix = 256 × 256, thickness = 1.0 mm, and (2) the blood oxygen on level depending (BOLD) sequence, TR = 2000ms, TE = 30 ms, number of slices = 35, voxel size = 3.00 mm × 3.00 mm × 4.00 mm, acquisition matrix = 64 × 62.

BOLD sequence preprocessing

DPARSF, a method based on the Resting-State Functional MR Imaging Toolkit (http://www.restfmri.net) and SPM (Statistical Parametric Mapping Software Package, www.fil.ion.ucl.ac.uk/spm) 12, was used for BOLD sequence preprocessing. The preprocessing steps included removing the first ten volumes of data, slice timing correction, realignment (subjects with head motion more than 3.0 mm of displacement in any direction, or 3.0 degrees of rotation in any angular dimension were excluded), reorientation, co-registration of T1 images to functional images, image segmentation for the high-resolution T1-weighted turbo gradient echo sequence and the BOLD sequence by DARTEL, normalization using a 12-parameter nonlinear transformation to the standard Montreal Neurological Institute (MNI) space (3 × 3 × 3 mm³), and smoothing functional images with a 4-mm full width at half maximum Gaussian spatial kernel. The segmentation and normalization of both structural and functional images were visually inspected and corrected manually if necessary.

Large-scale brain network analysis

Large-scale brain network analysis was performed based on spDCM. SpDCM differs from (the alternative) stochastic dynamic causal modelling by parameterising neuronal fluctuations using scale free (i.e., power law) forms, providing a constrained inversion of stochastic models, which enables the inference of direct (causal) connections between neural systems [16]. After preprocessing, 30 principal components were automatically recognized by GIFT (Group Independent Component Analysis fMRI Toolbox, https://trendscenter.org/software/gift/). Spatial distribution map and time series of principal components were estimated by group independent component analysis back reconstruction algorithm [17]. The infomax algorithm was iteratively executed 100 times using ICASSO (a built-in tool in GIFT) to ensure component reliability. The 30 principal components were then subjected to spatial correlation analysis with resting-state network templates, which contained 14 large-scale brain networks. Finally, five principal components were obtained: default mode network (peak MNI coordinate: 9, −82, 41), primary visual network (peak MNI coordinate: 9, −70, 8), somatomotor network (peak MNI coordinate: 0, −61, 56), left execution network (peak MNI coordinate: 48, −58, 47), posterior default mode network (peak MNI coordinate: 0, −70, 32).

Peak MNI coordinates of the five principal components were set to the center of the ROI spheres, with the radius of 6 mm. Following the regression of white matter signals (reference coordinate: 0, −24, −33), cerebrospinal fluid signals (reference coordinate: 0, −40, −5), and head motion parameters, the time series of ROIs were extracted and subsequently subjected to first-level DCM analysis.

SpDCM was constructed by SPM12, which was implemented to evaluate both direct (causal) connections between brain regions and intrinsic self-inhibition within each region (measured in Hertz). These connectivity parameters were subsequently employed in second-level DCM analysis. To ensure quality control of model inversion accuracy, models with variance ratio below 70% were excluded.

Correlation analysis between Raman intensities and spDCM was conducted through parametric empirical Bayes (PEB). PEB employs a two-level analysis, conducting statistical comparisons both within-group and between-group. Between-subject variability was estimated by a group-level general linear model implemented in the PEB framework, to which subject-specific connectivity indicators were transferred [8]. In this study, only connectivity effects with posterior probability exceeding 0.95 were reported. The statistical significance of effect sizes was assessed through free energy-based model comparison to validate the relationship between Raman intensities and connectivity strength.

Statistical analysis

Demographic data were analyzed using SPSS version 22.0. All continuous variables were tested for normality before comparison by observing whether the data points on the Probability-Probability Plot basically coincide with the diagonal. A Chi-Squared test or Fisher’s exact test was performed for the comparison of categorical variables. Analysis of variance (ANOVA), and Kruskal-Wallis tests were used to compare variations in continuous variables, in accordance with the results of the test of normality. A Spearman’s correlation or Pearson correlation were conducted to correlation analysis. Differences with P < 0.05 were considered statistically significant.

Results

Demographics and neuropsychological data

Gulou cohort

There were totally 220 subjects (93 subjects with NC, 70 subjects with MCI, and 57 subjects with dementia) in Gulou cohort. Demographically, there was no significant difference among subjects with NC, MCI, and dementia with regard to sex and education. The average age of subjects with NC was younger than that of MCI. Neuropsychologically, subjects with NC performed better in MMSE and MoCA than that of MCI and dementia (Table 1).

Table 1.

Participant characteristics in Gulou cohort

	GL-NC group (n = 93)	GL-MCI group (n = 70)	GL-dementia group (n = 57)	Statistics	P-value
Sex (male, %)	35 (37.6%)	33 (47.1%)	24 (42.1%)	1.487	0.476
Age (years)	66.753 ± 9.185	70.171 ± 7.173	73.702 ± 9.575	11.450	<0.001
Education (years)	12.000 (7.000)	12.000 (6.000)	12.000 (9.000)	3.084	0.214
Vascular risk factors
Hypertension, n (%)	33 (35.5%)	28 (40.0%)	35 (61.4%)	0.371	0.830
Diabetes, n (%)	13 (14.0%)	19 (27.1%)	10 (17.5%)	4.600	0.100
Hyperlipidemia, n (%)	14 (15.1%)	22 (31.4%)	6 (10.5%)	10.586	0.005
Smoking, n (%)	13 (14.0%)	8 (11.4%)	11 (19.3%)	1.607	0.448
Drinking, n (%)	10 (10.8%)	4 (5.7%)	4 (7.0%)	0.858	0.370
Neuropsychological evaluation
MMSE	29.000 (2.000)	28.000 (3.000)	19.000 (10.000)	128.385	<0.001
MoCA *	26.000 (2.000)	22.000 (6.000)	14.500 (8.000)	140.067	<0.001
HAMA #	3.000 (5.000)	5.000 (8.500)	4.000 (7.000)	5.986	0.050
HAMD #	3.000 (5.000)	4.000 (5.500)	3.000 (7.000)	1.893	0.371

Normal distribution data were analyzed by one-way analysis of variance and described as mean ± standard deviation expressed. Non-normal distribution data were analyzed by Kruskal Wallis H test and described as median (interquartile range). Categorical data were analyzed by Chi-square test or Fisher’s exact test

Abbreviations: GL Gulou cohort, NC normal cognition, MCI mild cognitive impairment, MMSE Mini-mental State Examination, MoCA Montreal Cognitive Assessment, HAMD Hamilton Depression Rating Scale, HAMA Hamilton Anxiety Rating Scale. *: Five subjects with dementia lacked MoCA; ^#: One subject with normal cognition, one subject with MCI, and four subjects with dementia lacked HAMA and HAMD

The community cohort

There are differences in the demographic characteristics between the community cohort and Gulou cohort. There were 14 subjects with NC, 16 subjects with MCI, and 10 subjects with dementia in the community cohort. Demographically, there was no significant difference among subjects with NC, MCI, and dementia with regard to sex, age and education. Neuropsychologically, subjects with NC performed better in MMSE and MoCA than that of MCI and dementia (Table 2).

Table 2.

Participant characteristics in the community cohort

	CM-NC group (n = 14)	CM-MCI group (n = 16)	CM-dementia group (n = 10)	Statistics	P-value
Sex (male, %)	5 (35.7%)	8 (50.0%)	5 (50.0%)	0.812	0.718
Age (years)	71.000 ± 6.839	66.688 ± 9.329	67.800 ± 8.290	1.059	0.357
Education (years)	10.000 ± 5.129	10.625 ± 4.829	9.400 ± 4.274	0.204	0.816
Vascular risk factors
Hypertension, n (%)	9 (64.3%)	4 (25.0%)	1 (10.0%)	7.910	0.005
Diabetes, n (%)	3 (21.4%)	5 (31.3%)	1 (10.0%)	0.287	0.636
Hyperlipidemia, n (%)	3 (21.4%)	7 (43.8%)	1 (10.0%)	0.168	0.821
Smoking, n (%)	1 (7.1%)	4 (25.0%)	2 (20.0%)	0.827	0.430
Drinking, n (%)	2 (14.3%)	2 (12.5%)	1 (10.0%)	0.094	1.000
Neuropsychological evaluation
MMSE	28.143 ± 2.070	27.750 ± 1.528	18.100 ± 6.173	30.990	<0.001
MoCA *	26.000 (3.250)	22.500 (6.000)	14.000 (9.500)	18.732	<0.001
HAMA #	1.000 (4.250)	6.500 (7.500)	2.000 (5.500)	6.231	0.044
HAMD #	2.500 (5.750)	10.500 (7.000)	2.000 (7.000)	7.288	0.026

Normal distribution data were analyzed by one-way analysis of variance and described as mean ± standard deviation expressed. Non-normal distribution data were analyzed by Kruskal Wallis H test and described as median (interquartile range). Categorical data were analyzed by Fisher’s exact test

Abbreviations: CM the community cohort, NC normal cognition, MCI mild cognitive impairment, MMSE Minimum Mental State Examination, MoCA Montreal Cognitive Assessment, HAMD Hamilton Depression Rating Scale, HAMA Hamilton Anxiety Rating Scale. *: One subject with dementia lacked MoCA. ^#: One subject with dementia lacked HAMA and HAMD

Raman spectral collection

After preprocessing, totally 10,877 Raman spectra for 220 subjects (approximately 49.4 Raman spectra per person on average) were retained for subsequent analysis in Gulou cohort. Considering the simplicity and efficiency in real-world applications, we reduced the collection of Raman spectra in the community cohort. After preprocessing, totally 800 Raman spectra for 40 subjects (20 Raman spectra per person on average) were retained for validation.

Unsupervised clustering of Raman spectra

Original Raman spectra with similar signal intensities were clustered together, ultimately forming three clusters using elbow method. Raman intensities were significantly different at 1002, 1602, and 1666 cm^− 1. Based on the Raman intensities, the three clusters were named low, medium, and high. The high cluster corresponds to higher peak intensities at 1002, 1602, and 1666 cm^− 1. After assigning cluster labels (low, medium, high) to all spectra, it was observed that the distribution of these three clusters differed across NC, MCI, and dementia groups. Specifically, the proportion of the low cluster gradually decreased with worsening cognitive performance, while the high cluster progressively increased in prevalence as cognitive impairment became more severe (Table S1., Fig. 2).

Fig. 2.

Unsupervised clustering of Raman spectra. Legends: The histogram illustrates the data as median with interquartile range. Detailed information was shown in supplementary material (Table S1). (A) Unsupervised clustering elbow plot; (B) Schematic diagram of the Raman spectra distribution in the low, middle, and high clusters in Gulou cohort; (C-E) The proportion of low, middle, and high cluster labels in the normal cognition, mild cognitive impairment, and dementia subjects in Gulou cohort; (H) The distribution of the Raman intensities of 1002, 1602, and 1666 cm^− 1 in the low, middle, and high clusters in Gulou cohort

Ensemble learning model

Final eigenvalue in the machine learning included (1) Raman spectra after normalization, (2) the proportion of Raman spectra belonging to high clusters in each serum sample, (3) highest peak of original Raman spectrum/reference peak (4800). The ensemble learning model was composed of LDA, ANN (single-layer), and MLP. In the testing set, the accuracy, average sensitivity, and average specificity was 0.85, 0.85, 0.92; the average area under curve was 0.92. In the validation set (the community cohort), the accuracy, average sensitivity, and average specificity was 0.80, 0.80, 0.89; the average area under curve was 0.89 (Table S2., Fig. 3, Table 3).

Fig. 3.

Ensemble learning model. Legends: (A-E) show the architecture of the ensemble learning model, the corresponding confusion matrices and ROC curves. The classification performance is evaluated by generating pseudo-probabilities based on the statistical distribution of multiple individual spectrum predictions within each subject. The subject-level classification performance is evaluated by generating pseudo-probabilities based on the statistical distribution of multiple individual spectrum predictions within each subject. Specifically, for each subject corresponding to multiple single spectra, the proportion of spectra predicted as each class (NC, MCI, or dementia) was calculated and used as pseudo-probabilities representing the likelihood of that subject belonging to each cognitive grade. These pseudo-probabilities were then used to generate ROC curves and calculate AUC values at the subject level. Abbreviations: NC, normal cognition; MCI, mild cognitive impairment; LDA, linear discriminant analysis; ANN, artificial neural network; MLP, multi-layer perceptron; ROC, receiver operating characteristic; AUC, area under curve

Table 3.

Group differences at 1602, 1002, and 1666 cm^− 1 in Gulou cohort

	GL-NC group (n = 93)	GL-MCI group (n = 70)	GL-dementia group (n = 57)	Statistics	P-value
Raman intensity at 1602 cm− 1	0.124 (0.024)	0.122 (0.023)	0.114 (0.028)	16.978	<0.001 a, b
Raman intensity at 1002 cm− 1	0.306 ± 0.042	0.296 ± 0.041	0.273 ± 0.041	11.280	<0.001 a, b
Raman intensity at 1666 cm− 1	0.282 ± 0.028	0.276 ± 0.028	0.259 ± 0.031	11.893	<0.001 a, b
Betweenness centrality at 1002 cm⁻¹	0.022 (0.033)	0.035 (0.048)	0.033 (0.066)	8.133	0.017 a
Clustering coefficient at 1666 cm⁻¹	0.810 (0.195)	0.815 (0.186)	0.762 (0.183)	7.550	0.023 a

The Raman intensities displayed here are the normalized Raman intensities input into the ensemble model. For each subject, the normalized spectra were averaged. The resulting mean value was used to represent the spectral intensity at each Raman shift for that subject. Normal distribution data were analyzed by one-way analysis of variance and described as mean ± standard deviation expressed (Bonferroni correction for multiple comparison correction). Non-normal distribution data were analyzed by Kruskal Wallis H test and described as median (interquartile range) (pairwise tests for multiple comparison correction)

Abbreviations: GL Gulou cohort, NC normal cognition, MCI mild cognitive impairment. ^a, Significant difference between GL-NC and GL-dementia groups; ^b, Significant difference between GL-MCI and GL-dementia groups

We began by observing the model’s high-weight features, which revealed 1602, 1002, and 1666 cm^− 1 as the top three weights in the LDA model. Raman intensities at 1602, 1002, and 1666 cm^− 1 progressively decrease across the subjects with NC, MCI, and dementia in Gulou cohort (Table 3).

Graph theory analysis for Raman spectra

We focused on the three Raman shifts with the highest contributions in the LDA model and calculated the clustering coefficient, degree centrality, closeness centrality, and betweenness centrality for the Raman shift at 1602, 1002, and 1666 cm^− 1 in the Gulou cohort. Significant inter-group differences were identified for two metrics: the betweenness centrality at 1002 cm⁻¹ was reduced in the subjects with MCI and dementia relative to the subjects with NC, and the clustering coefficient at 1666 cm⁻¹ was lowest in the subjects with dementia compared to subjects with NC and MCI (Table 3).

Large-scale brain network analysis

Large-scale brain network analysis was performed in Gulou cohort, in which totally 176 subjects received BOLD sequence scans. After quality control, 65 subjects with NC, 54 subjects with MCI, 35 subjects with dementia were included in the large-scale brain network analysis. Subjects with MCI or dementia were defined as subjects with cognitive impairment (CI). Considering the different age in subjects with NC, MCI, and dementia, age was added in PEB as covariates. Compared to subjects with NC, subjects with CI showed increased inhibitory influence of left executive control network (LECN) on posterior default mode network (pDMN), as well as decreased excitatory influence of dorsal attention network (DAN) on sensorimotor network (SMN), DAN on pDMN (Fig. 4). Detailed information was shown in supplementary material (Table S3.).

Fig. 4.

Differences in inter-network connectivity between the NC and CI groups in Gulou cohort. Legends: (A-D) Color indicates the size of the effect, while positive/negative corresponds to excitatory/inhibitory or increased/decreased. Additionally, the specific values of the effect sizes were also directly annotated. Abbreviations: NC, normal cognition; CI, cognitive impairment; DAN, dorsal attention network; PVN, primary visual network; SMN, sensorimotor network; LECN, left executive control network; pDMN, posterior default mode network

We also conducted large-scale brain network analysis between subjects with NC and dementia. Detailed information was shown in supplementary material (Figure S1., Table S4.).

Correlation analysis

The Raman intensities (after normalization, which were input into ensemble learning model) at 1602, 1002, and 1666 cm^− 1 had a strong positive correlation with each other. The correlation between effective connectivity and Raman intensities at 1602, 1002, and 1666 cm^− 1 was performed based on PEB. In the effective connectivity between networks with inter-group differences, Raman intensities at 1602 cm^− 1 were closely associated with effective connectivity of DAN on pDMN in subjects with NC. Moreover, Raman intensities at both 1002 and 1666 cm^− 1 in subjects with NC were related to effective connectivity of LECN on pDMN. Notably, these correlations, unique to subjects with NC, disappeared in subjects with CI. Additionally, in all subjects, Raman intensities at 1602 cm^− 1 were strongly negative-related to effective connectivity of primary visual network (PVN) on pDMN (effect size: −5.55) (Fig. 5).

Fig. 5.

Network analysis and correlation exploration in Gulou cohort. Legends: Color indicates the size of the effect. The specific values of the effect sizes were also directly annotated. (A-C) Correlation between the Raman intensities at 1602, 1002, 1666 cm^− 1; (D) Correlation between the Raman intensities at 1602 cm^− 1 and the connectivity strength of DAN on pDMN in the subjects with normal cognition (effect size = 6.19); (E) Correlation between the Raman intensities at 1002 cm^− 1 and the connectivity strength of LECN on pDMN in the subjects with normal cognition (effect size = −2.93); (F) Correlation between the Raman intensities at 1666 cm^− 1 and the connectivity strength of LECN on pDMN in the subjects with normal cognition (effect size = −3.98); (H) Correlation between the Raman intensities at 1602 cm^− 1 and the connectivity strength of PVN on pDMN in all subjects (effect size = −5.55). Abbreviations: NC, normal cognition; CI, cognitive impairment; DAN, dorsal attention network; PVN, primary visual network; SMN, sensorimotor network; LECN, left executive control network; pDMN, posterior default mode network

Compared to direct Raman peak intensity, the nodal properties of Raman shifts may emerge only in the later stages of cognitive impairment. We also explored the correlations between the betweenness centrality at 1002 cm⁻¹, the clustering coefficient at 1666 cm⁻¹, and the functional connectivity between networks. Detailed information was shown in supplementary material.

Discussion

In this manuscript, we promoted that Raman spectroscopy combined with ensemble learning model serves as a powerful tool for cognitive screening. This study presents the first Raman-based tool capable of comprehensive staging across the entire cognitive impairment progression. Raman shifts at 1602, 1002, and 1666 cm⁻¹ exhibited changes in intensity and nodal properties, which were associated with impaired cognition and altered pDMN-related inter-network interactions. Moreover, to further capture systemic pathological alterations, we applied graph theory to serum Raman spectra, rather than focusing on isolated peaks. This provides a new framework for deciphering the systemic state encoded in the spectra.

Lacking blood-based alternatives to neuropsychological scales has been a major challenge in cognitive screening in China for decades. Previous studies focused on applying Raman spectroscopy to single diseases or specific stages, such as AD dementia stage or Parkinson’s disease, resulting in difficulties in application for graded cognitive screening to the communities. This study firstly developed a Raman spectroscopy-based model for community-wide cognitive screening that covers NC, MCI, and dementia. By combining brief medical history inquiries with serum testing, this approach enables rapid grading of cognitive status, offering a feasible new pathway for large-scale early detection of cognitive impairment. This study was designed with a focus on practical applications. However, the reliance on self-reported medical histories may have limited accuracy, potentially contributing to the model’s accuracy plateauing at 80–85%. Nevertheless, in the validation set, the model still demonstrated robust identification performance despite the reduction in the number of spectra.

Recently, there has been growing interest in optically derived biomarkers for neurological diseases. For instance, in a 2025 study, Bin et al. developed models based on Attenuated Total Reflectance–Fourier Transform Infrared plasma spectroscopy to distinguish AD dementia and MCI from NC. In their feature engineering process, particular emphasis was placed on selecting spectral features guided by the A-T-N in AD [18]. Our study, in contrast, focuses on the application of spectroscopy in cognitive screening and recognizes the inherent value of spectral data as a data asset. Therefore, when performing feature selection on serum Raman spectra, we prioritized intrinsic spectral characteristics through detailed spectral analysis. Moreover, unlike previous studies that primarily addressed binary classification, our model performs a three-class classification task. Compared to binary classification, the three-class task is more challenging in terms of model performance and requires a higher level of refinement in data processing.

The Raman intensities at 1602, 1002, and 1666 cm^− 1 served as key features for unsupervised clustering and were the top three most important features in the cognitive staging diagnostic model, where the 1602 cm^− 1 carried the highest weight.

The Raman shift at 1602 cm^{− 1} is named “Raman spectroscopic signature of life”, which was first identified in yeast cells and linked to mitochondrial metabolism [19–21]. The Raman shift at 1602 cm^{− 1} is considered as a biomarker of cellular metabolic activity, whose intensity sensitively specifically reflects the metabolic state in the cell and closely tied to glucose metabolism [22, 23]. Glucose is an essential energy source for the brain. The impairment of ATP-dependent processes essential for cognition leads to synaptic dysfunction and neuronal death, ultimately causing the structural and functional alteration in critical brain regions [24]. Researchers have observed through FDG-PET that glucose metabolism continues to decline in patients with MCI and dementia. This phenomenon is present across various types of cognitive disorders, such as Alzheimer’s disease-related cognitive impairment, vascular cognitive impairment, as well as Lewy body dementia, and so on [24–26].

1002 cm^{− 1} is the representative Raman shift for phenylalanine, which reflects C-C aromatic ring stretching and relates to collagen assignment. Phenylalanine is closely related to central nervous system diseases. Phenylalanine-tyrosine metabolism, were associated with the status of arteriosclerotic cerebral small vessel disease [27]. The intensity of the 1002 cm⁻¹ Raman shift is a key fingerprint for assessing the level of neuronal damage and intervention potential in middle cerebral artery occlusion rats [28]. For cognitive function, both high and low levels of phenylalanine are associated with adverse effects. On one hand, phenylalanine exhibits neurotoxicity, such as phenylketonuria, in which the excessive accumulation of phenylalanine severely impairs cognitive function [29]. Phenylalanine disrupts the transit of amino acids across the blood-brain barrier by competitively inhibiting the uptake of large neutral amino acids—such as tyrosine and tryptophan—thereby compromising neurotransmitter synthesis. Moreover, phenylalanine itself exhibits neurotoxic properties: it suppresses adenosine triphosphate synthesis, elevates reactive oxygen species (ROS) production, promotes neuronal apoptosis, and impairs both myelination and myelin maintenance, culminating in dysmyelination. On the other hand, there is some controversy surrounding the changes in peripheral blood phenylalanine levels in patients with cognitive impairment. AD patients had higher D-phenylalanine in the peripheral blood serum than healthy controls [30]. Researchers at Italy pointed out that Alzheimer’s disease, frontotemporal dementia, vascular dementia, and MCI patients might share similar deficits in amino acid precursors, which was caused by phenylalanine reductions [31]. However, in our study, the Raman intensity (after normalization) at 1002 cm^{− 1} was positively correlated with cognitive performance. Similarly, a 5-year follow-up study of 1440 subjects in Shanghai (China) found that phenylalanine levels in the serum were negatively correlated with the incidence of dementia [32]. In a Raman spectroscopy study, the intensity of the phenylalanine-related Raman shift was increased in Lewy body dementia patients compared to healthy controls, whereas it was slightly lower in the late AD patients compared to healthy controls [5]. Reduced peripheral blood phenylalanine levels could be a potential biomarker for high risk of cognitive impairment in middle-aged and older adults.

In the study of neurology, the Raman shift at 1666 cm⁻¹ is seldomly reported. The Raman shift at 1666 cm⁻¹ falls within the protein Amide I, which arise as a result of the polypeptide backbone vibration of a random coil. Subjects with cognitive impairment performed lower Raman intensity at 1666 cm⁻¹ than normal controls in our study. Similarly, AD patients had lower intensity at the Raman shift at Amide I (1650 cm⁻¹) than the healthy controls.

Overall, the decreased intensities in 1602, 1002, and 1666 cm⁻¹ Raman shift are important Raman-biomarkers for cognitive impairment. Our findings implicate the Raman shifts at 1602, 1002, and 1666 cm⁻¹ in cognitive impairment, where progression of cognitive impairment was associated with both their intensity changes and a concomitant rewiring of the spectral network, as evidenced by altered hub properties of the 1002 and 1666 cm⁻¹ peaks. Graph theory analysis endows Raman spectroscopy with a new dimension of molecular network fingerprint information. Notably, the application of Raman spectroscopy in medicine is still in its infancy. Thus, we seek to advance Raman spectroscopy from a mere screening tool for cognitive impairment to a method capable of reflecting brain function and correlating with neuroimaging metrics.

Compared to traditional functional brain network analyses, spDCM is more capable of revealing the directionality of interactions between networks. Patients with cognitive impairment exhibit altered interaction of the pDMN by the DAN and LECN. Correlation analysis revealed that in patients with CI, the Raman intensities at 1602, 1002, and 1666 cm⁻¹ were directionally associated with the alteration of the pDMN. Although no significant group difference was found in PVN on pDMN between NC and CI subjects, it also showed a strong correlation with the Raman intensity at1602 cm⁻¹. The Raman intensities at 1602, 1002, and 1666 cm⁻¹ demonstrate potential as peripheral blood indicators of pDMN function. The pDMN involves regions such as the posterior cingulate cortex, precuneus, medial temporal lobe, and angular gyrus, and is primarily responsible for episodic memory, scene construction, and spatial navigation. Modulation of the pDMN—whether inter of intra functional network alterations—is a key clinical manifestation in various types of CI, particularly AD and vascular cognitive impairment [33–38]. Thus, the pDMN has been an effective target for repetitive transcranial magnetic stimulation therapy for AD [34]. Current treatments for cognitive impairment remain inadequate [39, 40]. Neuromodulation therapies, such as transcranial magnetic stimulation and temporal interference stimulation, are gaining acceptance in Chinese populations due to their non-invasive nature. These approaches have demonstrated promising potential to modify progression of CI and improve clinical outcomes. Personalized neuromodulation is key to enhancing efficacy, but tailoring protocols typically requires functional MRI to assess brain network dynamics [41]. However, a significant proportion of individuals with cognitive impairment in community settings across China do not undergo MRI. Should serum Raman spectroscopy prove capable of gauging posterior default mode network integrity, it could guide targeted neuromodulation, enabling personalized therapeutic strategies and optimizing the allocation of healthcare resources.

A key strength of our study is the development of an efficient and accessible blood-based screening approach for cognitive assessment. It is important to note several limitations. Our study did not include left-handed participants, nor did it pursue the etiological classification of cognitive impairment. Our future work will focus on the development of an integrated system to streamline the workflow from detection to analysis, and the execution of large-scale, multi-center clinical studies to expand the sample size and refine the model.

In conclusion, the Raman spectroscopy-based ensemble learning model serves as a powerful tool for cognitive screening, effectively distinguishing individuals with NC, MCI, and dementia. The decreased intensities of the 1602, 1002, and 1666 cm⁻¹ represent key Raman signatures of cognitive impairment, potentially reflecting impaired pDMN-related inter-network interactions.

Authors’ contributions

Yuting Mo and Chenglu Mao wrote the main manuscript text, prepared figures and tables, performed statistical analysis. Jialiu Jiang, Shuang Fang, Zhihong Ke, Zheqi Hu, Pei Xie, and Ruozhu Xiong was responsible for data collection for the research. Lili Huang, and Dan Yang performed statistical analysis. Yun Xu was primarily responsible for research design. All authors reviewed the manuscript.

Funding

This work was funded by the STI2030-Major Projects (2022ZD0211800), Basic Research Program of Jiangsu (BK20250239), Changping National Laboratory Research Project (2025B-07-31), the National Natural Science Foundation of China (82130036), the Jiangsu Province Key Medical Discipline (ZDXK202216), the Youth Training Program of the National Natural Science Foundation of Nanjing Drum Tower Hospital (2024-JCYJ-QP-40).

Data availability

The datasets used and analyzed in this study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This research complied with the ethical standards of the relevant institutional and/or national research committee, the 1964 Helsinki Declaration and its subsequent amendments, as well as all comparable ethical standards. Written informed consent was obtained from all individual participants enrolled in this study or their legal guardians. The study protocol was approved by the Ethics Committee of Nanjing Drum Tower Hospital (Ethical Approval Code: 2022-472-02).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.