Authors: Ingyu Park, Sang-Kyu Lee, Hui-Chul Choi, Moo-Eob Ahn, Ohk-Hyun Ryu, Yong Soon Park, Dong-Kyu Kim, Unjoo Lee, Yeo Jin Kim
Categories: Article, Machine learning, Mild cognitive impairment, Gait, Body composition, Sleep, Digital biomarker, Neuroscience, Neurology
Source: Scientific Reports
Authors: Ingyu Park, Sang-Kyu Lee, Hui-Chul Choi, Moo-Eob Ahn, Ohk-Hyun Ryu, Yong Soon Park, Dong-Kyu Kim, Unjoo Lee, Yeo Jin Kim
Mild cognitive impairment (MCI) is a prodromal stage of dementia, and its early detection is critical for improving clinical outcomes. However, current diagnostic tools such as brain magnetic resonance imaging (MRI) and neuropsychological testing have limited accessibility and scalability. Using machine-learning models, we aimed to evaluate whether multimodal physical and behavioral measures, specifically gait characteristics, body mass composition, and sleep parameters, could serve as digital biomarkers for estimating MCI severity. We recruited 80 patients diagnosed with MCI and classified them into early- and late-stage groups based on their Mini-Mental State Examination scores. Participants underwent clinical assessments, including the Consortium to Establish a Registry for Alzheimer’s Disease Assessment Packet Korean Version, gait analysis using GAITRite, body composition evaluation via dual-energy X-ray absorptiometry, and polysomnography-based sleep assessment. Brain MRI was also performed to obtain structural imaging data. We evaluated the classification performance across various models, including support vector machines, random forest, multilayer perceptron, and convolutional neural network, using unimodal and multimodal datasets. Machine learning models trained on physical and behavioral data alone achieved a high classification accuracy (AUC up to 94%), comparable to that of MRI-based models, in differentiating early- and late-stage MCI. Combining physical and behavioral and MRI features yielded marginal improvements in the prediction performance. Gait velocity, lean body mass, and sleep efficiency were among the top predictors of cognitive function. Multimodal digital biomarkers or multimodal physical and behavioral signals can effectively estimate MCI severity and may offer a scalable, low-cost approach for early detection and monitoring of cognitive decline in real-world settings.
The online version contains supplementary material available at 10.1038/s41598-025-19364-1.
Mild cognitive impairment (MCI) refers to a stage of cognitive decline that precedes dementia and is characterized by noticeable cognitive impairment while maintaining the ability to perform daily activities^1^. In clinical practice, the ability of patients with MCI to preserve daily functioning despite cognitive deterioration is closely linked to their quality of life and long-term prognosis. Consequently, the early detection of cognitive decline is essential, as it facilitates timely treatment and improves patient outcomes.
Recent advances in digital technology have enabled the measurement of individualized health data using wearable and mobile devices, thereby facilitating real-time health monitoring. As the interest in preventive healthcare grows, there is increasing research on the use of digital tools to measure health indicators for the assessment and prediction of various health conditions. Standardized neuropsychological tests provide relatively accurate assessments of cognitive function. However, these tests are time intensive, typically requiring 1.5 to 2 h—and are not conducive to continuous or real-time monitoring^2^.
Given these limitations, there is growing interest in identifying digital biomarkers that allow for simpler and more accessible assessments of cognitive status^3^. A well-designed health indicator is a critical tool for assessing an individual’s health status, and selecting appropriate indicators based on specific objectives is critical for effective assessment^4^. Health indicators encompass a wide range of factors, including clinical and physical and behavioral measures, as well as socioeconomic and behavioral components. Among these, physical and behavioral measures that are objective, quantifiable, and easily obtainable in everyday life are particularly promising for digital cognitive assessment.
In this study, we focused on three non-imaging physical and behavioral measures—gait, body composition, and sleep—which have been previously associated with cognitive function^5–7^. Compared to imaging-based markers, which are widely used in research but remain costly and less accessible, these measures offer a practical alternative. Therefore, we aimed to investigate whether these easily measurable health indicators could serve as digital biomarkers for predicting the degree of cognitive decline and reflect structural changes in the brain.
We prospectively recruited 80 patients with MCI at Chuncheon Sacred Heart Hospital between October 2020 and April 2021. The inclusion criteria were as (a) age 40–100 years; (b) patients with MCI met the diagnostic criteria for minor neurocognitive disorder according to the Diagnostic and Statistical Manual of Mental Disorders Fifth Edition (DSM-V); (c) absence of dementia as assessed by physicians; and (d) a Clinical Dementia Rating (CDR) global score of 0.5, with a memory domain score ≥ 0.5. The exclusion criteria were as (a) severe illness with an anticipated fatal outcome within 3 months, (b) language barrier, (c) deafness or blindness, and (d) inability to provide informed consent. Patients were classified into two groups based on their Mini-Mental State Examination (MMSE) (a) late-stage MCI, defined as performance more than 1.5 standard deviation (SD) below the normative mean; and (b) early-stage MCI, defined as performance ≥ 1.5 SD below the normative mean. Written informed consent was obtained from each participant, and the study protocol was approved by the Institutional Review Board (IRB) of Chuncheon Sacred Heart Hospital (IRB Chuncheon 2020–09-005). The study was conducted in accordance with the Declaration of Helsinki, and all methods were performed in accordance with the relevant guidelines and regulations.
The demographic data (age, sex, and years of education) and anthropometric measurements (height, weight, and waist circumference) were recorded. Blood pressure was measured and laboratory tests, including fasting glucose and total cholesterol levels, were conducted. Participants completed scales assessing depression (Short Form of Geriatric Depression Scale, SGDS), anxiety (Korean Geriatric Anxiety Inventory, K-GAI), and quality of life (Geriatric Quality of Life Dementia, GQOL-D). Additionally, information of health behaviors and medical comorbidities was obtained using the questionnaire from the Korean National Health and Nutrition Examination Survey (KNHANES). Information on other medical comorbidities was obtained during in-clinic interviews. The comorbidities considered in this study included diabetes mellitus, hypertension, dyslipidaemia, heart disease, stroke, liver disease, depression, kidney disease, pulmonary disease, arthritis, chronic gastrointestinal disease, cancer, and fractures.
All participants underwent the Consortium to Establish a Registry for Alzheimer’s Disease Assessment Packet, Korean version (CERAD-K) and the Short Blessed Test (SBT-K)^8,9^ for cognitive evaluation. The CERAD-K consists of Verbal Fluency, the Boston Naming Test, the Mini-Mental State Examination Korean version (MMSE-KC)^10^, Word List Memory, Constructional Praxis, Word List Recall, Word List Recognition, Constructional Praxis Recall, and the Trail Making Test (Parts A and B)^10,11^. All cognitive assessments were conducted during the participant’s two-day hospital stay, either on the same day or within one day before/after the gait, sleep, MRI, and body composition measurements.
Gait was assessed using GAITRite instrumentation (CIR systems Inc., Havertown, PA, USA), which consists of a 5.6 m long and 0.9 m wide electronic walkway. Each patient was instructed to walk across the walkway at a normal pace without gait aid in a single pass. The study coordinator observed each patient’s gait without any interference. Gait features, including spatial (stride and step length), temporal (gait speed, step count, cadence, and stance time), and spatiotemporal (cadence) parameters, were analyzed.
Body composition, including body fat percentage and appendicular skeletal muscle mass, was assessed using dual-energy X-ray absorptiometry (DXA; Lunar, GE Healthcare, Fairfield, CT, USA).
Polysomnography (PSG) was performed using the Nox A1 device (Nox Medical, Reykjavik, Iceland). Sleep stage and respiratory events were scored according to the American Academy of Sleep Medicine guidelines^12,13^. The following sleep parameters were total sleep time (TST), sleep efficiency (TST/time in bed × 100), sleep stage (N1, N2, N3, and R), oxygen desaturation index (ODI), apnea index (AI), hypopnea index (HI), and apnea–hypopnea index (AHI). Definitions of oxygen desaturation, apnea, and hypopnea followed standard guidelines, and the indices were computed for each participant.
Standardized T2 fluid-attenuated inversion recovery (FLAIR) and three-dimensional (3D) T1-weighted images were acquired using a 3.0 T MRI scanner (Skyra, Siemens Healthineers, Erlangen, Germany) at the Chuncheon Sacred Heart Hospital. T1-weighted structural brain images were obtained using a magnetization-prepared rapid acquisition gradient echo (MPRAGE) sequence with the following sagittal slice thickness, 1.0 mm (no gap); repetition time (TR), 2300 ms; echo time (TE), 2.98 ms; flip angle, 9°; inversion time (TI), 900 ms. FLAIR images were acquired with an axial slice thickness of 2 mm, TR of 11,000 ms, TE of 125 ms, and a flip angle of 90°.
The total volume and other morphometric features of the grey matter (average cortical thickness [ThickAvg], standard deviation of cortical thickness [ThickStd], mean curvature [MeanCurv], Gaussian curvature [GausCurv], folding index [FoldInd], and curvature index [CurvInd]) were measured using 3D Slicer (version 4.11) and FreeSurfer (version 7.1.1). Cortical parcellation was conducted using the Desikan–Killiany atlas (lh.aparc.annot and rh.aparc.annot for the left and right hemispheres, respectively)^14^, and each hemisphere was divided into 34 distinct cortical regions (68 regions in total). The subcortical structures were segmented based on 3D T1-weighted MRI data.
White matter hyperintensities (WMH) were quantified using the Brain Intensity Abnormality Classification Algorithm (BIANCA) software within the FMRIB Software Library (FSL). FLAIR MRI images were used for the segmentation and classification of the WMH. The non-default settings included a lesion model trained using the MICCAI WMH segmentation challenge dataset. Features derived from WMH segmentation included the total WMH volume, periventricular WMH, and deep WMH volumes, which were further categorized by the four lobes of the brain.
Data on gait characteristics, body composition, sleep, and MRI scans were collected at baseline. The following unimodal datasets were (a) Gait Dataset: derived from GAITRite (74 features per participants); (b) Gray Matter Dataset: derived from MRI (476 features per participant); (c) White Matter Dataset: derived from MRI (27 features per participant); (d) Body Composition Dataset: derived from DXA (37 features per participant); (e) Sleep Dataset: derived from polysomnography (12 features per participant). Three multimodal datasets were constructed by combining the following unimodal Gait + Sleep + Body Composition, Gray Matter + White Matter Hyperintensity, and Gait + Sleep + Body Composition + Gray Matter + White Matter Hyperintensity. Machine learning analyses were conducted on these multimodal datasets.
Descriptive statistics were computed for baseline characteristics, with continuous variables presented as mean ± standard deviation (SD) and categorical variables as percentages. Differences between the early and late-stage MCI groups were assessed using Student’s t-tests were applied for normally distributed continuous variables and the Mann–Whitney U test was applied for non-normally distributed continuous variables. Categorical variables were compared using the chi-squared test. Demographic and clinical characteristics were analysed using IBM SPSS Statictics, version 27 (IBM corp., Armonk, NY, USA).
Pearson’s correlation coefficient (r) was calculated to examine the linear relationship between the cognitive scores and each lifelog feature. Spearman’s rank correlation coefficient (ρ) was also calculated to assess any potential monotonic associations, particularly for variables that were not normally distributed. Data normality was assessed using the Kolmogorov–Smirnov test, and outliers, defined by Z-scores exceeding ± 3, were excluded. Both correlation coefficients and their corresponding p-values are reported, with p-values adjusted using False Discovery Rate (FDR) correction. The analysis was conducted using the SciPy library in Python 3.8, with the statistical significance set at p < 0.05.
We conducted exploratory data analysis to assess data completeness, distributional properties, and potential collinearity among features before implementing the independent component analysis (ICA), principal component analysis (PCA), or machine learning algorithms. No missing values were present across the modalities (gait, 74 features; body composition, 37 features; grey matter on MRI, 476 features; and sleep, 12 features). Data normality was assessed using the Kolmogorov–Smirnov test, and variables were standardised using z-score normalisation to account for different scales and distributions. Histograms and density plots were visually inspected for skewness and multimodality. While some skewed distributions were observed, no clear bimodal patterns indicative of latent subgroups were found. Collinearity was addressed implicitly using ICA and PCA, which transformed the correlated variables into orthogonal components.
Two unsupervised machine learning (ML) algorithms, independent component analysis (ICA) and principal component analysis (PCA), were used for feature reduction. We tested 20, 40, and 60 reduced features for each of the three multimodal datasets. Supervised ML algorithms—support vector machine (SVM), random forest (RF), multilayer perceptron (MLP), and convolutional neural network (CNN)—were trained for classification using these reduced features. For the CNN, a one-dimensional convolutional neural network (1D CNN) was implemented. Each subject’s feature vector, which was reduced using PCA/ICA, was reshaped into an array of [n_components, 1], before being fed into two Conv1D layers with ReLU activation, followed by dropout, max pooling, a fully connected hidden layer, and a sigmoid output unit for binary classification. The detailed hyperparameters and architectural specifications are summarised in Supplementary Table S1. The dataset was split into 80% training and 20% test sets with tenfold cross-validation for model evaluation. Hyperparameter tuning was performed using 60 variations of each algorithm. The model performance was evaluated based on the area under the receiver operating characteristic curve (AUC), accuracy, recall, precision, and F1 score. The computations were performed on an Ubuntu 18.04 system with an Intel Core i9-9820X CPU (10 cores), a 64 GB memory, and an NVIDIA GTX 1080 Ti GPU.
A multiple regression model was employed to predict the cognitive scores based on independent variables from the multimodal datasets by incorporating features derived from ICA and PCA. The model can be expressed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ Y_{ik} = \beta_{0} + \sum\limits_{j = 1}^{{m_{k} }} {\beta_{j} X_{ij} } + \varepsilon_{ik} ,
\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ Y_{ik} $$\end{document} represents the ith cognitive score of the kth multimodal dataset, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ X_{ij} $$\end{document} represents the jth feature associated with the ith cognitive score of the kth multimodal dataset, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \beta_{0} $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \beta_{j} $$\end{document} are the intercept and regression coefficients, respectively, indicating the strength of each feature’s effect on the ith cognitive score of the kth multimodal dataset, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \varepsilon_{ik} $$\end{document} is an error term that accounts for the unobserved factors influencing the ith cognitive score of the kth multimodal dataset, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ m_{k} $$\end{document} is the total number of features in the kth multimodal dataset. Multiple regression analysis was performed using the statsmodels library in R. The regression results were further analyzed in the context of network structure variables, such as centrality and network position, using Cytoscape version 3.9.1. Cytoscape provides advanced tools for visualizing and integrating network data. ## Results ### Baseline characteristics Detailed demographic and clinical characteristics of the participants are presented in Table 1. The early-stage MCI group exhibited a higher proportion of females as than the late-stage MCI group. The years of education were shorter in the early than in the late-stage MCI group. Additionally, the early-stage MCI group had more comorbidities than the late-stage MCI group. Body mass index (BMI) was higher in the early than in the late-stage MCI group.Table 1Demographic and clinical characteristics.Total (n = 80)Early-stage MCI (n = 53)Late-stage MCI (n = 27)*p* ValueMean age74.6 ± 5.7474.6 ± 5.3374.6 ± 6.580.760^c^Gender, female (number, %) ^a^58 (72.5)43 (81.1)15 (55.6)0.020Year of education6.88 ± 3.9805.94 ± 4.0698.72 ± 3.1140.003^c^Number of comorbidities ^b^4 (0–8)4 (0–8)3 (0–7)0.049^d^Depression scale ^b^4 (0–14)4 (0–14)4 (1–14)0.251^d^Anxiety scale ^b^4.5 (0–20)5 (0–20)4 (0–20)0.674^d^QOL scale34.4 ± 8.6733.3 ± 8.3936.5 ± 9.000.124Height, cm156.1 ± 8.11155.4 ± 7.98157.7 ± 8.300.232Weight, kg61.0 ± 10.9261.9 ± 11.4759.3 ± 9.720.324BMI (mean, SD)25.0 ± 3.7025.6 ± 3.8423.8 ± 3.170.045Waist circumference (cm)87.8 ± 10.1287.6 ± 10.3988.1 ± 9.750.825Systolic BP (mmHg)128.6 ± 18.27130.0 ± 18.44126.0 ± 17.970.349Diastolic BP (mmHg)77.0 ± 9.8277.4 ± 9.4176.1 ± 10.730.606Fasting glucose (mg/dL)108.8 ± 26.28111.3 ± 28.44104.0 ± 21.080.240Total cholesterol (mg/dL)162.4 ± 32.5434.0 ± 4.6730.1 ± 5.790.913Data are mean ± SD unless otherwise indicated.^a^Number (%).^b^Median (Range).^c^Variables not meeting normality assumptions were compared using the Mann–Whitney U test.^d^Variables compared using the Chi-squared test.*MCI* Mild cognitive impairment; *QOL* Quality of life; *BMI* Body mass index; *BP* Blood pressure; *SD* Standard deviation. ### Correlation between physical and behavioral measures and cognitive functions Figure 1 and Table 2 show the gait, body composition, and sleep parameters most strongly correlated with each cognitive function score. Overall, a longer stride length, faster gait velocity, and greater single support were all positively associated with better cognitive performance, whereas greater variability, increased step count, and longer double support were negatively associated. In Table 2, both Pearson’s *r* and Spearman’s ρ values are reported for the direct comparison between linear and rank-based correlations. The top five measures most strongly associated with each cognitive score are presented in Supplementary Table S2. Notably, features showing strong correlations also overlapped with the PCA components and ML model importance rankings, indicating consistency across analytic approaches (Supplementary Table S3). For reference, the corresponding Spearman correlation plots are shown in Supplementary Fig. S1.Fig. 1Heatmaps illustrating the correlation between cognitive function scores and (**a**) gait parameters, (**b**) body mass composition, and (**c**) sleep parameters. Red represents positive correlations, blue represents negative correlations, and the color gradient indicates the strength of the correlation. Asterisks denote statistical *p* < 0.05 (*), *p* < 0.01 (**). Abbreviations: MMSE-KC, mini-mental state examination, Korean version; SBT-K, short blessed test, Korean version; CERAD-K, Consortium to Establish a Registry for Alzheimer’s Disease Assessment Packet Korean Version; L, Left; R, Right; Supp, Support; GC, Gait Cycle; Std Dev, Standard Deviation; Len, Length; BMC, Bone Mineral Content; BMD, Bone Mineral Density; AHI, Apnea-Hyponea Index; AI, Apnea Index; ESS, Epworth Sleepiness Scale; HI, Hypopnea Index; NREM, Non-Rapid Eye Movement sleep; RDI, Respiratory Disturbance Index; REM, Rapid Eye Movement sleep; RERA, Respiratory Effort-Related Arousal.Table 2Top correlated gait, body composition, and sleep parameters for each cognitive score.Positively correlatedPearson’s rSpearman’s ρ*p* Value for r*p* Value for ρNegatively correlatedPearson’s rSpearman’s ρ*p* Value for r*p* Value for ρSBT-KDouble Support Load %GC R0.3850.3420.0020.009Stride Velocity Left − 0.392 − 0.4140.0030.003MMSE-KCStep Length(cm) R0.3910.4390.0010.003Step Count − 0.369 − 0.4610.0050.001Total score of CERAD-K––––Word List Memory––––Word List Recall––––Word List Recognition––––Constructional Praxis––––Constructional Praxis RecallStride Length(cm) R0.3900.3790.0030.006Step Count − 0.361 − 0.3610.010.007Verbal FluencyStride Length(cm) L0.3790.4170.0020.003Step Count − 0.384 − 0.3760.0040.007Boston Naming TestStep Length(cm) R0.3840.3980.0070.003Step Time Std Dev R − 0.370 − 0.4110.010.003*SBT-K* Short blessed test, Korean version; *MMSE-KC* Mini-mental state examination, Korean version; *CERAD-K* Consortium to Establish a Registry for Alzheimer’s Disease Assessment Packet Korean Version; Double Support Load %GC, double support load percentage of gait cycle; *R* Right side; *L* Left side; *std dev* Standard deviation. ### Correlation between brain structures and cognitive functions Figure 2 and Table 3 present the brain MRI parameters that were most strongly correlated with each cognitive function score. Significant correlations were observed among cognitive measures, grey matter morphometry, and white matter hyperintensity burden across multiple regions. Detailed results, including the top five structural features per cognitive domain, are provided in Supplementary Table S4. Both Pearson’s *r* and Spearman’s ρ values are reported to enable the comparison of linear and rank-based associations. For reference, the corresponding Spearman correlation plots are shown in Supplementary Fig. S2.Fig. 2Heatmaps illustrating the correlation between cognitive function scores and brain structural (**a**) gray matter parameters and (**b**) white matter hyperintensity parameters. Red represents positive correlations, blue represents negative correlations, and the color gradient indicates the strength of the correlation. Asterisks denote statistical *p* < 0.05 (*), *p* < 0.01 (**). Abbreviations: MMSE-KC, mini-mental state examination, Korean version; SBT-K, short blessed test, Korean version; CERAD-K, Consortium to Establish a Registry for Alzheimer’s Disease Assessment Packet Korean Version; lh, left hemisphere; rh, right hemisphere; GrayVol, Gray matter Volume; ThickAvg, Thickness Average; ThickStd, Thickness Standard deviation; MeanCurv, Mean Curvature; GausCurv, Gaussian Curvature; FoldInd, Folding Index; CurvInd, Curvature Index; WMH, White Matter Hyperintensity; LV, Lateral Ventricle.Table 3Top correlated brain structural changes for each cognitive score.Positively correlatedPearson’s rSpearman’s ρ*p* Value for r*p* Value for ρNegatively correlatedPearson’s rSpearman’s ρ*p* Value for r*p* Value for ρSBT-Krh_occipital(total WMHs)0.2830.2730.2930.030GrayVol_lh.rostralanteriorcingulate − 0.356 − 0.3190.0360.040MMSE-KCGrayVol_lh.caudalanteriorcingulate0.3950.3810.0320.024MeanCurv_lh.temporalpole − 0.345 − 0.2870.0360.053Total score of CERAD-KCurvInd_lh.bankssts0.2930.1340.0480.381CurvInd_lh.lingual − 0.2250.040.1050.809Word List MemoryFoldInd_lh.bankssts0.3200.0950.0380.563CurvInd_lh.lingual − 0.254 − 0.050.0720.775Word List RecallFoldInd_lh.bankssts0.3200.0030.0380.983CurvInd_lh.lingual − 0.272 − 0.080.0590.634Word List RecognitionFoldInd_lh.bankssts0.3220.030.0380.853CurvInd_lh.lingual − 0.261 − 0.1630.0660.277Constructional PraxisGrayVol_lh.parsorbitalis0.4000.3960.0320.024GausCurv_lh.bankssts − 0.347 − 0.130.0360.400Constructional Praxis RecallGrayVol_lh.rostralanteriorcingulate0.4270.4310.0320.024MeanCurv_lh.temporalpole − 0.354 − 0.3590.0360.028Verbal FluencyGrayVol_rh.rostralanteriorcingulate0.4030.3920.0320.024GausCurv_rh.inferiortemporal − 0.287 − 0.2960.0510.050Boston Naming TestGrayVol_lh.rostralanteriorcingulate0.3720.3490.0360.028lh_temporal (PV WMHs) − 0.268 − 0.2480.2930.059*SBT-K* Short blessed test, Korean version; *MMSE-KC* Mini-mental state examination, Korean version; *CERAD-K* Consortium to Establish a Registry for Alzheimer’s Disease Assessment Packet Korean Version; *rh* Right hemisphere; *WMH* White matter hyperintensity; *Vol* Volume; *lh* Left hemisphere; *CurvInd* Curvature index; *bankssts* banks of the superior temporal sulcus; *FoldInd* Folding index; *MeanCurv* Mean curvature; *GausCurv* Gaussian curvature; *PV* Periventricular. ### Predicting late MCI using machine learning and deep learning analyses The machine learning classifications of the participants are listed in Table 4. Four machine learning algorithms (SVM, RF, MLP, and CNN) were employed to identify the optimal feature pattern for distinguishing late-from early stage MCI. In the health metrics dataset, SVM achieved an AUC of 0.94, ACC of 0.80, recall of 0.86, precision of 0.69, and F1 of 0.75. For the MRI dataset, the CNN achieved an AUC, ACC of 0.86, Recall of 0.94, 0.86, 0.86, 0.79, and F1 of 0.81. In the health metrics and MRI-combined dataset, the MLP achieved an AUC of 0.94, ACC of 0.86, recall of 0.81, precision of 0.83, and an F1 of 0.82. In addition to the classification performance, the top 10 features identified by PCA and ICA for each dataset are provided in Supplementary Table S5. Although the feature sets derived from PCA and ICA did not directly overlap owing to methodological differences, several variables, including gait variability, step length, sleep fragmentation, and regional cortical thickness, consistently emerged as important predictors across different modalities and analytic approaches.Table 4Machine learning classification of patients with MCI.DatasetAlgorithm for classificationAlgorithm for feature reductionNo. of features reducedAUC (95% CI)ACC (95% CI)Recall (95% CI)Precision (95% CI)F1 (95% CI)Gait + Sleep + BCSVMPCA400.94 (0.91–0.97)0.80 (0.75–0.85)0.86 (0.80–0.92)0.69 (0.58–0.80)0.75 (0.70–0.80)Gait + Sleep + BCRFPCA400.87 (0.81–0.93)0.76 (0.69–0.83)0.74 (0.58–0.90)0.65 (0.58–0.72)0.67 (0.61–0.73)Gait + Sleep + BCMLPICA600.94 (0.90–0.98)0.78 (0.73–0.83)0.94 (0.90–0.98)0.63 (0.60–0.66)0.75 (0.72–0.78)Gait + Sleep + BCCNNPCA600.94 (0.91–0.97)0.79 (0.74–0.84)0.86 (0.80–0.92)0.70 (0.64–0.76)0.75 (0.70–0.80)GM + WMHSVMICA200.90 (0.86–0.94)0.81 (0.77–0.85)0.82 (0.72–0.92)0.71 (0.65–0.77)0.75 (0.71–0.79)GM + WMHRFICA600.85 (0.79–0.91)0.78 (0.72–0.84)0.73 (0.69–0.77)0.70 (0.64–0.76)0.70 (0.67–0.73)GM + WMHMLPICA400.90 (0.87–0.93)0.81 (0.77–0.85)0.90 (0.86–0.94)0.69 (0.65–0.73)0.78 (0.74–0.82)GM + WMHCNNICA400.94 (0.90–0.98)0.86 (0.83–0.89)0.87 (0.83–0.91)0.80 (0.73–0.87)0.82 (0.79–0.85)GM + WMH + Gait + Sleep + BCSVMICA200.93 (0.89–0.97)0.86 (0.82–0.90)0.85 (0.79–0.91)0.82 (0.70–0.94)0.81 (0.78–0.84)GM + WMH + Gait + Sleep + BCRFPCA200.92 (0.88–0.96)0.81 (0.76–0.86)0.79 (0.75–0.83)0.74 (0.67–0.81)0.75 (0.71–0.79)GM + WMH + Gait + Sleep + BCMLPICA400.94 (0.92–0.96)0.86 (0.82–0.90)0.82 (0.74–0.90)0.82 (0.76–0.88)0.80 (0.77–0.83)GM + WMH + Gait + Sleep + BCCNNPCA600.93 (0.89–0.97)0.89 (0.85–0.93)0.78 (0.70–0.86)0.91 (0.87–0.95)0.83 (0.80–0.86)*BC* Body composition; *SVM* Support vector machine; *RF* Random forest; *MLP* Multilayer perceptron; *CNN* Convolutional neural network; *ICA* Independent component analysis; *PCA* Principal component analysis, *No.* Number; *AUC* Area under the curve; *ACC* Accuracy; *F1* F1 score. ## Discussion Gait, body composition, sleep parameters, and brain structure were associated with cognitive function. In addition to brain MRI information, which is widely used as a predictor of cognitive function levels, health metrics such as gait, body composition, and sleep were also shown to be effective predictors of cognitive function. These health metrics demonstrated predictive accuracy comparable to that of brain MRI for estimating cognitive function levels. Moreover, combining health metrics with brain MRI data improves the prediction of cognitive function. Previous studies have demonstrated that gait characteristics differ between elderly individuals with normal cognitive function and those with cognitive impairment^15–17^. Furthermore, machine learning models utilizing gait parameters have been successfully employed to differentiate between various degrees of cognitive function impairment^18^. These findings indicate that gait parameters can serve as reliable biomarkers for assessing the extent of cognitive impairment. The cholinergic system plays a crucial role in both gait and cognitive function, supporting the hypothesis that gait function may indirectly reflect cognitive function. Specifically, the findings on the nucleus basalis of Meynert suggest that the majority of cholinergic input to the cerebral cortex modulates hippocampal activity and cognitive frontoparietal networks^19^. Additionally, the cholinergic pedunculopontine nucleus (PPN) is thought to play a key role in both cognitive function^20^ and locomotion^21^. Dysfunction in specific brain regions and associated network impairments affect both cognitive and gait functions, reinforcing the link between these two domains. Our neuropsychological data suggest that attention and executive function are involved in gait control, particularly in the regulation of gait speed and variability. In this context, frontal cognitive networks are crucial, as they integrate sensory information through cortical sensory association areas, such as the parietal and occipital cortices^22^. The frontal-visuospatial network, in particular is implicated in controlling both gait velocity and variability, parameters related to dynamic stability^23^. This network facilitates adaptive motor behavior, especially in response to changes in the environmental context. Additionally, reduced hippocampal volume and decreased metabolism further support the role of memory in gait control. The hippocampal function has been associated with timing or rhythmicity of locomotion^24^. Gait characteristics also associated with hippocampal metabolism vary depending on the degree of memory impairment^24^. Taken together, these findings suggest that gait parameters, influenced by neurocognitive mechanisms, can provide valuable insights into cognitive function and serve as potential biomarkers for cognitive impairment. Previous studies have shown that body composition parameters correlate with cognitive impairment. In midlife, obesity with a body mass index over 30 kg/m^2^ has been linked to an increased risk of dementia and Alzheimer’s disease (AD), even after adjusting for sociodemographic factors^25^. MRI-based studies have further shown that individuals with higher BMI exhibit greater brain atrophy in the frontal, temporal, parietal, and occipital lobes^26^. On the other hand, in individuals over 65 years of age, a higher BMI has paradoxically been associated with a lower risk of dementia^27^. This discrepancy suggests that BMI alone may not accurately reflect body composition, particularly in the elderly population. With aging, body composition changes as muscle mass decreases and fat mass increases, making BMI an unreliable indicator for assessing obesity in older adults^28^. Since BMI does not differentiate between fat and muscle mass, it may fail to capture the true metabolic risks associated with body composition changes in the elderly. More precise body composition assessments have demonstrated that increased fat mass and reduced muscle mass are more directly associated with cognitive impairment in older adults^29^. Additionally, reductions in whole-body soft tissue, particularly lean mass loss, have been linked to cognitive decline^30,31^. Our study also found a positive correlation between lean mass and cognitive function, which is consistent with previous research demonstrating an association between body composition and cognitive function. Additionally, our findings revealed a positive correlation between bone mineral composition and cognitive function, which is consistent with previous studies reporting that decreased bone mineral density (BMD) is linked to an increased risk of AD and cognitive decline^32,33^. One proposed mechanism underlying the relationship between BMD and cognitive function is that osteoporosis may induce calcium influx, leading to degenerative changes in the brain^32^. Additionally, bone mass loss may contribute to cognitive decline by increasing systemic inflammation^34^. In the present study, sleep efficiency was positively associated with visuospatial function. Polysomnography defines sleep efficiency as the percentage of time spent asleep while in bed, and reductions in sleep efficiency often reflect difficulties such as prolonged sleep latency or frequent awakening^35^. Prior studies have reported that sleep efficiency, rather than sleep duration, is more strongly associated with cognitive impairment^36^. A previous study using PSG evidence showed that disturbed sleep patterns precede the clinical onset of cognitive decline, appearing during the prodromal stage known as MCI^37^. Patients with MCI exhibit a reduced amount of REM sleep and increased fragmentation of slow-wave sleep compared to cognitively healthy old adults^37^. Sleep disturbances have been increasingly recognized as both a symptom and a potential contributor to cognitive impairment, particularly in the context of AD. One proposed mechanism is amyloid-beta deposition. Even in the preclinical stage of AD, poor sleep quality has been linked to amyloid-beta accumulation^38^. As amyloid aggregation disrupts circadian rhythms and reduces the sleep quality, the sleep efficiency declines, which in turn leads to increased neuronal activity and further promotes amyloid deposition, creating a self-perpetuating cycle. Animal studies have demonstrated that sleep–wake cycle disruptions, previously considered a consequence of amyloid deposition, may precede its accumulation. Notably, these disruptions were reversed following amyloid clearance, indicating that impaired sleep could serve as an early marker of amyloid pathology^39^. Another pathway involves sleep-induced neuroinflammation. Sleep disturbances elevate the levels of proinflammatory cytokines, leading to microglial activation in the hippocampus and impairing memory consolidation^40^. Neuroinflammation induced by poor sleep may therefore exacerbate cognitive deficits independently or synergistically with amyloid pathology. Finally, structural degeneration of the basal forebrain, a region vulnerable to atrophy in AD and a major source of acetylcholine, plays an essential role in both sleep regulation and cognitive function. This suggests a shared cholinergic mechanism underlying sleep disturbances and cognitive decline^41^. Together, these mechanisms underscore the complex, bidirectional relationship between sleep quality and cognitive function. Cognitive assessments are typically required to detect cognitive decline in patients with MCI. However, frequent cognitive testing is often impractical, which makes it difficult to identify early changes in cognitive function. Previous studies have attempted to address this issue by using brain MRI, a modality considered to be one of the most objective indicators of cognitive function, as input for machine learning models to detect MCI and estimate cognitive status^42^. Nonetheless, brain MRI has limitations in accurately reflecting cognitive function and is not easily accessible due to the high cost and limited availability of imaging equipment. By contrast, gait analysis, body composition measurements, and sleep monitoring can be easily performed in daily life. If cognitive deterioration can be accurately detected using easily accessible physical and behavioral data, it may enable the continuous monitoring of cognitive status and facilitate earlier intervention and treatment. This study has several limitations. First, the relatively small sample size and single-center design may have limited the generalizability of our results. Second, the cross-sectional nature of the study restricts the interpretation of causal relationships or longitudinal cognitive changes. Third, although we used MMSE-based cut-offs to classify early- and late-stage MCI because the MMSE has been widely applied as a global index of cognitive functioning in prior studies^43^ and is suitable for staging purposes, reliance on the MMSE alone may not fully capture the clinical complexity of cognitive impairment. Fourth, any differences in educational attainment and comorbidities between early- and late-stage MCI may have influenced MMSE performance as well as behavioural features such as gait and sleep, raising the possibility of residual confounding. Finally, despite the use of cross validation, the complexity of the models relative to a modest sample size may increase the risk of overfitting. This finding underscores the need for external validation in larger multicentre cohorts to confirm the robustness of our findings. Despite these limitations, this study demonstrated that easily accessible physical and behavioral measures such as gait, body composition, and sleep, can predict cognitive decline with a performance comparable to that of MRI-based models. This suggests a promising and cost-effective approach for early detection and continuous monitoring of cognitive impairment in real-world settings. ## Supplementary Information Supplementary Information.