MRI quantified perivascular space metrics as imaging biomarkers for assessing the severity of cognitive impairment and sleep disturbance in young adults with long-time mobile phone use through machine learning approaches

MRI quantified perivascular space metrics as imaging biomarkers for assessing the severity of cognitive impairment and sleep disturbance in young adults with long-time mobile phone use through machine learning approaches | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article MRI quantified perivascular space metrics as imaging biomarkers for assessing the severity of cognitive impairment and sleep disturbance in young adults with long-time mobile phone use through machine learning approaches Li Li, Jiaojiao Wu, Bin Li, Rui Hua, Feng Shi, Lizhou Chen, Yeke Wu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5384782/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 13 Jan, 2026 Read the published version in Scientific Reports → Version 1 posted 5 You are reading this latest preprint version Abstract Emerging evidence has linked long-time mobile phone use (LTMPU) with cognitive impairment and sleep issues, with MRI-detected enlarged perivascular spaces (EPVSs) serving as markers for these conditions. Our study seeks to develop predictive model using MRI-based PVS measurements and machine learning to assess cognitive impairment, subjective sleep quality, and excessive daytime sleepiness in young adults with LTMPU. Eighty-two participants were included, deep learning algorithms were used to segment EPVS lesions and extract quantitative metrics. Training and testing datasets were randomly assigned to perform radiomics analysis, where EPVS metrics combined with sex and age were used to select the most valuable features for model construction. Finally, a Gaussian process model was constructed based on six features for assessing cognitive impairment, yielding an AUC of 0.818 (95% confidence interval [CI] 0.610-1) in the testing dataset. For sleep quality and sleepiness, two decision tree (DT) models using six features achieved an AUC value of 0.826 (95% CI 0.616-1) and 0.875 (95% CI 0.718-1) in the testing dataset respectively. Our study leveraged MRI-based PVS metrics and machine learning to assess the severity of cognitive impairment and sleep problems in young adults with LTMPU, and sheds light on a potential link between PVS and sleepiness. Biological sciences/Computational biology and bioinformatics Health sciences/Biomarkers Health sciences/Diseases Health sciences/Medical research Health sciences/Nephrology Health sciences/Risk factors Physical sciences/Mathematics and computing MRI imaging biomarker cognitive impairment sleep disturbance LTMPU EPVS Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Mobile phones have become ubiquitous in contemporary society, offering unparalleled convenience while simultaneously sparking health-related inquiries to people’s lives. Approximately 80% of internet users engaged various social media platforms through their mobile phones with young adults aged 18–29, the predominant demographic, spending an average of 181 minutes daily to social media interactions 1 . Long-time mobile phone use (LTMPU), defined as engaging with a mobile device ≥ 4 hours/day is consistently linked to sleep disturbances and mental distress 2 . A recent systematic review published in 2023 has uncovered a significant association between diminished sleep quality and mobile phone usage patterns 3 . Recent evidence also underscores a robust correlation between mobile phone dependency and the prevalence of excessive daytime sleepiness 4 . Furthermore, there is emerging evidence suggesting the link between the overuse of mobile phone and the impairment of cognitive functions 5 . Dementia, which is characterized as impaired memory, language, problem-solving abilities, and an overall decline in cognitive function, stands as a prevalent cause of disability and mortality among the elderly 6 . As China progressively transitions into an aging society, with the elderly population reaching 14% of the total populace and projected to escalate to 22% by 2033 7 , the imperative for early detection and intervention in dementia thus becomes increasingly critical. Given the absence of a curative treatment, the emphasis is on the early identification of cognitive decline, underpinned by the recognition that sleep is integral to cognitive performance 8 . Suboptimal sleep quality is increasingly linked to cognitive deficits 9 , while excessive daytime sleepiness is recognized to exert adverse effects on cognitive functions 10 . Consequently, precise evaluation of cognitive impairment, sleep quality, and the severity of daytime sleepiness symptoms is essential for devising preventive strategies aimed at pinpointing modifiable risk factors, thereby potentially decelerating or mitigating the advancement of dementia. In clinical practice, the Montreal Cognitive Assessment (MoCA) 11 is commonly utilized to assess the severity of cognitive impairment. The Epworth Sleepiness Scale (ESS) is employed to gauge the severity of excessive daytime sleepiness symptoms 12 , while the subjective sleep quality can be assessed by either Pittsburg Sleep Quality Index (PSQI) or the Insomnia severity index (ISI) 13 . However, a significant drawback of these assessment tools is their reliance on subjective reports, which may lack precision, accuracy, and reliability in certain contexts 14 , 15 . This shortfall has intensified the quest for more objective and precise diagnostic techniques to evaluate the severity of cognitive impairment, subjective sleep quality, and daytime sleepiness symptoms. Machine learning, encompassing the development of predictive models and the discovery of meaningful patterns within datasets through computational techniques, offers an objective approach to data analysis 16 . Neuroimaging modalities present significant potential for elucidating the nexus between sleep disorders and the risk of dementia in vivo 17 . Perivascular spaces (PVSs), which are fluid-filled cavities encircling penetrating cerebral arterioles and venules, are hypothesized to facilitate a drainage network crucial for the clearance of metabolic byproducts and cerebrospinal fluid from the brain 18 , 19 . Enlarged perivascular spaces (EPVS), detectable through MRI as indicators of PVS dysfunction 20 , have been traditionally viewed as a physiological variant across a broad age spectrum. However, an excessive burden of PVSs is correlated with prevalent neurological conditions 19 . A substantial body of evidence implicates a negative correlation between PVSs and cognitive functions, as well as sleep processes. A pivotal population-based study has established that the presence of EPVS in the basal ganglia and white matter correlates with a notably elevated risk of developing dementia 21 . Although research on the association between PVSs and excessive daytime sleepiness is nascent, indirect evidence points to perivascular space dysfunction during sleep disruption 20 , which is intricately associated with sleep quality. Previous studies have indicated morphological changes, characterized by the enlargement of basal ganglia-PVSs, in individuals experiencing persistent poor sleep quality following coronavirus disease 22 . The quantification of PVSs is increasingly being addressed through automated methodologies 19 , which may enhance the precision and objectivity of assessment in this research domain. This study aims to develop a predictive model utilizing MRI-quantified PVS metrics and machine learning to evaluate the severity of cognitive impairment, self-reported subjective sleep quality, and the intensity of excessive daytime sleepiness symptoms in young adults with LTMPU. Through this innovative methodology, we hope to elucidate the potential correlation between PVS and cognitive, sleep quality, and excessive daytime sleepiness in individuals addicted to mobile phone use. This research may pave the way for more precise, objective assessments and could potentially inform preventive strategies and interventions in this demographic. 2. Materials and Methods 2.1 Participants This study was a school-based cross-sectional study, which was conducted from October 2021 to May 2022. A total of 165 students and young teachers aged 18 to 50 years in a medical college in Wen jiang District, Chengdu, China were recruited in this study. Among them, 146 (88.5%) responded with valid data. Questionnaires were distributed to the students and young teachers during the class period. This study was approved by the ethics committee of the Hospital of Chengdu University of Traditional Chinese Medicine, and all the research protocols and strategies were performed in accordance with the relevant guidelines and regulations. The inclusion criteria in this study were as follows: (a) with LTMPU. The duration of mobile phone use per day was obtained by the following question: How long do you usually spend on using a mobile phone per day? The response categories for this question were: less than 2 hours, 2 to 4 hours, 4 to 6 hours, and more than 6 hours. long-time mobile phone use (LTMPU) was defined as using a mobile phone ≥ 4 hours per day in consideration of the recent findings 2 . (b) ethnic Han. (c) free of any psychoactive medication at least 2 weeks before and during the study. (d) right-handedness assessed with the Edinburgh Handedness Inventory 23 . Exclusion criteria in this study were as follows: (a) with coronavirus disease 2019 (COVID-19) infections; (b) with any significant neuropsychiatric disease or brain structural abnormality; (c) with MRI contraindications. Furthermore, to evaluate cognitive and sleep status, all participants were asked to complete the MoCA, the ESS, the PSQI, and the ISI. The severity of cognitive impairment was assessed by the MoCA. The total score of MoCA is in the range of 0 to 30. when the score falls below 26, cognitive impairment is present. The lower the MoCA score is, the worse the cognitive function 24 . The severity of excessive daytime sleepiness symptoms was assessed by the ESS. The total score of ESS is in the range of 0 to 24. An ESS score of more than 6, 11, and 16 was defined as sleepiness, excessive sleepiness, and risky sleepiness, respectively 25 . The severity of subjective sleep quality was assessed by the PSQI. The total score of PSQI is in the range of 0 to 21. A score > 5 suggests poor sleep quality 26 . The severity of subjective sleep quality was assessed by the ISI. The total of ISI is in the range of 0 to 28. An ISI score ≤ 7 indicates absence of insomnia; 8–14 indicates sub-threshold insomnia; 15–21 indicates moderate insomnia; 22–28 indicates severe insomnia 26 . Sleep quality is a complex, multifaceted construct that poses challenges for objective quantification due to inter-individual variability and its inherently subjective nature. The PSQI and ISI are two commonly used instruments of subjective self-report sleep quality. The PSQI, a widely recognized questionnaire for gauging subjective sleep quality, has demonstrated robust reliability and validity, particularly in known-group comparisons. However, concerns regarding its factor model, the large recall period, and the scoring system challenge the value of the global PSQI score for distinguishing poor and good sleepers. The ISI, on the other hand, quantifies perceived insomnia severity by focusing on the level of disturbance to the sleep pattern, consequences of insomnia, and the degree of concern and distress related to the sleep problem. The ISI has exhibited significant correlations with various sleep questionnaires icnluding PSQI (albeit with low correlation coefficients with ESS), as well as with psychological, health, and psychopathological assessments. Future studies are needed to clarify the factor structure of ISI. In our study, PSQI and ISI are utilized to evaluated the severity of subjective sleep quality. At baseline, 91 out of 146 participants (62.3%) reported using a mobile phone ≥ 4 hours per day (LTMPU). Each participant with LTMPU completed informed written consent before undergoing magnetic resonance (MR) imaging (within two weeks of completing the scale). Nine participants were excluded because of MRI motion artifacts. Finally, 82 participants with LTMPU were included. 2.2 MR Imaging All patients were examined using a 3.0 T whole-body scanner (Discovery MR750, GE Healthcare, Milwaukee, WI) equipped with a 32-channel phased array head coil. T2-weighted images (T2WI) acquisition parameters were: TR = 5613 ms, TE = 116 ms, slice thickness = 5.0 mm, slice spacing = 1.5 mm, FOV = 26 mm. T2 FLAIR acquisition parameters were: repetition time = 8400 ms, echo time = 150 ms, flip angle = 111°, FOV = 24 cm × 24 cm, matrix size = 256 × 256, inversion time = 2100 ms, slice thickness = 5.0 mm with no gap between slices. 3D T1-weighted imaging (T1WI) was acquired using spoiled gradient echo sequence with repetition time = 2.9 ms, echo time = 3.0 ms, inversion time = 450 ms, flip angle = 8°, slice thickness = 1 mm, matrix = 250 × 250, FOV = 22 cm × 22 cm. 2.3 Data preprocessing and PVS quantification The image preprocessing procedure consisted of several steps, as outlined below. First, N4 bias field corrections were applied to both T1WI and T2WI to remove magnetic field inhomogeneity. Next, grayscale values were standardized by normalizing intensities to the range of [-1, 1] through clipping at 0.1%-99.9%. Utilizing a deep learning model (VB-Net 27 embedded in an image analysis tool named uAI research portal (United Imaging Intelligence) 28 , the skull was removed from T1WI and the whole brain was segmented into 109 regions of interest (ROIs) based on the DK atlas 29 .These regions were then consolidated into 17 brain subregions detailed in Supplementary Table 1 , including bilateral frontal lobes, parietal lobes, occipital lobes, temporal lobes, basal ganglia, cerebellum, thalamus, centrum semiovale, and brainstem. Subsequently, PVS lesions were automatically segmented from the T2WI image using a built-in VB-Net model 30 with AI-generated masks reviewed and modified by two experienced radiologists as necessary. Furthermore, T1WI and T2WI images were co-registered using a registration algorithm 31 , transforming the segmentation mask from the T1WI space to the T2WI space. Finally, a comprehensive analysis was conducted, computing a total of 70 quantitative metrics of PVS lesions. These metrics encompassed the total number and total volume of PVS lesions in the whole brain, as well as the number, volume, average length, and average curvature of PVS lesions for each brain subregion. 2.4 Radiomics analysis Machine learning-based radiomics analysis was used to investigate the ability of PVS characteristics to predict cognitive impairment, sleep quality, insomnia, and sleepiness symptoms in young adults with LTMPU. The radiomics pipeline was performed via the uAI research portal (United Imaging Intelligence) 28 , mainly consisting of feature selection, model construction, and performance evaluation. Data grouping. Among 82 participants, 80% served as the training dataset, used for feature selection and model construction. The rest 20% served as the testing dataset, used to evaluate the robustness and generalizability of the model. Feature selection. The 70 PVS quantitative features in conjunction with 2 available clinical features (i.e., sex and age) served as the input to identify the most valuable biomarkers for clinical outcomes. Notably, feature standardization was first conducted to eliminate the effect of magnitudes between different features. Then, the minimum redundancy maximum relevance (mRMR) method was employed to select the most relevant feature combinations. Model construction. Based on the selected features, multiple machine learning algorithms (e.g., support vector machine [SVM], random forest [RF], logistic regression [LR], and K nearest neighbors [KNN], decision tree [DT], Gaussian process [GP]) were used to construct the classification models. For each classification task, we retained the model with the highest discriminative performance, where the GP model was used for the MoCA and ISI classification, and the DT model for the PSQI and ESS classification. The hyperparameters of each model are detailed in Supplementary Table 2 . Model evaluation. The performance of models was evaluated in the testing dataset, which could reflect the robustness and generalizability of models. The receiver operating characteristic (ROC) curve was first plotted, where the area under the curve (AUC) could be calculated quantitatively. Five metrics were calculated to evaluate the consistency between the actual label and predictive label, including accuracy, sensitivity, specificity, precision, and F1-score. These metrics were defined as follows ( Equations 1 – 5 ): $$\:Accuracy=\:\frac{TP+TN}{TP+PF+TN+FN}\:,\:$$ 1 $$\:Sensitivity=\:Recall=\:\frac{TP}{TP+FN\:}\:,$$ 2 $$\:Specificity=\frac{TN}{TN+FP\:}\:,$$ 3 $$\:Precision=\frac{TP}{TP+FP\:}\:,$$ 4 $$\:F1score=\frac{2\ast\:Precision\ast\:Recall}{Precision+Recall\:}\:,$$ 5 where TP represented true positive, TN represented true negative, FP represented false positive, and FN represented false negative. Calibration curves were also used to compare the predictive output and the actual outcome. Finally, the decision curves were utilized to show the clinical net benefit for predicting outcomes. 2.5 Statistical analysis The Shapiro-Wilk tests were used to check the normal distribution of continuous variables. For continuous variables that were approximately normally distributed, they were represented as mean ± standard deviation. For continuous variables with asymmetrical distributions, they were represented as median (25th, 75th percentiles). Categorical variables were represented as counts (percentages), and compared using chi-square tests. The correlation analysis utilized Pearson’s method when both variables satisfied normal distribution assumptions; otherwise, Spearman’s method was applied. To evaluate the classification performance of machine learning models, six quantitative metrics (i.e., AUC, accuracy, sensitivity, specificity, precision, and F1-score) were calculated. All statistical analyses were implemented using SPSS (version 26.0, https://www.ibm.com/spss ) and R (version 4.2.2, https://www.R-project.org ). All figures were plotted using GraphPad Prism 9 ( https://www.graphpad.com/ ), Origin 2021 ( https://www.originlab.com/ ), and Adobe Illustrator CC 2019 ( https://www.adobe.com/products/illustrator.html ). 3. Results 3.1 Participants characteristics We recruited 82 participants who underwent MRI examinations from the Affiliated Hospital of Chengdu University of Traditional Chinese Medicine between October 2021 and May 2022. The demographics and clinical scales of each participant were collected and presented in Table 1 . The median age of all participants is 38.0 years (Figure 1a) , with 29.3% (24/82) being male. The distribution of cognitive impairment, poor sleep quality, insomnia, and sleepiness among all participants is visualized in Figure 1b , with occurrences of 55, 54, 36, and 40, respectively (Supplementary Figure 1) . It demonstrates that one participant suffers from multiple disorders at the same time, revealing an intrinsic correlation among them. The demographics of participants in each disorder are summarized in Table 1 . Notably, the median age of the cognitive normalization group (MoCA ≥ 26) is 33.0 years, whereas the median age of the cognitive impairment group (MoCA < 26) is 40.0 years, which is a significant difference between the two ages. There are no significant differences in sex distribution between the cognitive normalization and cognitive impairment groups, between the good sleep (PSQI ≤ 5) and poor sleep (PSQI > 5) groups, between the non-insomnia (ISI ≤ 7) and insomnia (ISI > 7) groups, and between the non-sleepiness (ESS ≤ 6) and sleepiness (ESS > 6) groups. Table 1. Demographics of participants. The continuous variables between the non-anxiety and the anxiety, and between the non-depression and depression were compared using Mann-Whitey U tests. The sex distribution was compared using the chi-square test. A two-tailed p-value < 0.05 was considered a significant difference. 3.2 Gaussian process model in predicting cognitive impairment To explore the ability of PVS features to predict cognitive impairment severity, subjective sleep quality, and excessive daytime sleepiness symptoms severity in young adults with LTMPU, machine learning-based radiomics analyses are conducted. A total of 70 PVS features combined with easily accessible participant demographics (i.e., sex, age) are used as inputs to select the most valuable features to construct the machine learning model. The cognitive function can be classified into two categories based on MoCA scores, with MoCA ≥ 26 being the cognitive normalization group and MoCA < 26 being the cognitive impairment group. To identify participants with cognitive impairment, the mRMR method is used to select the six most valuable features (Figure 2a) , whose correlation matrix is shown in Figure 2b . It is clear that three pairs of features are significantly correlated. The distribution of the six features used to construct the machine learning model in the training and testing datasets is shown in Supplementary Figure 1 . There are significant differences between the cognitively normal and cognitively impaired groups in terms of age (Supplementary Figure 1a) and the average length of PVS lesions in the left centrum semiovale (Supplementary Figure 1f) . Subsequently, a classification model is constructed using the Gaussian process (GP) algorithm, with receiver operating characteristic (ROC) curves plotted in Figure 2c . Specifically, the area under the ROC curve (AUC) values of the GP model are 0.949 with a 95% confidence interval (CI) of 0.900-0.998 and 0.818 (95% CI 0.610-1.000) in the training and testing datasets, respectively. The calibration curves show that the positive incidence predicted by the GP model deviates somewhat from the actual incidence of cognitive impairment in the training and testing datasets, indicating that the accuracy of the model prediction needs to be further improved (Figure 2d) . Nonetheless, the model is still able to achieve a net clinical benefit within a threshold range of 0.3 to 0.9 (Figure 2e) . Additionally, the classification performance of the GP model is evaluated by the other five quantitative metrics, as detailed in Table 2 , calculated from the confusion matrix (Figures 2f, 2h) . It is easy to find that the specificity and precision are higher than 0.80 in both the training and testing datasets (Figures 2g, 2i) , while the sensitivity is only 0.636 in the testing dataset, which implies that there is a certain degree of false-negative rate for the GP model. Table 2. Performance of four machine learning models in predicting cognitive and sleep disorders. 3.3 Decision tree model in predicting subjective sleep quality(PSQI) Radiomics analysis is also used to categorize participants with poor subjective sleep quality (PSQI > 5) and good subjective sleep quality (PSQI ≤ 5). Six features are chosen through the mRMR method ( Figure 3a ), and the correlation matrix, displayed in Figure 3b , reveals significant correlations between five pairs of features. The distribution of the six features used to construct the machine learning model in the training and testing datasets is shown in Supplementary Figure 2 . A decision tree (DT) algorithm is employed to build a classification model, with ROC curves depicted in Figure 3c . The DT model exhibits AUC values of 0.865 (95% CI 0.770 – 0.959) and 0.826 (95% CI 0.616 – 1.000) in the training and testing datasets, respectively. The calibration curve demonstrates strong agreement between the predicted positive incidence rates by the DT model and the actual incidence rates of poor subjective sleep quality in the training dataset, with slight deviation observed in the testing dataset ( Figure 3d ). Overall, the model yields high prediction accuracy rates of 0.846 and 0.824 in the two datasets, as shown in Table 2 . Furthermore, decision curves illustrated in Figure 3e indicate that the model delivers substantial clinical net benefits across a range of thresholds (0.2-1.0) in both training and testing datasets, showcasing its potential for enhancing patient care and decision-making support effectively. Detailed quantitative metrics from confusion matrices visualized in Figures 3f, 3g for training data and Figures 3h, 3i for testing datasets are presented in Table 2 . Overall, except for specificity, all metrics surpass a threshold of above 0.8 in both datasets, indicating a relatively high prediction performance by the model despite some false positives being present. 3.4 Gaussian process model in predicting subjective sleep quality(ISI) Similar feature selection and modeling procedures are conducted to distinguish between the non-insomnia group (ISI ≤ 7) and the insomnia group (ISI > 7). Using the mRMR method, a total of six features are selected (Figure 4a) , with the corresponding correlation matrix presented in Figure 4b , indicating no significant correlation among the selected features, underscoring their unique and independent information contribution. The distribution of the six features used to construct the machine learning model in the training and testing datasets is shown in Supplementary Figure 3 . Subsequently, a classification model is built using the GP algorithm. The AUC values for the GP model are calculated as 0.947 (95% CI 0.888-1.000) in the training dataset and 0.757 (95% CI 0.492-1.000) in the testing dataset, as shown in Figure 4c . Although calibration curves show narrower prediction intervals and acceptable accuracy (Figure 4d) , decision curves reveal a clinical net benefit in both datasets, particularly within a threshold range of 0.3-0.6, emphasizing the model's potential in informing clinical decisions (Figure 4e) . The confusion matrices depicted in Figure 4f for the training dataset and Figure 4h for the testing dataset are utilized to compute quantitative metrics such as accuracy, F1-score, sensitivity, specificity, and precision. These metrics for both datasets are graphically presented in Figure 4g and Figure 4i . While the model displays robust precision and specificity, indicating its proficiency in identifying true negatives accurately, its lower sensitivity suggests a higher false negative rate that may lead to missed true positive cases. 3.5 Decision tree model in predicting excessive daytime sleepiness symptoms Feature selection and modeling procedures are applied to differentiate between the non-sleepiness group (ESS ≤ 6) and the sleepiness group (ESS > 6). Six features are selected using the mRMR method, as illustrated in Figure 5a , with the correlation matrix presented in Figure 5b showing no significant correlations among these selected features. This absence of significant correlation highlights the unique and independent information contributed by each feature. The distribution of the six features used to construct the machine learning model in the training and testing datasets is shown in Supplementary Figure 4 . There is a significant difference between the non-sleepiness and sleepiness groups in terms of the average length of PVS lesions in the left centrum semiovale (Supplementary Figure 4h) . Following this, a classification model is constructed using the DT algorithm. The AUC values for the DT model are determined as 0.923 (95% CI 0.867-0.978) in the training dataset and 0.875 (95% CI 0.718-1.000) in the testing dataset, as depicted in Figure 5c . The calibration curves demonstrate a strong alignment between the model’s predicted likelihood of sleepiness and the actual prevalence of sleepiness in both the training and testing datasets, as shown in Figure 5d . This alignment underscores the model's ability to accurately estimate an individual’s probability of belonging to the sleepiness group across different datasets, indicating its reliability in evaluating sleep disorders. Furthermore, decision curves showcase that the model provides clinical utility and benefit across a wide range of thresholds (0.1-0.8), suggesting its potential positive impact on clinical decision-making processes (Figure 5e) . Quantitative metrics such as accuracy, F1-score, sensitivity, specificity, and precision are computed based on confusion matrices depicted in Figure 5f for the training dataset and Figure 5h for the testing dataset. These metrics for both datasets are visually represented in Figure 5g and Figure 5i . Overall performance evaluation reveals that except for sensitivity, all metrics exceed a threshold of 0.8 in both datasets, indicating a strong performance across various assessment criteria. 4. Discussion To our knowledge, this paper is the first study that presents a novel approach to classify cognitive impairment severity, subjective sleep quality, and excessive daytime sleepiness symptoms severity in young adults with LTMPU by integrating MRI-based quantification of PVS and machine learning algorithms. Our model has exhibited remarkable accuracy in these classifications, presenting a promising path for non-invasive and objective assessment methodologies. The integration of MRI data with advanced computational models represents a significant advancement in the field, as no prior studies have been known to harness machine learning to such an end with respect to sleep and cognitive function. Furthermore, our research provides preliminary evidence suggesting a correlation between PVS and subjective symptoms of excessive daytime sleepiness, an area that has received scant attention in previous studies, as far as we know. PVSs are integral to the functionality of the glymphatic system, a network responsible for brain waste clearance 32 . Under normal conditions, PVSs are not discernible on structural MRI; however, the visibility of EPVS may indicate a dysfunction in glymphatic clearance 33 . It is plausible to hypothesize that EPVS may be associated with poor sleep quality and excessive daytime sleepiness, given the influence of sleep on glymphatic system efficacy 32 , although current evidence is limited. An early study has demonstrated that poor sleep quality was independently associated with the EPVS in basal ganglia and white matter 34 . The glymphatic system, with PVSs as a key component, plays a crucial role in the clearance of brain amyloid β (Aβ) and tauopathy, which are linked to neurodegenerative conditions 35 . A 2019 study has showed that a substantial perivascular space burden is associated with common neurological diseases, such as Alzheimer’s disease 36 . The quantification of PVSs has been enhanced through computational methods, which have shown heightened sensitivity in associating PVSs with white matter hyperintensities and retinal vessel diameters 37 . In our current research, we referenced a previous study for the precise measurement of PVS quantification 30 , thereby aiming to advance the understanding of PVSs in the context of subjective sleep quality, cognitive function, and their potential implications in neurodegenerative processes. In recent years, many studies have employed machine learning techniques to develop predictive models for classifying the severity of cognitive impairment 38 . This study represents the first attempt to utilize MRI quantified EPVS volumes and machine learning to accurately classify subjective sleep quality and the severity of excessive daytime sleepiness symptoms in young adults with LTMPU, which may hold significant potential for clinical applications. Our study provides preliminary evidence suggesting a relationship between EPVS, a biomarker indicative of glymphatic dysfunction, and the severity of cognitive impairment, sleep quality, and the severity of excessive daytime sleepiness symptoms in young adults with LTMPU. This is in line with the burgeoning body of research that points to a bidirectional relationship between sleep disturbances and the risk of dementia 17 . The absence of curative treatments for dementia underscores the critical need for preventive interventions 17 . By integrating MRI-quantified EPVS volumes with machine learning algorithms, our model may offer insights into the early stages of Alzheimer’s disease, potentially identifying syndromal conversion in cognitively unimpaired subjects—a domain where data are exceedingly scarce. This approach harnesses the power of neuroimaging to detect preclinical neurodegenerative changes, facilitating both the early diagnosis of Alzheimer’s and the monitoring of sleep health. Excessive daytime sleepiness is a public health issue, which is often undervalued, infrequently diagnosed, and inadequately addressed 12 . The observed relationship between EPVS and excessive daytime sleepiness symptoms in our study suggests that EPVS could serve as a promising biomarker for this condition. Further exploration of this association could deepen our understanding of the neurobiology underlying excessive daytime sleepiness, ultimately aiding in the development of improved diagnostic and therapeutic strategies for affected patients. This study, while pioneering in its approach, has several limitations. Firstly, the modest sample size employed may restrict the generalizability of our findings, warranting a need for validation on a larger scale to enhance the model's robustness and applicability. Secondly, our observational design does not permit the determination of causality between the persistent symptoms of excessive daytime sleepiness and the presence of EPVS. Additionally, the study’s single-center nature necessitates broader validation across diverse populations and settings to ensure the findings are representative and reliable. Furthermore, the current research did not include an analysis of EPVS location, which is imposed by the sample size. Future studies should aim to incorporate a more detailed examination of EPVS distribution to provide a comprehensive understanding of their potential impact on cognitive and sleep-related symptoms. 5. Conclusions Our study introduces an innovative analytical framework by integrating MRI-quantified PVS metrics with machine learning algorithms, offering a new methodological paradigm for classifying the severity of cognitive impairment, subjective sleep quality and excessive daytime sleepiness symptoms in young adults with LTMPU. The insights gained from this preliminary investigation set the stage for more extensive inquiries into the complex interplay between EPVS, cognitive function, and sleep quality in the context of LTMPU. Further research with expanded cohorts and multi-centric approaches are imperative for substantiating the reliability and generalizability of our model. Such endeavors will be pivotal in validating the predictive capabilities of our model across various populations and healthcare settings, thereby enhancing its potential impact on clinical diagnostics and therapeutic interventions. Declarations Acknowledgements Funding This work was supported by the National Natural Science Foundation of China (grant number: 82074303, 82174345, 81973684), Sichuan Key Research and Development Project (2023YFS0226), and Natural Science Foundation of Sichuan Province (2023NSFSC1760). Competing interests J.W., R.H., and F.S. are employees of United Imaging Intelligence. The company has no role in designing and performing the surveillance and analyzing and interpreting the data. All other authors report no conflicts of interest relevant to this article. Disclosure statement The authors report no biomedical financial interests or potential conflicts of interest. Data availability statement All data generated or analysed during this study are included in this published article and its supplementary information files. References Joshi, S. C., Woltering, S. & Woodward, J. Cell Phone Social Media Use and Psychological Well-Being in Young Adults: Implications for Internet-Related Disorders. Int J Environ Res Public Health 20 , (2023). Liu, S. et al. The associations of long-time mobile phone use with sleep disturbances and mental distress in technical college students: a prospective cohort study. Sleep 42 , (2019). Brautsch, L. A. et al. Digital media use and sleep in late adolescence and young adulthood: A systematic review. Sleep Med Rev 68 , 101742 (2023). Nowak, M. et al. Correlations between Problematic Mobile Phone Use and Depressiveness and Daytime Sleepiness, as Well as Perceived Social Support in Adolescents. Int J Environ Res Public Health 19 , (2022). Wacks, Y. & Weinstein, A. M. Excessive Smartphone Use Is Associated With Health Problems in Adolescents and Young Adults. Front Psychiatry 12 , 669042 (2021). Sadeghmousavi, S., Eskian, M., Rahmani, F. & Rezaei, N. The effect of insomnia on development of Alzheimer's disease. J Neuroinflammation 17 , 289 (2020). Zhao, M., Li, J., Xiang, L., Zhang, Z. H. & Peng, S. L. A diagnosis model of dementia via machine learning. Front Aging Neurosci 14 , 984894 (2022). Mason, G. M., Lokhandwala, S., Riggins, T. & Spencer, R. M. C. Sleep and human cognitive development. Sleep Med Rev 57 , 101472 (2021). Li, M., Wang, N. & Dupre, M. E. Association between the self-reported duration and quality of sleep and cognitive function among middle-aged and older adults in China. J Affect Disord 304 , 20-27 (2022). Slonkova, J. et al. Hypocretin-1/orexin-A, sleep and excessive daytime sleepiness in patients with nonconvulsive status epilepticus: A cross-sectional cohort study. Sleep Med 119 , 192-200 (2024). Cai, Y. et al. Comparing machine learning-derived MRI-based and blood-based neurodegeneration biomarkers in predicting syndromal conversion in early AD. Alzheimers Dement 19 , 4987-4998 (2023). Pérez-Carbonell, L., Mignot, E., Leschziner, G. & Dauvilliers, Y. Understanding and approaching excessive daytime sleepiness. Lancet 400 , 1033-1046 (2022). McCarter, S. J. et al. Physiological markers of sleep quality: A scoping review. Sleep Med Rev 64 , 101657, doi:10.1016/j.smrv.2022.101657 (2022). Liu, X. et al. Gait can reveal sleep quality with machine learning models. PLoS One 14 , e0223012 (2019). Jeon, J. et al. Accuracy of Machine Learning Using the Montreal Cognitive Assessment for the Diagnosis of Cognitive Impairment in Parkinson's Disease. J Mov Disord 15 , 132-139 (2022). Greener, J. G., Kandathil, S. M., Moffat, L. & Jones, D. T. A guide to machine learning for biologists. Nat Rev Mol Cell Biol 23 , 40-55 (2022). Ward, S. A. & Pase, M. P. Advances in pathophysiology and neuroimaging: Implications for sleep and dementia. Respirology 25 , 580-592 (2020). Ding, J. et al. Large Perivascular Spaces Visible on Magnetic Resonance Imaging, Cerebral Small Vessel Disease Progression, and Risk of Dementia: The Age, Gene/Environment Susceptibility-Reykjavik Study. JAMA Neurol 74 , 1105-1112 (2017). Duering, M. et al. Neuroimaging standards for research into small vessel disease-advances since 2013. Lancet Neurol 22 , 602-618 (2023). Wardlaw, J. M. et al. Perivascular spaces in the brain: anatomy, physiology and pathology. Nat Rev Neurol 16 , 137-153 (2020). Zhu, Y. C. et al. High degree of dilated Virchow-Robin spaces on MRI is associated with increased risk of dementia. J Alzheimers Dis 22 , 663-672 (2010). Del Brutto, O. H. et al. Long coronavirus disease-related persistent poor sleep quality and progression of enlarged perivascular spaces. A longitudinal study. Sleep 45 , (2022). Oldfield, R. C. The assessment and analysis of handedness: the Edinburgh inventory. Neuropsychologia 9 , 97-113 (1971). Chen, X. Y., Chen, Z. Y., Dong, Z., Liu, M. Q. & Yu, S. Y. Regional volume changes of the brain in migraine chronification. Neural Regen Res 15 , 1701-1708 (2020). Hou, A. et al. Widespread aberrant functional connectivity throughout the whole brain in obstructive sleep apnea. Front Neurosci 16 , 920765 (2022). Morin, C. M., Belleville, G., Bélanger, L. & Ivers, H. The Insomnia Severity Index: psychometric indicators to detect insomnia cases and evaluate treatment response. Sleep 34 , 601-608 (2011). Shi, F. et al. Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy. Nat Commun 13 , 6566 (2022). Wu, J. et al. uRP: An integrated research platform for one-stop analysis of medical images. Front Radiol 3 , 1153784 (2023). Desikan, R. S. et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage 31 , 968-980 (2006). Zhang, Z. et al. Quantitative Analysis of Multimodal MRI Markers and Clinical Risk Factors for Cerebral Small Vessel Disease Based on Deep Learning. Int J Gen Med 17 , 739-750 (2024). Avants, B. B. et al. The Insight ToolKit image registration framework. Front Neuroinform 8 , 44 (2014). Dredla, B. K., Del Brutto, O. H. & Castillo, P. R. Sleep and Perivascular Spaces. Curr Neurol Neurosci Rep 23 , 607-615 (2023). Ramirez, J. et al. MRI-visible perivascular space volumes, sleep duration and daytime dysfunction in adults with cerebrovascular disease. Sleep Med 83 , 83-88 (2021). Yang, S., Yin, J., Qin, W., Yang, L. & Hu, W. Poor Sleep Quality Associated With Enlarged Perivascular Spaces in Patients With Lacunar Stroke. Front Neurol 12 , 809217 (2021). Wang, X. X. et al. MRI-visible enlarged perivascular spaces: imaging marker to predict cognitive impairment in older chronic insomnia patients. Eur Radiol 32 , 5446-5457 (2022). Francis, F., Ballerini, L. & Wardlaw, J. M. Perivascular spaces and their associations with risk factors, clinical disorders and neuroimaging features: A systematic review and meta-analysis. Int J Stroke 14 , 359-371 (2019). Ballerini, L. et al. Computational quantification of brain perivascular space morphologies: Associations with vascular risk factors and white matter hyperintensities. A study in the Lothian Birth Cohort 1936. Neuroimage Clin 25 , 102120 (2020). Levy, B. et al. Machine Learning Enhances the Efficiency of Cognitive Screenings for Primary Care. J Geriatr Psychiatry Neurol 32 , 137-144 (2019). Tables Tables 1 to 2 are available in the Supplementary Files section Additional Declarations No competing interests reported. Supplementary Files SupplementaryInformation.docx Tables.docx Cite Share Download PDF Status: Published Journal Publication published 13 Jan, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 19 Mar, 2025 Editor assigned by journal 11 Feb, 2025 Editor invited by journal 19 Nov, 2024 Submission checks completed at journal 18 Nov, 2024 First submitted to journal 03 Nov, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5384782","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":384273102,"identity":"66f01f97-05d0-422b-a81e-eefbf6183edb","order_by":0,"name":"Li Li","email":"","orcid":"","institution":"Department of Radiology, Hospital of Chengdu University of Traditional Chinese Medicine, Chengdu 610072","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Li","suffix":""},{"id":384273103,"identity":"1777ba06-da18-4517-8b04-5f69cb16665c","order_by":1,"name":"Jiaojiao Wu","email":"","orcid":"","institution":"Department of Research and Development, Shanghai United Imaging Intelligence Co., Ltd., Shanghai 200232","correspondingAuthor":false,"prefix":"","firstName":"Jiaojiao","middleName":"","lastName":"Wu","suffix":""},{"id":384273104,"identity":"9850dfd2-85ef-419c-b6e2-14f90be05c9b","order_by":2,"name":"Bin Li","email":"","orcid":"","institution":"Department of Geriatrics, Hospital of Chengdu University of Traditional Chinese Medicine, Chengdu 610072","correspondingAuthor":false,"prefix":"","firstName":"Bin","middleName":"","lastName":"Li","suffix":""},{"id":384273105,"identity":"240652c4-7f3d-48b0-94fa-f0a1f9888f97","order_by":3,"name":"Rui Hua","email":"","orcid":"","institution":"Department of Research and Development, Shanghai United Imaging Intelligence Co., Ltd., Shanghai 200232","correspondingAuthor":false,"prefix":"","firstName":"Rui","middleName":"","lastName":"Hua","suffix":""},{"id":384273106,"identity":"f6d89169-5f11-4c6f-a352-93d42fa02580","order_by":4,"name":"Feng Shi","email":"","orcid":"","institution":"Department of Research and Development, Shanghai United Imaging Intelligence Co., Ltd., Shanghai 200232","correspondingAuthor":false,"prefix":"","firstName":"Feng","middleName":"","lastName":"Shi","suffix":""},{"id":384273107,"identity":"58ac4536-8232-4c3b-af59-d25e1da03091","order_by":5,"name":"Lizhou Chen","email":"","orcid":"","institution":"Department of Radiology, West China Hospital of Sichuan University, Chengdu, 610041","correspondingAuthor":false,"prefix":"","firstName":"Lizhou","middleName":"","lastName":"Chen","suffix":""},{"id":384273109,"identity":"53164e02-9e46-4a53-a334-811981768f75","order_by":6,"name":"Yeke Wu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0klEQVRIiWNgGAWjYJACxgYGGwaGA3AOcVrSgFqYSdNymAQtBjdyDB/OqDgvz3cj/+CnGww2shsOMD97QECLseGGM7cNZ95IZpbOYUgz3nCAzdwAnxazGzlmkg/bbicY3EhmY85hOJy44QAPmwQBLeY/H/47B9PynygtZowbGw7AtBwgrMX+zLNiyRnHkg1nnnlsLJ1jkGw88zCbGV4tku3JGz/21NjJ8x1PfPg5p8JOtu948zO8WhgYOJCDB8Rmxq8eCNgfEFQyCkbBKBgFIxwAAJCkTnX7DaB6AAAAAElFTkSuQmCC","orcid":"","institution":"Department of Stomatology, Hospital of Chengdu University of Traditional Chinese Medicine, Chengdu, 610072","correspondingAuthor":true,"prefix":"","firstName":"Yeke","middleName":"","lastName":"Wu","suffix":""}],"badges":[],"createdAt":"2024-11-04 04:38:35","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5384782/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5384782/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-35845-3","type":"published","date":"2026-01-13T16:29:25+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":72299853,"identity":"72b64352-3440-417f-b789-1a3240470207","added_by":"auto","created_at":"2024-12-25 01:13:19","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":31648,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCharacteristics of the participants.\u003c/strong\u003e \u003cstrong\u003e(a)\u003c/strong\u003e Age distribution of all participants. \u003cstrong\u003e(b)\u003c/strong\u003e Distribution of four disorders in all participants.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5384782/v1/b2ef06ecfaa1c4572a0f7320.png"},{"id":72299857,"identity":"97073ace-8824-448c-b83f-5ed11b099b42","added_by":"auto","created_at":"2024-12-25 01:13:19","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":154906,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel construction and evaluation in identifying cognitive impairment. (a) \u003c/strong\u003eSix features selected by the mRMR method. \u003cstrong\u003e(b)\u003c/strong\u003e Correlation heatmap of selected features. Pearson or Spearman correlation analyses were performed and * indicates \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05. Numbers labeled in the plots represent correlation coefficients. \u003cstrong\u003e(c) \u003c/strong\u003eROC curves evaluating the trade-off between sensitivity and specificity of the GP model, with a higher AUC indicating a better discrimination ability of the model across different threshold settings. \u003cstrong\u003e(d) \u003c/strong\u003eCalibration curves evaluating the consistency of predicted probability and the actual cognitive impairment rate. \u003cstrong\u003e(e)\u003c/strong\u003e Decision curves showing the clinical net benefit. \u003cstrong\u003e(f) \u003c/strong\u003eConfusion matrix of the training dataset. The “0”\u003cstrong\u003e \u003c/strong\u003eand “1” represent cognitive normalization and impairment, respectively.\u003cstrong\u003e (g)\u003c/strong\u003eMetrics of the training dataset. \u003cstrong\u003e(h) \u003c/strong\u003eConfusion matrix and \u003cstrong\u003e(i) \u003c/strong\u003emetrics of the testing dataset.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5384782/v1/f14f10c9fe68e04abed26334.png"},{"id":72299856,"identity":"fbd3cf9d-50b3-4587-a5f7-89922230bc61","added_by":"auto","created_at":"2024-12-25 01:13:19","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":157610,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel construction and evaluation in classifying poor \u003c/strong\u003esubjective \u003cstrong\u003esleep quality and good \u003c/strong\u003esubjective \u003cstrong\u003esleep quality\u003c/strong\u003e(PSQI)\u003cstrong\u003e. (a) \u003c/strong\u003eSix features selected by the mRMR method. \u003cstrong\u003e(b)\u003c/strong\u003eCorrelation heatmap of selected features. Pearson or Spearman correlation analyses were performed and * indicates \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05. Numbers labeled in the plots represent correlation coefficients. \u003cstrong\u003e(c) \u003c/strong\u003eROC curves evaluating the trade-off between sensitivity and specificity of the DT model. \u003cstrong\u003e(d) \u003c/strong\u003eCalibration curves evaluating the consistency of predicted probability and the actual poor subjective sleep quality rate. \u003cstrong\u003e(e)\u003c/strong\u003e Decision curves showing the clinical net benefit. \u003cstrong\u003e(f) \u003c/strong\u003eConfusion matrix of the training dataset. The “0”\u003cstrong\u003e \u003c/strong\u003erepresents good sleep and “1” represents poor sleep. \u003cstrong\u003e(g)\u003c/strong\u003e Six metrics of the training dataset. \u003cstrong\u003e(h) \u003c/strong\u003eConfusion matrix of the testing dataset.\u003cstrong\u003e (i)\u003c/strong\u003e Six metrics of the testing dataset.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5384782/v1/917c8fb84c7aceda3a8dcf0d.png"},{"id":72299863,"identity":"d4ddc8f3-37a3-447b-b1c9-7653bdc7e7eb","added_by":"auto","created_at":"2024-12-25 01:13:19","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":149506,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel construction and evaluation in identifying \u003c/strong\u003esubjective sleep quality(ISI).\u003cstrong\u003e (a) \u003c/strong\u003eSix features selected by the mRMR method. \u003cstrong\u003e(b)\u003c/strong\u003e Correlation heatmap of selected features. Pearson or Spearman correlation analyses were performed and * indicates \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05. Numbers labeled in the plots represent correlation coefficients. \u003cstrong\u003e(c) \u003c/strong\u003eROC curves evaluating the trade-off between sensitivity and specificity of the GP model. \u003cstrong\u003e(d) \u003c/strong\u003eCalibration curves evaluating the consistency of predicted probability and the actual insomnia rate. \u003cstrong\u003e(e)\u003c/strong\u003e Decision curves showing the clinical net benefit. \u003cstrong\u003e(f) \u003c/strong\u003eConfusion matrix of the training dataset. The “0”\u003cstrong\u003e \u003c/strong\u003erepresents non-insomnia and “1” represents insomnia. \u003cstrong\u003e(g)\u003c/strong\u003e Six metrics of the training dataset. \u003cstrong\u003e(h) \u003c/strong\u003eConfusion matrix of the testing dataset.\u003cstrong\u003e (i)\u003c/strong\u003eSix metrics of the testing dataset.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5384782/v1/b7288d44d0cff16f7a660452.png"},{"id":72299859,"identity":"f22b412d-97a7-4faa-a1f8-b9e82c402bff","added_by":"auto","created_at":"2024-12-25 01:13:19","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":155375,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel construction and evaluation in identifying sleepiness. (a) \u003c/strong\u003eSix features selected by the mRMR method. \u003cstrong\u003e(b)\u003c/strong\u003e Correlation heatmap of selected features. Pearson or Spearman correlation analyses were performed and * indicates \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05. Numbers labeled in the plots represent correlation coefficients. \u003cstrong\u003e(c) \u003c/strong\u003eROC curves evaluating the trade-off between sensitivity and specificity of the DT model. \u003cstrong\u003e(d) \u003c/strong\u003eCalibration curves evaluating the consistency of predicted probability and the actual sleepiness rate. \u003cstrong\u003e(e)\u003c/strong\u003e Decision curves showing the clinical net benefit. \u003cstrong\u003e(f) \u003c/strong\u003eConfusion matrix of the training dataset. The “0”\u003cstrong\u003e \u003c/strong\u003erepresents non-sleepiness and “1” represents sleepiness. \u003cstrong\u003e(g)\u003c/strong\u003e Six metrics of the training dataset. \u003cstrong\u003e(h) \u003c/strong\u003eConfusion matrix of the testing dataset.\u003cstrong\u003e (i)\u003c/strong\u003eSix metrics of the testing dataset.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5384782/v1/730dd7510833457053130580.png"},{"id":100617370,"identity":"57a9126f-bcea-4b22-be50-92406c224bba","added_by":"auto","created_at":"2026-01-19 17:51:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1798576,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5384782/v1/f1bc2ca0-8012-47bb-8bc7-7b95e2a4ce65.pdf"},{"id":72300210,"identity":"7cef0a96-5f4f-4d39-ba3d-7cd4032a3f2a","added_by":"auto","created_at":"2024-12-25 01:21:19","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":2353232,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-5384782/v1/f37dc68a132a79699d095a84.docx"},{"id":72299854,"identity":"3cb583a3-3c57-438d-a928-0669c9cdfbe4","added_by":"auto","created_at":"2024-12-25 01:13:19","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":24817,"visible":true,"origin":"","legend":"","description":"","filename":"Tables.docx","url":"https://assets-eu.researchsquare.com/files/rs-5384782/v1/2b2feca9892c1b0e1eee0b3c.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"MRI quantified perivascular space metrics as imaging biomarkers for assessing the severity of cognitive impairment and sleep disturbance in young adults with long-time mobile phone use through machine learning approaches","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eMobile phones have become ubiquitous in contemporary society, offering unparalleled convenience while simultaneously sparking health-related inquiries to people\u0026rsquo;s lives. Approximately 80% of internet users engaged various social media platforms through their mobile phones with young adults aged 18\u0026ndash;29, the predominant demographic, spending an average of 181 minutes daily to social media interactions\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Long-time mobile phone use (LTMPU), defined as engaging with a mobile device\u0026thinsp;\u0026ge;\u0026thinsp;4 hours/day is consistently linked to sleep disturbances and mental distress\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. A recent systematic review published in 2023 has uncovered a significant association between diminished sleep quality and mobile phone usage patterns\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Recent evidence also underscores a robust correlation between mobile phone dependency and the prevalence of excessive daytime sleepiness\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Furthermore, there is emerging evidence suggesting the link between the overuse of mobile phone and the impairment of cognitive functions\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eDementia, which is characterized as impaired memory, language, problem-solving abilities, and an overall decline in cognitive function, stands as a prevalent cause of disability and mortality among the elderly\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. As China progressively transitions into an aging society, with the elderly population reaching 14% of the total populace and projected to escalate to 22% by 2033\u003csup\u003e\u003cb\u003e7\u003c/b\u003e\u003c/sup\u003e, the imperative for early detection and intervention in dementia thus becomes increasingly critical. Given the absence of a curative treatment, the emphasis is on the early identification of cognitive decline, underpinned by the recognition that sleep is integral to cognitive performance\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Suboptimal sleep quality is increasingly linked to cognitive deficits\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, while excessive daytime sleepiness is recognized to exert adverse effects on cognitive functions\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Consequently, precise evaluation of cognitive impairment, sleep quality, and the severity of daytime sleepiness symptoms is essential for devising preventive strategies aimed at pinpointing modifiable risk factors, thereby potentially decelerating or mitigating the advancement of dementia.\u003c/p\u003e \u003cp\u003eIn clinical practice, the Montreal Cognitive Assessment (MoCA) \u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e is commonly utilized to assess the severity of cognitive impairment. The Epworth Sleepiness Scale (ESS) is employed to gauge the severity of excessive daytime sleepiness symptoms\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, while the subjective sleep quality can be assessed by either Pittsburg Sleep Quality Index (PSQI) or the Insomnia severity index (ISI) \u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. However, a significant drawback of these assessment tools is their reliance on subjective reports, which may lack precision, accuracy, and reliability in certain contexts\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. This shortfall has intensified the quest for more objective and precise diagnostic techniques to evaluate the severity of cognitive impairment, subjective sleep quality, and daytime sleepiness symptoms.\u003c/p\u003e \u003cp\u003eMachine learning, encompassing the development of predictive models and the discovery of meaningful patterns within datasets through computational techniques, offers an objective approach to data analysis\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Neuroimaging modalities present significant potential for elucidating the nexus between sleep disorders and the risk of dementia in vivo\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Perivascular spaces (PVSs), which are fluid-filled cavities encircling penetrating cerebral arterioles and venules, are hypothesized to facilitate a drainage network crucial for the clearance of metabolic byproducts and cerebrospinal fluid from the brain\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Enlarged perivascular spaces (EPVS), detectable through MRI as indicators of PVS dysfunction\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, have been traditionally viewed as a physiological variant across a broad age spectrum. However, an excessive burden of PVSs is correlated with prevalent neurological conditions\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. A substantial body of evidence implicates a negative correlation between PVSs and cognitive functions, as well as sleep processes. A pivotal population-based study has established that the presence of EPVS in the basal ganglia and white matter correlates with a notably elevated risk of developing dementia\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Although research on the association between PVSs and excessive daytime sleepiness is nascent, indirect evidence points to perivascular space dysfunction during sleep disruption\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, which is intricately associated with sleep quality. Previous studies have indicated morphological changes, characterized by the enlargement of basal ganglia-PVSs, in individuals experiencing persistent poor sleep quality following coronavirus disease\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. The quantification of PVSs is increasingly being addressed through automated methodologies\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, which may enhance the precision and objectivity of assessment in this research domain.\u003c/p\u003e \u003cp\u003eThis study aims to develop a predictive model utilizing MRI-quantified PVS metrics and machine learning to evaluate the severity of cognitive impairment, self-reported subjective sleep quality, and the intensity of excessive daytime sleepiness symptoms in young adults with LTMPU. Through this innovative methodology, we hope to elucidate the potential correlation between PVS and cognitive, sleep quality, and excessive daytime sleepiness in individuals addicted to mobile phone use. This research may pave the way for more precise, objective assessments and could potentially inform preventive strategies and interventions in this demographic.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Participants\u003c/h2\u003e \u003cp\u003eThis study was a school-based cross-sectional study, which was conducted from October 2021 to May 2022. A total of 165 students and young teachers aged 18 to 50 years in a medical college in Wen jiang District, Chengdu, China were recruited in this study. Among them, 146 (88.5%) responded with valid data. Questionnaires were distributed to the students and young teachers during the class period. This study was approved by the ethics committee of the Hospital of Chengdu University of Traditional Chinese Medicine, and all the research protocols and strategies were performed in accordance with the relevant guidelines and regulations.\u003c/p\u003e \u003cp\u003eThe inclusion criteria in this study were as follows: (a) with LTMPU. The duration of mobile phone use per day was obtained by the following question: How long do you usually spend on using a mobile phone per day? The response categories for this question were: less than 2 hours, 2 to 4 hours, 4 to 6 hours, and more than 6 hours. long-time mobile phone use (LTMPU) was defined as using a mobile phone\u0026thinsp;\u0026ge;\u0026thinsp;4 hours per day in consideration of the recent findings\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. (b) ethnic Han. (c) free of any psychoactive medication at least 2 weeks before and during the study. (d) right-handedness assessed with the Edinburgh Handedness Inventory\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Exclusion criteria in this study were as follows: (a) with coronavirus disease 2019 (COVID-19) infections; (b) with any significant neuropsychiatric disease or brain structural abnormality; (c) with MRI contraindications.\u003c/p\u003e \u003cp\u003eFurthermore, to evaluate cognitive and sleep status, all participants were asked to complete the MoCA, the ESS, the PSQI, and the ISI. The severity of cognitive impairment was assessed by the MoCA. The total score of MoCA is in the range of 0 to 30. when the score falls below 26, cognitive impairment is present. The lower the MoCA score is, the worse the cognitive function\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. The severity of excessive daytime sleepiness symptoms was assessed by the ESS. The total score of ESS is in the range of 0 to 24. An ESS score of more than 6, 11, and 16 was defined as sleepiness, excessive sleepiness, and risky sleepiness, respectively\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. The severity of subjective sleep quality was assessed by the PSQI. The total score of PSQI is in the range of 0 to 21. A score\u0026thinsp;\u0026gt;\u0026thinsp;5 suggests poor sleep quality\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. The severity of subjective sleep quality was assessed by the ISI. The total of ISI is in the range of 0 to 28. An ISI score\u0026thinsp;\u0026le;\u0026thinsp;7 indicates absence of insomnia; 8\u0026ndash;14 indicates sub-threshold insomnia; 15\u0026ndash;21 indicates moderate insomnia; 22\u0026ndash;28 indicates severe insomnia\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eSleep quality is a complex, multifaceted construct that poses challenges for objective quantification due to inter-individual variability and its inherently subjective nature. The PSQI and ISI are two commonly used instruments of subjective self-report sleep quality. The PSQI, a widely recognized questionnaire for gauging subjective sleep quality, has demonstrated robust reliability and validity, particularly in known-group comparisons. However, concerns regarding its factor model, the large recall period, and the scoring system challenge the value of the global PSQI score for distinguishing poor and good sleepers. The ISI, on the other hand, quantifies perceived insomnia severity by focusing on the level of disturbance to the sleep pattern, consequences of insomnia, and the degree of concern and distress related to the sleep problem. The ISI has exhibited significant correlations with various sleep questionnaires icnluding PSQI (albeit with low correlation coefficients with ESS), as well as with psychological, health, and psychopathological assessments. Future studies are needed to clarify the factor structure of ISI. In our study, PSQI and ISI are utilized to evaluated the severity of subjective sleep quality.\u003c/p\u003e \u003cp\u003eAt baseline, 91 out of 146 participants (62.3%) reported using a mobile phone\u0026thinsp;\u0026ge;\u0026thinsp;4 hours per day (LTMPU). Each participant with LTMPU completed informed written consent before undergoing magnetic resonance (MR) imaging (within two weeks of completing the scale). Nine participants were excluded because of MRI motion artifacts. Finally, 82 participants with LTMPU were included.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 MR Imaging\u003c/h2\u003e \u003cp\u003eAll patients were examined using a 3.0 T whole-body scanner (Discovery MR750, GE Healthcare, Milwaukee, WI) equipped with a 32-channel phased array head coil. T2-weighted images (T2WI) acquisition parameters were: TR\u0026thinsp;=\u0026thinsp;5613 ms, TE\u0026thinsp;=\u0026thinsp;116 ms, slice thickness\u0026thinsp;=\u0026thinsp;5.0 mm, slice spacing\u0026thinsp;=\u0026thinsp;1.5 mm, FOV\u0026thinsp;=\u0026thinsp;26 mm. T2 FLAIR acquisition parameters were: repetition time\u0026thinsp;=\u0026thinsp;8400 ms, echo time\u0026thinsp;=\u0026thinsp;150 ms, flip angle\u0026thinsp;=\u0026thinsp;111\u0026deg;, FOV\u0026thinsp;=\u0026thinsp;24 cm \u0026times; 24 cm, matrix size\u0026thinsp;=\u0026thinsp;256 \u0026times; 256, inversion time\u0026thinsp;=\u0026thinsp;2100 ms, slice thickness\u0026thinsp;=\u0026thinsp;5.0 mm with no gap between slices. 3D T1-weighted imaging (T1WI) was acquired using spoiled gradient echo sequence with repetition time\u0026thinsp;=\u0026thinsp;2.9 ms, echo time\u0026thinsp;=\u0026thinsp;3.0 ms, inversion time\u0026thinsp;=\u0026thinsp;450 ms, flip angle\u0026thinsp;=\u0026thinsp;8\u0026deg;, slice thickness\u0026thinsp;=\u0026thinsp;1 mm, matrix\u0026thinsp;=\u0026thinsp;250 \u0026times; 250, FOV\u0026thinsp;=\u0026thinsp;22 cm \u0026times; 22 cm.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Data preprocessing and PVS quantification\u003c/h2\u003e \u003cp\u003eThe image preprocessing procedure consisted of several steps, as outlined below. First, N4 bias field corrections were applied to both T1WI and T2WI to remove magnetic field inhomogeneity. Next, grayscale values were standardized by normalizing intensities to the range of [-1, 1] through clipping at 0.1%-99.9%. Utilizing a deep learning model (VB-Net\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e embedded in an image analysis tool named uAI research portal (United Imaging Intelligence)\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, the skull was removed from T1WI and the whole brain was segmented into 109 regions of interest (ROIs) based on the DK atlas\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e.These regions were then consolidated into 17 brain subregions detailed in \u003cb\u003eSupplementary Table\u0026nbsp;1\u003c/b\u003e, including bilateral frontal lobes, parietal lobes, occipital lobes, temporal lobes, basal ganglia, cerebellum, thalamus, centrum semiovale, and brainstem. Subsequently, PVS lesions were automatically segmented from the T2WI image using a built-in VB-Net model\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e with AI-generated masks reviewed and modified by two experienced radiologists as necessary. Furthermore, T1WI and T2WI images were co-registered using a registration algorithm\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, transforming the segmentation mask from the T1WI space to the T2WI space. Finally, a comprehensive analysis was conducted, computing a total of 70 quantitative metrics of PVS lesions. These metrics encompassed the total number and total volume of PVS lesions in the whole brain, as well as the number, volume, average length, and average curvature of PVS lesions for each brain subregion.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Radiomics analysis\u003c/h2\u003e \u003cp\u003eMachine learning-based radiomics analysis was used to investigate the ability of PVS characteristics to predict cognitive impairment, sleep quality, insomnia, and sleepiness symptoms in young adults with LTMPU. The radiomics pipeline was performed \u003cem\u003evia\u003c/em\u003e the uAI research portal (United Imaging Intelligence)\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, mainly consisting of feature selection, model construction, and performance evaluation.\u003c/p\u003e \u003cp\u003e \u003cem\u003eData grouping.\u003c/em\u003e Among 82 participants, 80% served as the training dataset, used for feature selection and model construction. The rest 20% served as the testing dataset, used to evaluate the robustness and generalizability of the model.\u003c/p\u003e \u003cp\u003e \u003cem\u003eFeature selection.\u003c/em\u003e The 70 PVS quantitative features in conjunction with 2 available clinical features (i.e., sex and age) served as the input to identify the most valuable biomarkers for clinical outcomes. Notably, feature standardization was first conducted to eliminate the effect of magnitudes between different features. Then, the minimum redundancy maximum relevance (mRMR) method was employed to select the most relevant feature combinations.\u003c/p\u003e \u003cp\u003e \u003cem\u003eModel construction.\u003c/em\u003e Based on the selected features, multiple machine learning algorithms (e.g., support vector machine [SVM], random forest [RF], logistic regression [LR], and K nearest neighbors [KNN], decision tree [DT], Gaussian process [GP]) were used to construct the classification models. For each classification task, we retained the model with the highest discriminative performance, where the GP model was used for the MoCA and ISI classification, and the DT model for the PSQI and ESS classification. The hyperparameters of each model are detailed in \u003cb\u003eSupplementary Table\u0026nbsp;2\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003cem\u003eModel evaluation.\u003c/em\u003e The performance of models was evaluated in the testing dataset, which could reflect the robustness and generalizability of models. The receiver operating characteristic (ROC) curve was first plotted, where the area under the curve (AUC) could be calculated quantitatively. Five metrics were calculated to evaluate the consistency between the actual label and predictive label, including accuracy, sensitivity, specificity, precision, and F1-score. These metrics were defined as follows \u003cb\u003e(\u003c/b\u003eEquations \u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Equ5\" class=\"InternalRef\"\u003e5\u003c/span\u003e):\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:Accuracy=\\:\\frac{TP+TN}{TP+PF+TN+FN}\\:,\\:$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:Sensitivity=\\:Recall=\\:\\frac{TP}{TP+FN\\:}\\:,$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:Specificity=\\frac{TN}{TN+FP\\:}\\:,$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:Precision=\\frac{TP}{TP+FP\\:}\\:,$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\:F1score=\\frac{2\\ast\\:Precision\\ast\\:Recall}{Precision+Recall\\:}\\:,$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere TP represented true positive, TN represented true negative, FP represented false positive, and FN represented false negative. Calibration curves were also used to compare the predictive output and the actual outcome. Finally, the decision curves were utilized to show the clinical net benefit for predicting outcomes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Statistical analysis\u003c/h2\u003e \u003cp\u003eThe Shapiro-Wilk tests were used to check the normal distribution of continuous variables. For continuous variables that were approximately normally distributed, they were represented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation. For continuous variables with asymmetrical distributions, they were represented as median (25th, 75th percentiles). Categorical variables were represented as counts (percentages), and compared using chi-square tests. The correlation analysis utilized Pearson\u0026rsquo;s method when both variables satisfied normal distribution assumptions; otherwise, Spearman\u0026rsquo;s method was applied. To evaluate the classification performance of machine learning models, six quantitative metrics (i.e., AUC, accuracy, sensitivity, specificity, precision, and F1-score) were calculated. All statistical analyses were implemented using SPSS (version 26.0, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ibm.com/spss\u003c/span\u003e\u003cspan address=\"https://www.ibm.com/spss\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) and R (version 4.2.2, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.R-project.org\u003c/span\u003e\u003cspan address=\"https://www.R-project.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). All figures were plotted using GraphPad Prism 9 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.graphpad.com/\u003c/span\u003e\u003cspan address=\"https://www.graphpad.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), Origin 2021 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.originlab.com/\u003c/span\u003e\u003cspan address=\"https://www.originlab.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), and Adobe Illustrator CC 2019 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.adobe.com/products/illustrator.html\u003c/span\u003e\u003cspan address=\"https://www.adobe.com/products/illustrator.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e).\u003c/span\u003e\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003e3.1 Participants characteristics\u003c/p\u003e\n\u003cp\u003eWe recruited 82 participants who underwent MRI examinations from the Affiliated Hospital of Chengdu University of Traditional Chinese Medicine between October 2021 and May 2022. The demographics and clinical scales of each participant were collected and presented in \u003cstrong\u003eTable 1\u003c/strong\u003e. The median age of all participants is 38.0 years \u003cstrong\u003e(Figure 1a)\u003c/strong\u003e, with 29.3% (24/82) being male. The distribution of cognitive impairment, poor sleep quality, insomnia, and sleepiness among all participants is visualized in \u003cstrong\u003eFigure 1b\u003c/strong\u003e, with occurrences of 55, 54, 36, and 40, respectively\u003cstrong\u003e\u0026nbsp;(Supplementary Figure 1)\u003c/strong\u003e. It demonstrates that one participant suffers from multiple disorders at the same time, revealing an intrinsic correlation among them. The demographics of participants in each disorder are summarized in \u003cstrong\u003eTable 1\u003c/strong\u003e. Notably, the median age of the cognitive normalization group (MoCA\u0026nbsp;\u0026ge;\u0026nbsp;26) is 33.0 years, whereas the median age of the cognitive impairment group (MoCA \u0026lt; 26) is 40.0 years, which is a significant difference between the two ages. There are no significant differences in sex distribution between the cognitive normalization and cognitive impairment groups, between the good sleep (PSQI \u0026le; 5) and poor sleep (PSQI \u0026gt; 5) groups, between the non-insomnia (ISI \u0026le; 7) and insomnia (ISI \u0026gt; 7) groups, and between the non-sleepiness (ESS \u0026le; 6) and sleepiness (ESS \u0026gt; 6) groups.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Demographics of participants.\u003c/strong\u003e The continuous variables between the non-anxiety and the anxiety, and between the non-depression and depression were compared using Mann-Whitey \u003cem\u003eU\u003c/em\u003e tests. The sex distribution was compared using the chi-square test. A two-tailed \u003cem\u003ep-value\u003c/em\u003e \u0026lt; 0.05 was considered a significant difference.\u003c/p\u003e\n\u003cp\u003e3.2 Gaussian process model in predicting cognitive impairment\u003c/p\u003e\n\u003cp\u003eTo explore the ability of PVS features to predict cognitive impairment severity, subjective sleep quality, and excessive daytime sleepiness symptoms severity in young adults with LTMPU, machine learning-based radiomics analyses are conducted. A total of 70 PVS features combined with easily accessible participant demographics (i.e., sex, age) are used as inputs to select the most valuable features to construct the machine learning model.\u003c/p\u003e\n\u003cp\u003eThe cognitive function can be classified into two categories based on MoCA scores, with MoCA \u0026ge; 26 being the cognitive normalization group and MoCA \u0026lt; 26 being the cognitive impairment group. To identify participants with cognitive impairment, the mRMR method is used to select the six most valuable features \u003cstrong\u003e(Figure 2a)\u003c/strong\u003e, whose correlation matrix is shown in \u003cstrong\u003eFigure 2b\u003c/strong\u003e. It is clear that three pairs of features are significantly correlated. The distribution of the six features used to construct the machine learning model in the training and testing datasets is shown in \u003cstrong\u003eSupplementary Figure 1\u003c/strong\u003e. There are significant differences between the cognitively normal and cognitively impaired groups in terms of age \u003cstrong\u003e(Supplementary Figure 1a)\u0026nbsp;\u003c/strong\u003eand the average length of PVS lesions in the left centrum semiovale \u003cstrong\u003e(Supplementary Figure 1f)\u003c/strong\u003e. Subsequently, a classification model is constructed using the Gaussian process (GP) algorithm, with receiver operating characteristic (ROC) curves plotted in \u003cstrong\u003eFigure 2c\u003c/strong\u003e. Specifically, the area under the ROC curve (AUC) values of the GP model are 0.949 with a 95% confidence interval (CI) of 0.900-0.998 and 0.818 (95% CI 0.610-1.000) in the training and testing datasets, respectively. The calibration curves show that the positive incidence predicted by the GP model deviates somewhat from the actual incidence of cognitive impairment in the training and testing datasets, indicating that the accuracy of the model prediction needs to be further improved\u003cstrong\u003e\u0026nbsp;(Figure 2d)\u003c/strong\u003e. Nonetheless, the model is still able to achieve a net clinical benefit within a threshold range of 0.3 to 0.9 \u003cstrong\u003e(Figure 2e)\u003c/strong\u003e. Additionally, the classification performance of the GP model is evaluated by the other five quantitative metrics, as detailed in \u003cstrong\u003eTable 2\u003c/strong\u003e, calculated from the confusion matrix \u003cstrong\u003e(Figures 2f, 2h)\u003c/strong\u003e. It is easy to find that the specificity and precision are higher than 0.80 in both the training and testing datasets \u003cstrong\u003e(Figures 2g, 2i)\u003c/strong\u003e, while the sensitivity is only 0.636 in the testing dataset, which implies that there is a certain degree of false-negative rate for the GP model.\u003c/p\u003e\n\u003cp\u003eTable 2. Performance of four machine learning models in predicting cognitive and sleep disorders.\u003c/p\u003e\n\u003cp\u003e3.3 Decision tree model in predicting subjective sleep quality(PSQI)\u003c/p\u003e\n\u003cp\u003eRadiomics analysis is also used to categorize participants with poor subjective sleep quality (PSQI \u0026gt; 5) and good subjective sleep quality (PSQI\u0026nbsp;\u0026le;\u0026nbsp;5). Six features are chosen through the mRMR method (\u003cstrong\u003eFigure 3a\u003c/strong\u003e), and the correlation matrix, displayed in \u003cstrong\u003eFigure 3b\u003c/strong\u003e, reveals significant correlations between five pairs of features. The distribution of the six features used to construct the machine learning model in the training and testing datasets is shown in \u003cstrong\u003eSupplementary Figure 2\u003c/strong\u003e. A decision tree (DT) algorithm is employed to build a classification model, with ROC curves depicted in \u003cstrong\u003eFigure 3c\u003c/strong\u003e. The DT model exhibits AUC values of 0.865 (95% CI 0.770 \u0026ndash; 0.959) and 0.826 (95% CI 0.616 \u0026ndash; 1.000) in the training and testing datasets, respectively. The calibration curve demonstrates strong agreement between the predicted positive incidence rates by the DT model and the actual incidence rates of poor\u0026nbsp;subjective sleep quality in the training dataset, with slight deviation observed in the testing dataset (\u003cstrong\u003eFigure 3d\u003c/strong\u003e). Overall, the model yields high prediction accuracy rates of 0.846 and 0.824 in the two datasets, as shown in \u003cstrong\u003eTable 2\u003c/strong\u003e. Furthermore, decision curves illustrated in \u003cstrong\u003eFigure 3e\u003c/strong\u003e indicate that the model delivers substantial clinical net benefits across a range of thresholds (0.2-1.0) in both training and testing datasets, showcasing its potential for enhancing patient care and decision-making support effectively. Detailed quantitative metrics from confusion matrices visualized in \u003cstrong\u003eFigures 3f, 3g\u003c/strong\u003e for training data and \u003cstrong\u003eFigures 3h, 3i\u003c/strong\u003e for testing datasets are presented in \u003cstrong\u003eTable 2\u003c/strong\u003e. Overall, except for specificity, all metrics surpass a threshold of above 0.8 in both datasets, indicating a relatively high prediction performance by the model despite some false positives being present.\u003c/p\u003e\n\u003cp\u003e3.4 Gaussian process model in predicting subjective sleep quality(ISI)\u003c/p\u003e\n\u003cp\u003eSimilar feature selection and modeling procedures are conducted to distinguish between the non-insomnia group (ISI \u0026le; 7) and the insomnia group (ISI \u0026gt; 7). Using the mRMR method, a total of six features are selected \u003cstrong\u003e(Figure 4a)\u003c/strong\u003e, with the corresponding correlation matrix presented in \u003cstrong\u003eFigure 4b\u003c/strong\u003e, indicating no significant correlation among the selected features, underscoring their unique and independent information contribution. The distribution of the six features used to construct the machine learning model in the training and testing datasets is shown in \u003cstrong\u003eSupplementary Figure 3\u003c/strong\u003e. Subsequently, a classification model is built using the GP algorithm. The AUC values for the GP model are calculated as 0.947 (95% CI 0.888-1.000) in the training dataset and 0.757 (95% CI 0.492-1.000) in the testing dataset, as shown in \u003cstrong\u003eFigure 4c\u003c/strong\u003e. Although calibration curves show narrower prediction intervals and acceptable accuracy \u003cstrong\u003e(Figure 4d)\u003c/strong\u003e, decision curves reveal a clinical net benefit in both datasets, particularly within a threshold range of 0.3-0.6, emphasizing the model\u0026apos;s potential in informing clinical decisions\u003cstrong\u003e\u0026nbsp;(Figure 4e)\u003c/strong\u003e. The confusion matrices depicted in \u003cstrong\u003eFigure 4f\u003c/strong\u003e for the training dataset and \u003cstrong\u003eFigure 4h\u003c/strong\u003e for the testing dataset are utilized to compute quantitative metrics such as accuracy, F1-score, sensitivity, specificity, and precision. These metrics for both datasets are graphically presented in \u003cstrong\u003eFigure 4g\u003c/strong\u003e and \u003cstrong\u003eFigure 4i\u003c/strong\u003e. While the model displays robust precision and specificity, indicating its proficiency in identifying true negatives accurately, its lower sensitivity suggests a higher false negative rate that may lead to missed true positive cases.\u003c/p\u003e\n\u003cp\u003e3.5 Decision tree model in predicting excessive daytime sleepiness symptoms\u003c/p\u003e\n\u003cp\u003eFeature selection and modeling procedures are applied to differentiate between the non-sleepiness group (ESS \u0026le; 6) and the sleepiness group (ESS \u0026gt; 6). Six features are selected using the mRMR method, as illustrated in \u003cstrong\u003eFigure 5a\u003c/strong\u003e, with the correlation matrix presented in \u003cstrong\u003eFigure 5b\u003c/strong\u003e showing no significant correlations among these selected features. This absence of significant correlation highlights the unique and independent information contributed by each feature. The distribution of the six features used to construct the machine learning model in the training and testing datasets is shown in \u003cstrong\u003eSupplementary Figure 4\u003c/strong\u003e. There is a significant difference between the non-sleepiness and sleepiness groups in terms of the average length of PVS lesions in the left centrum semiovale \u003cstrong\u003e(Supplementary Figure 4h)\u003c/strong\u003e. Following this, a classification model is constructed using the DT algorithm. The AUC values for the DT model are determined as 0.923 (95% CI 0.867-0.978) in the training dataset and 0.875 (95% CI 0.718-1.000) in the testing dataset, as depicted in \u003cstrong\u003eFigure 5c\u003c/strong\u003e. The calibration curves demonstrate a strong alignment between the model\u0026rsquo;s predicted likelihood of sleepiness and the actual prevalence of sleepiness in both the training and testing datasets, as shown in \u003cstrong\u003eFigure 5d\u003c/strong\u003e. This alignment underscores the model\u0026apos;s ability to accurately estimate an individual\u0026rsquo;s probability of belonging to the sleepiness group across different datasets, indicating its reliability in evaluating sleep disorders. Furthermore, decision curves showcase that the model provides clinical utility and benefit across a wide range of thresholds (0.1-0.8), suggesting its potential positive impact on clinical decision-making processes \u003cstrong\u003e(Figure 5e)\u003c/strong\u003e. Quantitative metrics such as accuracy, F1-score, sensitivity, specificity, and precision are computed based on confusion matrices depicted in \u003cstrong\u003eFigure 5f\u003c/strong\u003e for the training dataset and \u003cstrong\u003eFigure 5h\u003c/strong\u003e for the testing dataset. These metrics for both datasets are visually represented in \u003cstrong\u003eFigure 5g\u003c/strong\u003e and \u003cstrong\u003eFigure 5i\u003c/strong\u003e. Overall performance evaluation reveals that except for sensitivity, all metrics exceed a threshold of 0.8 in both datasets, indicating a strong performance across various assessment criteria.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eTo our knowledge, this paper is the first study that presents a novel approach to classify cognitive impairment severity, subjective sleep quality, and excessive daytime sleepiness symptoms severity in young adults with LTMPU by integrating MRI-based quantification of PVS and machine learning algorithms. Our model has exhibited remarkable accuracy in these classifications, presenting a promising path for non-invasive and objective assessment methodologies. The integration of MRI data with advanced computational models represents a significant advancement in the field, as no prior studies have been known to harness machine learning to such an end with respect to sleep and cognitive function. Furthermore, our research provides preliminary evidence suggesting a correlation between PVS and subjective symptoms of excessive daytime sleepiness, an area that has received scant attention in previous studies, as far as we know.\u003c/p\u003e \u003cp\u003ePVSs are integral to the functionality of the glymphatic system, a network responsible for brain waste clearance\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. Under normal conditions, PVSs are not discernible on structural MRI; however, the visibility of EPVS may indicate a dysfunction in glymphatic clearance\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. It is plausible to hypothesize that EPVS may be associated with poor sleep quality and excessive daytime sleepiness, given the influence of sleep on glymphatic system efficacy\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, although current evidence is limited. An early study has demonstrated that poor sleep quality was independently associated with the EPVS in basal ganglia and white matter\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. The glymphatic system, with PVSs as a key component, plays a crucial role in the clearance of brain amyloid β (Aβ) and tauopathy, which are linked to neurodegenerative conditions\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. A 2019 study has showed that a substantial perivascular space burden is associated with common neurological diseases, such as Alzheimer\u0026rsquo;s disease\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. The quantification of PVSs has been enhanced through computational methods, which have shown heightened sensitivity in associating PVSs with white matter hyperintensities and retinal vessel diameters\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. In our current research, we referenced a previous study for the precise measurement of PVS quantification\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e, thereby aiming to advance the understanding of PVSs in the context of subjective sleep quality, cognitive function, and their potential implications in neurodegenerative processes.\u003c/p\u003e \u003cp\u003eIn recent years, many studies have employed machine learning techniques to develop predictive models for classifying the severity of cognitive impairment\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. This study represents the first attempt to utilize MRI quantified EPVS volumes and machine learning to accurately classify subjective sleep quality and the severity of excessive daytime sleepiness symptoms in young adults with LTMPU, which may hold significant potential for clinical applications. Our study provides preliminary evidence suggesting a relationship between EPVS, a biomarker indicative of glymphatic dysfunction, and the severity of cognitive impairment, sleep quality, and the severity of excessive daytime sleepiness symptoms in young adults with LTMPU. This is in line with the burgeoning body of research that points to a bidirectional relationship between sleep disturbances and the risk of dementia\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. The absence of curative treatments for dementia underscores the critical need for preventive interventions\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. By integrating MRI-quantified EPVS volumes with machine learning algorithms, our model may offer insights into the early stages of Alzheimer\u0026rsquo;s disease, potentially identifying syndromal conversion in cognitively unimpaired subjects\u0026mdash;a domain where data are exceedingly scarce. This approach harnesses the power of neuroimaging to detect preclinical neurodegenerative changes, facilitating both the early diagnosis of Alzheimer\u0026rsquo;s and the monitoring of sleep health. Excessive daytime sleepiness is a public health issue, which is often undervalued, infrequently diagnosed, and inadequately addressed\u003csup\u003e\u003cb\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/b\u003e\u003c/sup\u003e. The observed relationship between EPVS and excessive daytime sleepiness symptoms in our study suggests that EPVS could serve as a promising biomarker for this condition. Further exploration of this association could deepen our understanding of the neurobiology underlying excessive daytime sleepiness, ultimately aiding in the development of improved diagnostic and therapeutic strategies for affected patients.\u003c/p\u003e \u003cp\u003eThis study, while pioneering in its approach, has several limitations. Firstly, the modest sample size employed may restrict the generalizability of our findings, warranting a need for validation on a larger scale to enhance the model's robustness and applicability. Secondly, our observational design does not permit the determination of causality between the persistent symptoms of excessive daytime sleepiness and the presence of EPVS. Additionally, the study\u0026rsquo;s single-center nature necessitates broader validation across diverse populations and settings to ensure the findings are representative and reliable. Furthermore, the current research did not include an analysis of EPVS location, which is imposed by the sample size. Future studies should aim to incorporate a more detailed examination of EPVS distribution to provide a comprehensive understanding of their potential impact on cognitive and sleep-related symptoms.\u003c/p\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eOur study introduces an innovative analytical framework by integrating MRI-quantified PVS metrics with machine learning algorithms, offering a new methodological paradigm for classifying the severity of cognitive impairment, subjective sleep quality and excessive daytime sleepiness symptoms in young adults with LTMPU. The insights gained from this preliminary investigation set the stage for more extensive inquiries into the complex interplay between EPVS, cognitive function, and sleep quality in the context of LTMPU. Further research with expanded cohorts and multi-centric approaches are imperative for substantiating the reliability and generalizability of our model. Such endeavors will be pivotal in validating the predictive capabilities of our model across various populations and healthcare settings, thereby enhancing its potential impact on clinical diagnostics and therapeutic interventions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the National Natural Science Foundation of China (grant number:\u0026nbsp;82074303, 82174345, 81973684), Sichuan Key Research and Development Project (2023YFS0226), and Natural Science Foundation of Sichuan Province (2023NSFSC1760).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJ.W., R.H., and F.S. are employees of United Imaging Intelligence. The company has no role in designing and performing the surveillance and analyzing and interpreting the data. All other authors report no conflicts of interest relevant to this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDisclosure statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors report no biomedical financial interests or potential conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003estatement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analysed during this study are included in this published article and its supplementary information files.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eJoshi, S. C., Woltering, S. \u0026amp; Woodward, J. Cell Phone Social Media Use and Psychological Well-Being in Young Adults: Implications for Internet-Related Disorders. \u003cem\u003eInt J Environ Res Public Health\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, (2023).\u003c/li\u003e\n \u003cli\u003eLiu, S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e The associations of long-time mobile phone use with sleep disturbances and mental distress in technical college students: a prospective cohort study. \u003cem\u003eSleep\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, (2019).\u003c/li\u003e\n \u003cli\u003eBrautsch, L. A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Digital media use and sleep in late adolescence and young adulthood: A systematic review. \u003cem\u003eSleep Med Rev\u003c/em\u003e \u003cstrong\u003e68\u003c/strong\u003e, 101742 (2023).\u003c/li\u003e\n \u003cli\u003eNowak, M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Correlations between Problematic Mobile Phone Use and Depressiveness and Daytime Sleepiness, as Well as Perceived Social Support in Adolescents. \u003cem\u003eInt J Environ Res Public Health\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, (2022).\u003c/li\u003e\n \u003cli\u003eWacks, Y. \u0026amp; Weinstein, A. M. Excessive Smartphone Use Is Associated With Health Problems in Adolescents and Young Adults. \u003cem\u003eFront Psychiatry\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 669042 (2021).\u003c/li\u003e\n \u003cli\u003eSadeghmousavi, S., Eskian, M., Rahmani, F. \u0026amp; Rezaei, N. The effect of insomnia on development of Alzheimer\u0026apos;s disease. \u003cem\u003eJ Neuroinflammation\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 289 (2020).\u003c/li\u003e\n \u003cli\u003eZhao, M., Li, J., Xiang, L., Zhang, Z. H. \u0026amp; Peng, S. L. A diagnosis model of dementia via machine learning. \u003cem\u003eFront Aging Neurosci\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 984894 (2022).\u003c/li\u003e\n \u003cli\u003eMason, G. M., Lokhandwala, S., Riggins, T. \u0026amp; Spencer, R. M. C. Sleep and human cognitive development. \u003cem\u003eSleep Med Rev\u003c/em\u003e \u003cstrong\u003e57\u003c/strong\u003e, 101472 (2021).\u003c/li\u003e\n \u003cli\u003eLi, M., Wang, N. \u0026amp; Dupre, M. E. Association between the self-reported duration and quality of sleep and cognitive function among middle-aged and older adults in China. \u003cem\u003eJ Affect Disord\u003c/em\u003e \u003cstrong\u003e304\u003c/strong\u003e, 20-27 (2022).\u003c/li\u003e\n \u003cli\u003eSlonkova, J.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Hypocretin-1/orexin-A, sleep and excessive daytime sleepiness in patients with nonconvulsive status epilepticus: A cross-sectional cohort study. \u003cem\u003eSleep Med\u003c/em\u003e \u003cstrong\u003e119\u003c/strong\u003e, 192-200 (2024).\u003c/li\u003e\n \u003cli\u003eCai, Y.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Comparing machine learning-derived MRI-based and blood-based neurodegeneration biomarkers in predicting syndromal conversion in early AD. \u003cem\u003eAlzheimers Dement\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 4987-4998 (2023).\u003c/li\u003e\n \u003cli\u003eP\u0026eacute;rez-Carbonell, L., Mignot, E., Leschziner, G. \u0026amp; Dauvilliers, Y. Understanding and approaching excessive daytime sleepiness. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e400\u003c/strong\u003e, 1033-1046 (2022).\u003c/li\u003e\n \u003cli\u003eMcCarter, S. J.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Physiological markers of sleep quality: A scoping review. \u003cem\u003eSleep Med Rev\u003c/em\u003e \u003cstrong\u003e64\u003c/strong\u003e, 101657, doi:10.1016/j.smrv.2022.101657 (2022).\u003c/li\u003e\n \u003cli\u003eLiu, X.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Gait can reveal sleep quality with machine learning models. \u003cem\u003ePLoS One\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, e0223012 (2019).\u003c/li\u003e\n \u003cli\u003eJeon, J.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Accuracy of Machine Learning Using the Montreal Cognitive Assessment for the Diagnosis of Cognitive Impairment in Parkinson\u0026apos;s Disease. \u003cem\u003eJ Mov Disord\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 132-139 (2022).\u003c/li\u003e\n \u003cli\u003eGreener, J. G., Kandathil, S. M., Moffat, L. \u0026amp; Jones, D. T. A guide to machine learning for biologists. \u003cem\u003eNat Rev Mol Cell Biol\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 40-55 (2022).\u003c/li\u003e\n \u003cli\u003eWard, S. A. \u0026amp; Pase, M. P. Advances in pathophysiology and neuroimaging: Implications for sleep and dementia. \u003cem\u003eRespirology\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 580-592 (2020).\u003c/li\u003e\n \u003cli\u003eDing, J.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Large Perivascular Spaces Visible on Magnetic Resonance Imaging, Cerebral Small Vessel Disease Progression, and Risk of Dementia: The Age, Gene/Environment Susceptibility-Reykjavik Study. \u003cem\u003eJAMA Neurol\u003c/em\u003e \u003cstrong\u003e74\u003c/strong\u003e, 1105-1112 (2017).\u003c/li\u003e\n \u003cli\u003eDuering, M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Neuroimaging standards for research into small vessel disease-advances since 2013. \u003cem\u003eLancet Neurol\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 602-618 (2023).\u003c/li\u003e\n \u003cli\u003eWardlaw, J. M.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Perivascular spaces in the brain: anatomy, physiology and pathology. \u003cem\u003eNat Rev Neurol\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 137-153 (2020).\u003c/li\u003e\n \u003cli\u003eZhu, Y. C.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e High degree of dilated Virchow-Robin spaces on MRI is associated with increased risk of dementia. \u003cem\u003eJ Alzheimers Dis\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 663-672 (2010).\u003c/li\u003e\n \u003cli\u003eDel Brutto, O. H.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Long coronavirus disease-related persistent poor sleep quality and progression of enlarged perivascular spaces. A longitudinal study. \u003cem\u003eSleep\u003c/em\u003e \u003cstrong\u003e45\u003c/strong\u003e, (2022).\u003c/li\u003e\n \u003cli\u003eOldfield, R. C. The assessment and analysis of handedness: the Edinburgh inventory. \u003cem\u003eNeuropsychologia\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 97-113 (1971).\u003c/li\u003e\n \u003cli\u003eChen, X. Y., Chen, Z. Y., Dong, Z., Liu, M. Q. \u0026amp; Yu, S. Y. Regional volume changes of the brain in migraine chronification. \u003cem\u003eNeural Regen Res\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 1701-1708 (2020).\u003c/li\u003e\n \u003cli\u003eHou, A.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Widespread aberrant functional connectivity throughout the whole brain in obstructive sleep apnea. \u003cem\u003eFront Neurosci\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 920765 (2022).\u003c/li\u003e\n \u003cli\u003eMorin, C. M., Belleville, G., B\u0026eacute;langer, L. \u0026amp; Ivers, H. The Insomnia Severity Index: psychometric indicators to detect insomnia cases and evaluate treatment response. \u003cem\u003eSleep\u003c/em\u003e \u003cstrong\u003e34\u003c/strong\u003e, 601-608 (2011).\u003c/li\u003e\n \u003cli\u003eShi, F.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 6566 (2022).\u003c/li\u003e\n \u003cli\u003eWu, J.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e uRP: An integrated research platform for one-stop analysis of medical images. \u003cem\u003eFront Radiol\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 1153784 (2023).\u003c/li\u003e\n \u003cli\u003eDesikan, R. S.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. \u003cem\u003eNeuroimage\u003c/em\u003e \u003cstrong\u003e31\u003c/strong\u003e, 968-980 (2006).\u003c/li\u003e\n \u003cli\u003eZhang, Z.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Quantitative Analysis of Multimodal MRI Markers and Clinical Risk Factors for Cerebral Small Vessel Disease Based on Deep Learning. \u003cem\u003eInt J Gen Med\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 739-750 (2024).\u003c/li\u003e\n \u003cli\u003eAvants, B. B.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e The Insight ToolKit image registration framework. \u003cem\u003eFront Neuroinform\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 44 (2014).\u003c/li\u003e\n \u003cli\u003eDredla, B. K., Del Brutto, O. H. \u0026amp; Castillo, P. R. Sleep and Perivascular Spaces. \u003cem\u003eCurr Neurol Neurosci Rep\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 607-615 (2023).\u003c/li\u003e\n \u003cli\u003eRamirez, J.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e MRI-visible perivascular space volumes, sleep duration and daytime dysfunction in adults with cerebrovascular disease. \u003cem\u003eSleep Med\u003c/em\u003e \u003cstrong\u003e83\u003c/strong\u003e, 83-88 (2021).\u003c/li\u003e\n \u003cli\u003eYang, S., Yin, J., Qin, W., Yang, L. \u0026amp; Hu, W. Poor Sleep Quality Associated With Enlarged Perivascular Spaces in Patients With Lacunar Stroke. \u003cem\u003eFront Neurol\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 809217 (2021).\u003c/li\u003e\n \u003cli\u003eWang, X. X.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e MRI-visible enlarged perivascular spaces: imaging marker to predict cognitive impairment in older chronic insomnia patients. \u003cem\u003eEur Radiol\u003c/em\u003e \u003cstrong\u003e32\u003c/strong\u003e, 5446-5457 (2022).\u003c/li\u003e\n \u003cli\u003eFrancis, F., Ballerini, L. \u0026amp; Wardlaw, J. M. Perivascular spaces and their associations with risk factors, clinical disorders and neuroimaging features: A systematic review and meta-analysis. \u003cem\u003eInt J Stroke\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 359-371 (2019).\u003c/li\u003e\n \u003cli\u003eBallerini, L.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Computational quantification of brain perivascular space morphologies: Associations with vascular risk factors and white matter hyperintensities. A study in the Lothian Birth Cohort 1936. \u003cem\u003eNeuroimage Clin\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 102120 (2020).\u003c/li\u003e\n \u003cli\u003eLevy, B.\u003cem\u003e\u0026nbsp;et al.\u003c/em\u003e Machine Learning Enhances the Efficiency of Cognitive Screenings for Primary Care. \u003cem\u003eJ Geriatr Psychiatry Neurol\u003c/em\u003e \u003cstrong\u003e32\u003c/strong\u003e, 137-144 (2019).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 to 2 are available in the Supplementary Files section\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"MRI, imaging biomarker, cognitive impairment, sleep disturbance, LTMPU, EPVS","lastPublishedDoi":"10.21203/rs.3.rs-5384782/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5384782/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEmerging evidence has linked long-time mobile phone use (LTMPU) with cognitive impairment and sleep issues, with MRI-detected enlarged perivascular spaces (EPVSs) serving as markers for these conditions. Our study seeks to develop predictive model using MRI-based PVS measurements and machine learning to assess cognitive impairment, subjective sleep quality, and excessive daytime sleepiness in young adults with LTMPU. Eighty-two participants were included, deep learning algorithms were used to segment EPVS lesions and extract quantitative metrics. Training and testing datasets were randomly assigned to perform radiomics analysis, where EPVS metrics combined with sex and age were used to select the most valuable features for model construction. Finally, a Gaussian process model was constructed based on six features for assessing cognitive impairment, yielding an AUC of 0.818 (95% confidence interval [CI] 0.610-1) in the testing dataset. For sleep quality and sleepiness, two decision tree (DT) models using six features achieved an AUC value of 0.826 (95% CI 0.616-1) and 0.875 (95% CI 0.718-1) in the testing dataset respectively. Our study leveraged MRI-based PVS metrics and machine learning to assess the severity of cognitive impairment and sleep problems in young adults with LTMPU, and sheds light on a potential link between PVS and sleepiness.\u003c/p\u003e","manuscriptTitle":"MRI quantified perivascular space metrics as imaging biomarkers for assessing the severity of cognitive impairment and sleep disturbance in young adults with long-time mobile phone use through machine learning approaches","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-25 01:13:14","doi":"10.21203/rs.3.rs-5384782/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-03-19T06:24:31+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-02-11T13:37:25+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-11-19T12:00:45+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-11-18T15:17:59+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2024-11-04T04:34:51+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a087a80e-6235-46c2-884e-f56b55a7db12","owner":[],"postedDate":"December 25th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":40939666,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":40939667,"name":"Health sciences/Biomarkers"},{"id":40939668,"name":"Health sciences/Diseases"},{"id":40939669,"name":"Health sciences/Medical research"},{"id":40939670,"name":"Health sciences/Nephrology"},{"id":40939671,"name":"Health sciences/Risk factors"},{"id":40939672,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2026-01-19T17:21:40+00:00","versionOfRecord":{"articleIdentity":"rs-5384782","link":"https://doi.org/10.1038/s41598-026-35845-3","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-01-13 16:29:25","publishedOnDateReadable":"January 13th, 2026"},"versionCreatedAt":"2024-12-25 01:13:14","video":"","vorDoi":"10.1038/s41598-026-35845-3","vorDoiUrl":"https://doi.org/10.1038/s41598-026-35845-3","workflowStages":[]},"version":"v1","identity":"rs-5384782","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5384782","identity":"rs-5384782","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}