Introduction

Although combination antiretroviral therapy (cART) effectively suppresses human immunodeficiency virus type 1 (HIV-1) replication in the central nervous system (CNS), HIV-associated neurocognitive disorders (HAND) remain highly prevalent.1,2 HAND encompasses three clinical stages: asymptomatic neurocognitive impairment (ANI), mild neurocognitive disorder (MND), and HIV-associated dementia (HAD).3 In the cART era, the prevalence of the severe form, HAD, has markedly declined, whereas milder forms of cognitive impairment, particularly ANI and MND, have become the predominant clinical manifestations.4 Patients with these milder forms often present only with subtle subclinical manifestations, such as mild psychomotor slowing or memory decline, which are easily overlooked and thereby hinder early identification.

Currently, the diagnosis of HAND still relies primarily on standardized neuropsychological testing (NPT), which requires comprehensive assessment across multiple cognitive domains. However, this approach is procedurally complex, time-consuming, and highly dependent on specialized assessment conditions, limiting its feasibility for efficient screening of early-stage HAND in routine clinical practice.5 Notably, ANI is typically characterized by the absence of overt impairment in everyday functioning and may therefore be overlooked in routine clinical practice, even though early neuropathological alterations may already be present.6 Moreover, longitudinal evidence indicates that individuals with ANI are at an increased risk of subsequent symptomatic cognitive decline compared with cognitively intact individuals with HIV.7 Therefore, ANI may represent an important stage for early intervention and risk stratification in HAND, underscoring the clinical importance of its early identification. Accordingly, objective, sensitive, and scalable biomarkers are urgently needed to facilitate early identification and risk stratification of HAND.

In recent years, non-invasive MRI techniques have been widely applied in neuroHIV research, providing important evidence for structural and functional brain abnormalities associated with HAND. Resting-state functional MRI (rs-fMRI) characterizes brain functional activity from the perspectives of both functional segregation and functional integration. Functional segregation reflects spontaneous neural activity within local brain regions, as measured by indices such as the amplitude of low-frequency fluctuation (ALFF) and regional homogeneity (ReHo), whereas functional integration reflects interregional coordination and the organization of large-scale brain networks. Previous studies have shown that HIV infection is associated with functional abnormalities in key nodes of the default mode network (DMN), central executive network (CEN), and salience network (SN), including altered posterior DMN connectivity and aberrant SN-DMN coupling. These disruptions in inter-network coordination have been associated cognitive decline.8–11

Our group and other investigators have explored the neuroimaging basis of HIV-associated neurocognitive disorder–asymptomatic neurocognitive impairment (HIV-ANI) from multiple brain network perspectives, including whole-brain connectomic characteristics, multimodal connectomics analyses, multilayer network dynamic reconfiguration, and abnormalities in structural-functional coupling.12–15 These studies consistently indicate that HIV infection leads to widespread alterations in brain network organization, accompanied by disrupted inter-network coordination. However, most existing research has focused on abnormalities at the network or connectome level, primarily examining alterations in interregional connectivity. In contrast, it remains insufficiently understood whether the key hub nodes constituting these networks already exhibit local structural and functional abnormalities. From a brain-network perspective, highly connected hub nodes may be particularly vulnerable to pathological disturbances because of their high information-processing demands, and local structural or functional abnormalities within these hubs may further affect network-level organization and function.16,17 Therefore, characterizing local abnormalities within core nodes of the DMN, CEN, and SN may provide complementary information for understanding the large-scale network dysfunction previously reported in early HAND.

Radiomics enables the high-throughput extraction of multidimensional quantitative features from medical images, thereby characterizing within-region signal distributions and spatial heterogeneity.18,19 Conventional regions of interest (ROIs) averaged measures and functional connectivity metrics primarily characterize overall regional properties or interregional relationships, respectively, and provide limited information about within-node heterogeneity. Radiomics can provide richer quantitative information within individual nodes and may therefore complement conventional approaches in detecting subtle structural and functional alterations. Although radiomics has been applied in neurodegenerative diseases such as Alzheimer’s disease and Parkinson’s disease,20,21 its application in the field of HAND remains limited. Recent studies combining multimodal MRI with machine learning suggest that integrating structural, functional, and clinical features may improve the early identification of HAND and facilitate the identification of key brain regions associated with cognitive impairment.22

Given the important roles of core nodes within the DMN, CEN, and SN in attention, executive function, memory, and cognitive control, as well as their repeatedly reported structural and functional abnormalities in HAND, the present study selected these nodes as ROIs. Node-level radiomic analyses were performed using 3D T1-weighted images and ALFF and ReHo maps. Based on previous evidence implicating core hub regions of the DMN, CEN, and SN in HAND, We hypothesized that, compared with HIV cognitively intact (HIV-IC) individuals, individuals with HIV-ANI would exhibit detectable subtle structural and functional abnormalities within core cognitive network nodes and that these abnormalities could be characterized using multimodal MRI radiomics. Accordingly, this study aimed to: (1) identify key node-level structural and functional radiomic features associated with HIV-ANI; (2) develop and compare unimodal, multimodal, and imaging–clinical fusion machine-learning models and evaluate their potential for identifying HIV-ANI; and (3) use SHapley Additive exPlanations (SHAP) to characterize the contributions of key predictive features and their corresponding nodes and to explore the potential association between node-level abnormalities and brain network dysfunction in early HAND.

Materials and Methods Study Participants

Between June 2023 and February 2026, 67 individuals with HIV-associated asymptomatic neurocognitive impairment (HIV-ANI group) and 70 cognitively intact individuals with HIV (HIV-IC group) were enrolled. Written informed consent was obtained from all participants. The study was approved by the Ethics Committees of Beijing Youan Hospital, Capital Medical University, and the Affiliated Infectious Diseases Hospital of Zhengzhou University (approval numbers: LL-2020-047-K and IEC-KY-2025-14-02). Participants were primarily recruited from the Sexually Transmitted Diseases and AIDS Clinic of Beijing Youan Hospital and the Affiliated Infectious Diseases Hospital of Zhengzhou University.

The inclusion criteria were as follows: (1) confirmed HIV infection; (2) Han Chinese ethnicity; (3) age between 20 and 60 years; (4) right-handedness; and (5) completion of standardized neuropsychological assessment and MRI examination. The exclusion criteria were as follows: (1) central nervous system disorders that could affect cognitive function or brain MRI findings, including central nervous system tumors, infections, or cerebrovascular diseases; (2) a history of neurological or psychiatric disorders that could affect cognitive assessment; (3) a history of alcohol or other substance abuse or dependence; and (4) contraindications to MRI.

HIV-ANI was diagnosed according to the Frascati criteria, requiring performance at least 1 standard deviation below demographically adjusted normative means in at least two cognitive domains, without impairment in activities of daily living (ADL). Raw neuropsychological test scores were converted to T scores using Chinese normative data corrected for age, years of education, and demographic background, and everyday functional status was assessed using the ADL scale;3 HIV-IC was defined as HIV-positive individuals who did not meet the diagnostic threshold for any HAND subtype and no impairment in daily functioning.

Neuropsychological Testing

A comprehensive neuropsychological test battery was used to assess six major cognitive domains, with Chinese normative data corrected for age, years of education, and urban–rural background.23 The assessment battery consisted of eight subtests covering six cognitive domains: (1) speed of information processing, assessed using the Trail Making Test-A (TMT-A); (2) learning and memory, assessed using the Hopkins Verbal Learning Test-Revised (HVLT-R) and the Brief Visuospatial Memory Test-Revised (BVMT-R); (3) abstraction and executive function, assessed using the 64-card Wisconsin Card Sorting Test (WCST-64); (4) attention and working memory, assessed using the Wechsler Memory Scale-Third Edition (WMS-III) and the Paced Auditory Serial Addition Test (PASAT); (5) fine motor coordination, assessed using the Grooved Pegboard Test; and (6) language and verbal fluency, assessed using the animal naming test. Raw scores from the eight subtests were converted into T scores for the six cognitive domains according to Chinese normative data. Lower T scores indicated poorer cognitive performance, with lower values reflecting greater impairment in the corresponding domain. In addition, all individuals with HIV underwent assessment using the ADL scale to further determine whether functional impairment was present.

MRI Acquisition

MRI data for all participants were acquired using a Siemens 3.0-T MRI scanner (Magnetom Trio, Siemens Healthineers). Before study implementation, the MRI acquisition protocols and key imaging parameters at both centers were standardized according to a unified study protocol. The Zhengzhou branch center used the same scanning protocol and acquisition parameters as the Beijing center to ensure consistency in image acquisition across the two centers and to minimize potential inter-center and scanner-related variability. A 32-channel phased-array orthogonal coil was used. The subject lay quietly on the scanning bed with the head secured to a foam pad, eyes closed, remaining awake but without any systematic mental activity. The head was stabilized with the foam pad, and earplugs were worn to reduce the impact of MRI noise. A routine MRI scan was performed first, including 3D-T1WI structural images, to rule out organic intracranial lesions in the subjects. The 3D-T1WI structural images covered the entire head. The scanning sequence was a three-dimensional magnetization-prepared gradient echo sequence with a repetition time of 1800 ms; Echo time = 4.5 ms; Inversion time = 900 ms; Flip angle = 9°; Field of view 160 mm × 130 mm; Matrix 320 × 320, slice thickness 1 mm; Subsequently, a transverse scan is performed using a single-excitation planar echo sequence to acquire fMRI data, with scan parameters: repetition time = 2000 ms; Echo time = 20 ms; Flip angle = 90°; Field of view 160 mm × 128 mm; Matrix 82 × 82, slice thickness 2 mm, number of scan periods 180, scanning 240 time points.

Clinical Data Collection

Clinical and immunological data were systematically collected from electronic health records (EHR), including age, educational level, CD4 count, CD4 nadir, duration of HIV infection, and duration of antiretroviral therapy.

Resting-State fMRI Preprocessing

Resting-state fMR data were preprocessed using the Data Processing and Analysis for Brain Imaging toolbox, version 5.1 (DPABI V5.1).24 The main preprocessing steps included DICOM-to-NIfTI conversion, removal of the first 10 time points, slice-timing correction, head-motion correction, coregistration with T1-weighted images, DARTEL-based spatial normalization, nuisance regression, and band-pass filtering at 0.01–0.10 Hz. Nuisance covariates included linear drift, white matter signal, cerebrospinal fluid signal, global signal, and Friston 24-parameter head-motion regressors. Mean framewise displacement (mean FD) was included as a covariate in group-level analyses to further minimize the influence of head motion;25–27Participants with head displacement >1.0 mm in any direction or rotation >1.0°during the scan were excluded. Standardized ReHo (mReHo) and ALFF (mALFF) maps were then calculated. ReHo was computed using Kendall’s coefficient of concordance to assess the temporal similarity between each voxel and its 26 neighboring voxels;28 whereas ALFF reflected the amplitude of low-frequency signal fluctuations within the 0.01–0.10 Hz frequency band in specific brain regions during the resting state29. For ALFF analysis, spatial smoothing was performed using a 4-mm full width at half maximum (FWHM) Gaussian kernel before ALFF calculation. ReHo was calculated from the unsmoothed functional images, and the resulting ReHo maps were subsequently smoothed using a 4-mm FWHM Gaussian kernel. ALFF and ReHo values were then normalized to generate mean-normalized ALFF (mALFF) and ReHo (mReHo) maps.

Imaging Feature Extraction

The 3D T1WI, mALFF, and mReHo images used for feature extraction were spatially normalized and transformed into the same standard space as the Automated Anatomical Labeling (AAL) atlas. Anatomical parcellation was performed using the AAL atlas, and core network nodes of the default mode network (DMN), central executive network (CEN), and salience network (SN) were predefined based on prior literature. The corresponding AAL labels were used to define the regions of interest (ROIs), including the cingulate cortex, precuneus, and medial prefrontal cortex within the DMN; the dorsolateral prefrontal cortex and posterior parietal cortex within the CEN; and the insula and anterior cingulate cortex within the SN.

The predefined ROIs were consistently applied to the 3D T1WI, mALFF, and mReHo images in standard space. Quantitative imaging features within each ROI were automatically extracted using an in-house Python script (Python version 3.10.11). NIfTI images were loaded using NiBabel (version 5.3.3), and voxel values within each predefined ROI were extracted. First-order statistical features were subsequently calculated using NumPy (version 2.2.6). The same spatial normalization, ROI definition, spatial mapping, and feature-extraction procedures were applied uniformly to all participants. No manual ROI delineation or adjustment was involved, thereby minimizing operator-related variability and improving the reproducibility of feature extraction.

For each predefined ROI, seven quantitative imaging features were extracted, comprising six first-order statistical features—mean, standard deviation, median, interquartile range, 10th percentile, and 90th percentile—and voxel count. Although voxel count was included as a quantitative ROI descriptor in the subsequent feature-selection pipeline, it was not classified as a first-order statistical feature. No conventional texture or higher-order radiomic features were extracted. According to the Image Biomarker Standardisation Initiative (IBSI),30 first-order statistical features constitute an established category within the standardized radiomics feature framework. By characterizing within-node signal distributions and heterogeneity beyond conventional ROI-averaged measures, this approach provides complementary quantitative information for subsequent multimodal radiomics and machine-learning analyses.

Dimensionality Reduction and Radiomic Feature Selection

The imaging dataset was randomly partitioned into training and testing sets using stratified sampling at a 7:3 ratio. The training set was used for feature selection, model construction, and hyperparameter optimization, whereas the test set was used exclusively for model performance evaluation. During data preprocessing, features with more than 30% missing values were first removed from the training set. Missing values were then imputed using the median values of the corresponding features in the training set, and the same imputation parameters were applied to the test set. Variance-based feature selection was subsequently performed to retain features with sufficient variability and remove non-informative or redundant features. To reduce multicollinearity, pairwise Pearson correlation coefficients were calculated among candidate features in the training set, and highly correlated redundant features were removed using a threshold of |r| > 0.85. The retained features were then standardized using Z-score normalization so that each feature had a mean of 0 and a standard deviation of 1. Standardization parameters were fitted only in the training set and subsequently applied to the test set. The Mann–Whitney U-test was then performed in the training set to assess distributional differences between the two classes. Features were ranked in ascending order according to their P values, and the top K = 25 features were retained for embedded feature selection. This step was used solely as a univariate prescreening procedure within the sequential dimensionality-reduction pipeline and was not intended for independent statistical inference on individual features. An ElasticNet-regularized logistic regression model was subsequently used for further feature selection, with hyperparameters optimized using five-fold stratified cross-validation within the training set. For the regularization parameter C, 35 logarithmically spaced candidate values ranging from 10−3 to 101·3 were evaluated, while the L1 mixing parameter (l1_ratio) was set to 0.4, 0.5, 0.6, 0.7, or 0.8. The optimal hyperparameters were selected based on the cross-validated area under the receiver operating characteristic curve (ROC-AUC). Features with nonzero coefficients were retained, and the 10 features with the largest absolute coefficient values were subsequently selected for final model construction. All data preprocessing, feature selection, and hyperparameter optimization procedures were performed exclusively within the training set, and the held-out test set was not involved in any of these steps.

A multivariable logistic regression model was established using the final selected radiomic features, and a radiomics score (RadScore) was calculated accordingly. The RadScore was defined as the linear predictor of the model, calculated as the weighted sum of the standardized values of the selected features multiplied by their corresponding regression coefficients, plus the intercept (Equation 1). The RadScore represented an individual-level model prediction score derived from the radiomic features and was used for subsequent between-group comparison and model evaluation.

(1)

Where β0 represents the intercept, βi represents the regression coefficient of the i-th selected feature, and Xi represents the standardized value of the corresponding feature.

To evaluate the incremental discriminative value of multimodal imaging information beyond conventional clinical variables, a clinical-only baseline model was additionally constructed. Six clinical variables were prespecified: age, years of education, duration of HIV infection, current CD4 count, nadir CD4 count, and duration of antiretroviral therapy (ART). All six variables were entered directly into the logistic regression model without additional feature selection. The clinical model used the same 7:3 stratified random split into training and test sets as the imaging model. Missing values were imputed using the median values derived from the training set, and continuous variables were standardized using the means and standard deviations estimated from the training set. The preprocessing parameters derived from the training set were then applied to the held-out test set. The held-out test set was used exclusively for final model performance evaluation. Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC) with its 95% confidence interval (CI). The DeLong test was used to compare the AUCs of the clinical baseline model and the imaging–clinical fusion model in the same held-out test set.

In addition, given the statistically significant difference in ART duration between the HIV-ANI and HIV-IC groups, a sensitivity analysis was performed to assess the potential influence of this variable on model discrimination. While keeping the feature set, training–test split, and all other modeling procedures of the original final model unchanged, ART duration was included as an additional covariate irrespective of feature selection, without repeating the feature-selection procedure. ART duration was processed using the same preprocessing strategy as in the primary analysis, with missing-value imputation and Z-score standardization performed on the basis of the training set and the resulting parameters subsequently applied to the held-out test set. The AUCs of the models before and after adjustment for ART duration were then calculated in the same held-out test set and compared using the DeLong test to assess the stability of model discrimination after adjustment for ART duration.

Model Construction and Evaluation

The study constructed five categories of predictive models: a clinical model based on clinical variables, an fMRI model based on ALFF- and ReHo-derived functional radiomic features, a T1WI model based on structural radiomic features from T1-weighted imaging, a multimodal MRI radiomics model integrating fMRI and T1WI features, and an imaging-clinical fusion model further incorporating radiomic features and clinical variables. The five models were refitted using the training set and subsequently evaluated in the held-out test set using AUC, sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). The workflow of the multimodal MRI radiomics framework is shown in Figure 1. In addition, to further assess the internal stability of the final imaging–clinical fusion model and the optimism in its performance estimates, internal validation was performed using 1000 bootstrap resamples of the training set. For each bootstrap resample, the prespecified final model was refitted, and its AUC was calculated in both the bootstrap sample and the original training set; the difference between these AUCs was used to estimate optimism. The mean optimism was then subtracted from the apparent AUC to obtain the optimism-corrected AUC.

Figure 1 Node-based multimodal MRI radiomics framework for identifying asymptomatic neurocognitive impairment in HIV-positive individuals. (A) MRI data acquisition, including resting-state fMRI–derived ALFF/ReHo maps and 3D T1-weighted structural MRI. (B) Definition of core cognitive network regions of interest (ROIs), including the salience network (SN; insula/anterior cingulate cortex), default mode network (DMN; posterior cingulate cortex/precuneus/medial prefrontal cortex), and central executive network (CEN; dorsolateral prefrontal cortex/posterior parietal cortex). (C) ROI-based radiomics feature extraction from ALFF, ReHo, and 3D T1WI images using voxel intensity distribution characteristics within each ROI. (D) Feature selection performed in the training set only, including stratified train-test split, preprocessing, feature filtering, and ElasticNet with five-fold cross-validation for optimal feature selection. (E) Model construction and evaluation using multimodal and imaging–clinical fusion models, followed by SHAP-based model interpretation and test-set performance assessment.

Abbreviations: MRI, magnetic resonance imaging; fMRI, functional magnetic resonance imaging; ALFF, amplitude of low-frequency fluctuation; ReHo, regional homogeneity; 3D T1WI, three-dimensional T1-weighted imaging; ROI, region of interest; SN, salience network; DMN, default mode network; CEN, central executive network; SHAP, SHapley Additive exPlanations.

Model Interpretability and Visualization

To improve model interpretability and visually demonstrate the contribution of each selected feature to the prediction results, SHapley Additive exPlanations (SHAP) were further applied to interpret the final model. The sign of the SHAP value indicates whether a given feature shifts the model prediction toward the ANI group or the HIV-IC group, whereas its absolute value reflects the magnitude of the feature contribution. SHAP summary plots were generated to visualize the overall importance of each feature and the relationship between feature values and prediction direction. The mean absolute SHAP value was used to quantify the global contribution of each feature to model output, with larger values indicating a stronger influence on model prediction.

A nomogram was constructed based on the final multivariable logistic regression model. Regression coefficients of the features included in the imaging–clinical fusion model were converted into corresponding point scores, and the total score was mapped to the predicted probability of the target outcome, enabling intuitive quantification of individualized risk. The discrimination, calibration, and clinical net benefit of the nomogram model were evaluated using calibration curves and decision curve analysis.

Statistical Analysis

Statistical analyses and model construction were performed using IBM SPSS Statistics version 25.0, R version 3.6.1, and Python version 3.10.11. Normally distributed variables were reported as the mean ± standard deviation(x ± s), whereas non-normally distributed variables were expressed as the median and interquartile range [M(IQR)]. The chi-square test was used for categorical variables. Independent-samples t tests were used for normally distributed continuous variables, whereas the Kruskal–Wallis test and Mann–Whitney U-test were used for non-normally distributed variables. A two-sided P value <0.05 was considered statistically significant. For radiomics analysis, the sample was divided into training and test sets at a ratio of 7:3 using stratified random sampling. After missing-value processing, low-variance feature removal, Pearson correlation filtering, Z-score normalization, and univariate screening, radiomic and clinical features were subjected to embedded feature selection and model construction using ElasticNet-regularized logistic regression. Elastic Net-regularized logistic regression was selected because of the relatively high dimensionality and potential collinearity of the candidate features in relation to the sample size. By combining L1 and L2 regularization, Elastic Net enables feature selection while mitigating multicollinearity and controlling model complexity. Model parameters were optimized using five-fold stratified cross-validation. Model performance was evaluated using ROC curves, AUCs, and corresponding 95% confidence intervals. Classification metrics, including sensitivity, specificity, accuracy, PPV, NPV, were also calculated. SHAP analysis was performed to interpret the imaging–clinical fusion model, and a nomogram was constructed to visually present individualized predicted probabilities. Calibration curves and decision curve analysis were further used to assess model calibration and clinical net benefit, respectively.

Results Clinical Characteristics

There were no significant differences between the two groups in age (P = 0.711) or years of education (P = 0.094). No significant differences were observed between the HIV-IC and HIV-ANI groups in CD4 count (P = 0.172), CD4 nadir (P = 0.673), duration of HIV infection (P = 0.084), or the proportion of participants with unsuppressed viral load (P = 1.000). The duration of antiretroviral therapy was shorter in the HIV-ANI group than in cognitively intact people living with HIV. Among participants receiving stable cART, one individual in the HIV-IC group had a viral load of 44 copies/mL, whereas one individual in the HIV-ANI group had a viral load of 94 copies/mL. The clinical characteristics of participants in the HIV-IC and HIV-ANI groups and the between-group comparisons are presented in Table 1.

Table 1 Clinical Characteristics and Neurocognitive Assessment Results of the HIV-IC and HIV-ANI Groups

Diagnostic Performance of Radiomic Features

A total of 98 quantitative imaging features were extracted from each modality (T1WI, ALFF, and ReHo), yielding 294 imaging features. Together with six prespecified clinical variables, 300 candidate variables were initially included in the imaging–clinical analysis. After preprocessing and feature selection, three imaging-based models were constructed: an fMRI model based on ALFF- and ReHo-derived features, a T1WI model based on structural imaging features, and a multimodal fMRI+T1WI model integrating structural and functional imaging features. In addition, a clinical-only model was constructed using the six prespecified clinical variables, and an imaging–clinical fusion model was developed by integrating the selected imaging features with clinical variables. Rad-scores were calculated for the imaging-based and imaging–clinical fusion model, and model performance was evaluated using AUC, specificity, sensitivity, accuracy, PPV, and NPV (Table 2).

Table 2 Comparative Analysis of the Diagnostic Performance of the Four Prediction Models

In the training set, the fMRI model achieved the highest AUC value of 0.843 (95% CI: 0.763–0.914), with sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of 0.761, 0.857, 0.811, 0.833, and 0.792. The second-best model was the image-clinical fusion model, with an AUC of 0.827 (95% CI: 0.737–0.902), and sensitivity, specificity, and accuracy of 0.739, 0.816, and 0.779. The training set AUCs for the fMRI+T1WI multimodal model, the T1WI model, and the clinical model were 0.794 (95% CI: 0.698–0.883), 0.783 (95% CI: 0.687–0.831), 0.700 (0.594, 0.802). In the test set, the image-clinical fusion model demonstrated the best diagnostic performance, with an AUC of 0.735 (95% CI: 0.574–0.880), and sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of 0.619, 0.810, 0.714, 0.765, and 0.680, respectively. The fMRI+T1WI model ranked second, with an AUC of 0.705 (95% CI: 0.527–0.855), a sensitivity of 0.429, a specificity of 0.857, and a precision of 0.643. Among the single-modality models, the fMRI model had an AUC of 0.651 (95% CI: 0.468–0.815) in the test group, while the T1WI model had an AUC of 0.519 (95% CI: 0.336–0.702), and the clinical model had an AUC of 0.599 (95% CI: 0.416–0.779); the T1WI model exhibited higher specificity (0.857) but lower sensitivity (0.286). The ROC curve analysis results for the five predictive models in the training and test sets are shown in Figure 2.

Figure 2 ROC curves of the five prediction models in the training and test sets. (A) ROC curves in the training set. (B) ROC curves in the test set. ROC curves were used to evaluate the classification performance of the clinical model, fMRI model, T1WI model, fMRI+T1WI model, and imaging-clinical fusion model. The dashed line represents the reference line indicating no discriminative ability.

Abbreviations: ROC, receiver operating characteristic; fMRI, functional magnetic resonance imaging; T1WI, T1-weighted imaging; AUC, area under the receiver operating characteristic curve.

The imaging–clinical fusion model achieved a higher AUC than the clinical only model (0.735 vs. 0.599; ΔAUC = 0.136). However, the DeLong test indicated that this difference did not reach statistical significance (Z = 1.354, P = 0.176).

In the sensitivity analysis, after ART duration was included as an additional covariate in the final model, the test-set AUC decreased slightly from 0.735 to 0.721 (ΔAUC = −0.014). The DeLong test showed no statistically significant difference in AUC between the models before and after adjustment (Z = −0.313, P = 0.754), indicating that model discrimination did not change significantly after adjustment for ART duration.

In the internal validation of the final model using 1,000 bootstrap resamples, the apparent AUC was 0.825, with a mean optimism of 0.081, resulting in an optimism-corrected AUC of 0.744, which was close to the AUC of 0.735 observed in the held-out test set. These results provided additional information for assessing the internal stability of model performance and the potential optimism in its performance estimates.

Optimal Features of the Imaging-Clinical Model

Among the 300 candidate variables described above, all were retained after missing-value screening. Low-variance filtering reduced the number of variables to 258, and subsequent Pearson correlation filtering further reduced the candidate set to 93 variables. Subsequently, univariate screening was performed on the training set using the Mann–Whitney U-test, and the top 25 candidate variables ranked by P-value were retained for inclusion in the ElasticNet-regularized logistic regression model. The ElasticNet model identified 18 variables with non-zero coefficients, and the 10 variables with the largest absolute coefficients were ultimately selected for the final model construction based on the absolute values of the regression coefficients. The final selected features included ReHo_Frontal_Sup_Medial_R_p10, ReHo_Cingulum_Ant_R_std, T1WI_Cingulum_Post_R_mean, ReHo_Cingulum_Post_L_p10, T1WI_Insula_L_std, Education, Duration of infection, ALFF_Parietal_Inf_L_p10, T1WI_Parietal_Inf_R_p10, ReHo_Insula_L_std. These features comprised four ReHo-derived features, three T1WI-derived features, one ALFF-derived feature, and two clinical features.

Interpretation of the Imaging-Clinical Model

SHAP analysis was performed to interpret the imaging-clinical model (Figure 3). Anatomical and methodological interpretation of the features included in the final imaging-clinical fusion model is provided in Supplementary Table S1. Because SHAP values reflect the contribution of each feature to the raw model output, a positive SHAP value indicates that the feature shifts the model output toward the ANI group (label = 1), whereas a negative SHAP value indicates that the feature shifts the output toward the HIV-IC group (label = 0). The SHAP beeswarm plot illustrated the relationship between feature values and the direction of model output. High values of ReHo_Frontal_Sup_Medial_R_p10 and ReHo_Cingulum_Ant_R_std were mainly distributed in the positive SHAP value region, indicating that increases in these features tended to increase the probability of model prediction toward the ANI group. By contrast, high values of T1WI_Cingulum_Post_R_mean and ReHo_Cingulum_Post_L_p10 were mainly distributed in the negative SHAP value region, suggesting that increases in these features shifted model prediction toward the HIV-IC group. High values of T1WI_Insula_L_std mainly showed a positive contribution, suggesting that increased dispersion of the T1WI signal in the left insula may be associated with prediction of the ANI group. Global SHAP importance analysis showed that ReHo_Frontal_Sup_Medial_R_p10 had the greatest contribution to the model, followed by ReHo_Cingulum_Ant_R_std, T1WI_Cingulum_Post_R_mean, ReHo_Cingulum_Post_L_p10, and T1WI_Insula_L_std. The top five features were all radiomic features, mainly involving the medial superior frontal gyrus, cingulate cortex, and insula. This finding suggests that the discriminative information of the imaging–clinical fusion model was primarily derived from ReHo-based functional radiomic features and T1WI-based structural radiomic features. The remaining selected features included educational level, duration of HIV infection, ALFF_Parietal_Inf_L_p10, T1WI_Parietal_Inf_R_p10, and ReHo_Insula_L_std. Although their mean absolute SHAP values were relatively low, they still contributed to model output, indicating that imaging features related to the inferior parietal lobule and insula, together with clinical background information, provided complementary discriminative information.

Figure 3 SHAP plots of the imaging-clinical fusion model. (A) SHAP beeswarm plot showing the direction and magnitude of the contribution of each feature to the output of the final fusion model. The x-axis represents SHAP values; SHAP values greater than 0 indicate an increase in model output, whereas SHAP values less than 0 indicate a decrease in model output. Colors indicate relative feature values according to the cividis color scale, ranging from low to high as shown in the color bar. (B) SHAP global feature importance plot showing the mean absolute SHAP value of each feature. Larger values indicate greater overall contribution to model prediction.

Abbreviation: SHAP, SHapley Additive exPlanations.

Nomogram of the Imaging–Clinical Fusion Model

A nomogram was constructed based on the 10 final selected features in the imaging–clinical fusion model (Figure 4). The nomogram maps the value of each predictor to a corresponding point score and calculates the predicted probability of ANI based on the total score, thereby providing a visual representation of the imaging–clinical fusion model. The nomogram model was evaluated using calibration curves and decision curve analysis (Figure 5). Calibration curve analysis showed that both the apparent and bootstrap-corrected calibration curves in the training set were generally close to the ideal reference line, indicating good agreement between predicted probabilities and observed outcomes. In the test set, the calibration curve also showed a generally consistent trend between predicted probabilities and observed outcomes. The AUC was 0.735 and the Brier score was 0.217, indicating moderate discriminative ability and probability calibration performance. Decision curve analysis further showed that, within a certain range of threshold probabilities, the imaging–clinical fusion model provided a higher net benefit than both the “treat-all” and “treat-none” strategies, indicating its potential clinical utility for individualized ANI risk prediction.

Figure 4 Nomogram of the imaging-clinical fusion model. The nomogram was constructed using the final selected features included in the imaging–clinical fusion model and was used to estimate the individualized predicted probability of ANI.

Abbreviation: HIV-ANI, HIV-associated asymptomatic neurocognitive impairment.

Figure 5 Calibration curves and decision curve analysis of the imaging-clinical fusion model. (A and B) Calibration curves in the training (A) and test (B) sets, respectively, used to assess the agreement between predicted probabilities and observed outcomes. (C and D) Decision curve analysis in the training (C) and test (D) sets, respectively, used to evaluate the clinical net benefit of the model across different threshold probabilities. The red curve represents the fusion model. “All” denotes the treat-all strategy, in which all participants are assumed to have ANI, whereas “None” denotes the treat-none strategy, in which no participants are assumed to have ANI.

Abbreviation: ANI, asymptomatic neurocognitive impairment.

Discussion

In this study, we constructed a multimodal model integrating structural radiomic features derived from 3D T1-weighted imaging and functional radiomic features derived from ALFF and ReHo maps, focusing on core hub nodes of the DMN, CEN, and SN, to identify HIV-ANI. The results showed that the imaging–clinical fusion model achieved the best performance in the test set, with an AUC of 0.735, achieving the highest test-set AUC among the evaluated models. SHAP analysis revealed that core hub nodes, including the medial prefrontal cortex, anterior and posterior cingulate cortices, and insula, made the greatest contributions to model prediction. Reduced function in the posterior DMN was associated with model-based prediction of ANI, whereas increased prefrontal functional activity may reflect the contribution of local functional alterations to prediction. Clinical factors, including years of education and duration of HIV infection, also contributed to model performance.

Previous studies have shown that HIV infection can induce large-scale brain network dysfunction, with abnormal functional connectivity and altered network topology in core DMN regions, such as the posterior cingulate cortex, precuneus, and medial prefrontal cortex. These alterations are associated with impairments in attention, memory, and executive function.9,31 Interactions between the SN and DMN are disrupted in patients with HAND, particularly in regions involving the anterior cingulate cortex and prefrontal cortex, suggesting impaired inter-network coordination.10 In parallel, reduced functional connectivity in CEN-related regions has been associated with deficits in working memory and executive function.32 Functional connectomics studies further suggest that abnormalities in core hub nodes, such as the posterior cingulate cortex, may play an important role in network dysfunction, with altered nodal efficiency being associated with declines in executive function and abstraction ability.33,34 Impaired coordination across multiple networks is closely associated with overall cognitive performance.9,32 Therefore, these findings suggest that network abnormalities in HAND may arise not only from disrupted connectivity but also from local structural and functional abnormalities within hub nodes, providing a complementary perspective on the potential relationship between node-level alterations and broader network dysfunction.

The mechanism of node-driven network dysfunction suggests that local structural alterations may constrain or reshape regional functional dynamics and contribute to the progression from local abnormalities to whole-brain network reorganization.35 Multimodal radiomic features can provide information on within-node signal distributions and heterogeneity beyond conventional ROI-averaged measures, thereby offering complementary insights into the potential relationship between local node abnormalities and broader network dysfunction. SHAP analysis showed that core hub nodes of the DMN, SN, and CEN, including the posterior cingulate cortex, anterior cingulate cortex, medial prefrontal cortex, and insula, made the largest contributions to ANI prediction, as indicated by the mean absolute SHAP values. Specifically, reduced function in posterior DMN nodes was associated with model-based prediction of ANI, whereas increased prefrontal functional activity also contributed positively to model prediction, suggesting functional heterogeneity across brain regions in the early disease stage. This pattern is partly consistent with previous findings showing posterior DMN dysfunction and abnormal structure–function coupling in early HAND, suggesting that functional impairment and compensatory, or inefficient, activation may coexist during the early stage of the disease.12,15 In addition, structural radiomic features, such as 3D T1WI-derived features related to the insula and parietal regions, also made substantial contributions, suggesting that structural signal heterogeneity and functional abnormalities may coexist and potentially reflect alterations in structure–function relationships.36–38

Compared with previous HAND studies that have primarily relied on ROI-averaged measures, functional connectivity, or connectomic features, the present study further extracted structural and functional radiomic features from within core network nodes to characterize within-node signal distributions and heterogeneity. Previous studies have combined clinical information with multimodal MRI features for machine-learning-based prediction of HIV-associated neurocognitive impairment,22 More recent studies have further applied DTI-based radiomics and rs-fMRI radiomic features to the early identification of HAND or ANI,39,40 highlighting the potential value of multimodal information integration for the early identification of HAND. Similar strategies have also been applied to neurodegenerative diseases such as Alzheimer’s disease and Parkinson’s disease, where multiparametric or multimodal radiomics models integrate complementary information from different imaging sources to improve quantitative characterization and classification of disease phenotypes.41–43 Compared with these studies, the incremental contribution of the present work lies primarily in two aspects. First, we integrated information from 3D T1WI, ALFF, and ReHo to characterize early HIV-ANI-related abnormalities from both structural and local functional perspectives. Second, we further focused the analysis on within-node characteristics of core nodes in the DMN, CEN, and SN and used SHAP to interpret the predictive contributions of key node-level features, thereby providing complementary information on within-node heterogeneity that is not fully captured by conventional connectomic analyses. Thus, rather than replacing existing connectomic frameworks, the present approach provides a complementary characterization of HAND from the perspective of within-node features.

The imaging–clinical fusion model demonstrated some discriminative value in the present study, highlighting the importance of integrating structural, functional, and clinical information. Functional indices, such as ALFF and ReHo, primarily reflect local neural activity,28,29 whereas 3D T1WI captures macroscopic morphological alterations in brain structure.44 Their integration enables a more comprehensive characterization of the multidimensional pathological features of the disease from both structural and functional perspectives. Clinical variables, including years of education and duration of HIV infection, also contributed to the model, suggesting potential interactions among imaging features, individual cognitive reserve, and disease exposure that may jointly influence the risk of ANI. Previous studies have shown that years of education is an independent factor associated with HAND, particularly ANI. Lower educational attainment is associated with a higher risk of cognitive impairment, potentially through mechanisms related to cognitive reserve and resilience to neural injury.45–47 The duration of HIV infection is also associated with the occurrence and progression of HAND. However, in individuals receiving cART, its independent predictive effect may be modulated by nadir CD4 count, age, and treatment-related factors. This suggests that the duration of HIV infection may reflect cumulative neurotoxic exposure, the cognitive impact of which depends on immune status and treatment background.7,48–50 In addition, sensitivity analysis showed that model performance remained largely unchanged after adjustment for ART duration, suggesting that the between-group difference in ART duration may have had a limited impact on model discrimination.

Taken together, the integration of imaging and clinical information may provide a complementary perspective for identifying individuals at risk of HIV-ANI. From a clinical perspective, however, the current model is not intended to replace standardized neuropsychological assessment. With further validation, multimodal MRI-derived features may serve as an objective adjunct for identifying individuals at increased risk of HIV-ANI and for selecting those who may benefit from more comprehensive neurocognitive evaluation and longitudinal follow-up. Thus, this approach may ultimately complement, rather than replace, existing clinical assessment strategies for HAND.

This study has several limitations. First, the sample size was limited and the cohort was predominantly male. No formal sample-size or statistical power analysis was performed, and the relatively small held-out test set may have increased uncertainty in the model performance estimates. Although an AUC of 0.735 in the test set indicated some discriminative ability, the decline in performance from the training to the test set observed in some models suggests that overfitting cannot be entirely excluded. Second, the cross-sectional design limits inferences regarding longitudinal changes and causal relationships. Third, radiomic features provide only indirect representations of the underlying biological alterations, and the predefined ROI-based strategy may have overlooked other brain regions vulnerable to HIV-related injury, such as the basal ganglia and cerebellum. Future exploratory whole-brain analyses may enable a more comprehensive characterization of regional brain abnormalities in early HAND. Finally, the lack of an external validation cohort limits the generalizability of the present findings. Accordingly, the current model should be regarded as an exploratory imaging framework, and its robustness and generalizability require further validation in larger, independent, multicenter cohorts.

Conclusion

In summary, multimodal MRI analysis revealed structural and functional signal heterogeneity within core cognitive network nodes in individuals with HIV-ANI. The integration of imaging-derived features with clinical factors showed potential for distinguishing HIV-ANI from cognitively intact individuals with HIV. These findings provide preliminary evidence that node-level abnormalities may be associated with broader network dysfunction in early HAND. Further validation in larger, independent, multicenter cohorts is warranted to establish the robustness and generalizability of these findings.

Abbreviations

cART, Combination antiretroviral therapy; HIV, Human immunodeficiency virus; CNS, Central nervous system; HAND, HIV-associated neurocognitive disorders; ANI, Asymptomatic neurocognitive impairment; MND, Mild neurocognitive disorder; HAD, HIV-associated dementia; NPT, Neuropsychological testing; ROI, Regions of interest; AUC, Area under the curve; RadScore, Radiomics score; PPV, Positive predictive value; NPV, Negative predictive value.

Data Sharing Statement

The datasets generated and analyzed during the current study are not publicly available because they contain clinical and neuroimaging data from human participants and are subject to patient privacy protection, ethical approval requirements, and institutional data management regulations. De-identified data may be made available from the corresponding author upon reasonable request and with approval from the relevant ethics committees and institutional data governance bodies.

Ethics Approval and Informed Consent

This study was approved by the Ethics Committees of Beijing Youan Hospital, Capital Medical University, and the Affiliated Infectious Diseases Hospital of Zhengzhou University (approval numbers: LL-2020-047-K and IEC-