Early-stage diagnosis of HIV-associated neurocognitive disorders via multiple learning models based on resting-state functional magnetic resonance imaging

Introduction

The widespread use of combination antiretroviral therapy (cART) has brought human immunodeficiency virus (HIV) under control in much of the world (1). However, viruses in the early stages of infection are capable of crossing the blood-brain barrier (BBB) into the central nervous system (CNS) to induce inflammation, infect microglia and astrocytes, and form viral reservoirs (2,3). In addition, inflammation in the CNS system provides a suitable environment for HIV replication, further facilitating cerebrospinal fluid viral escape and prolonging exposure of the brain to circulating virus. This ultimately leads to apoptosis and axonal damage, which can progress to structural brain damage and cognitive limitations in people living with HIV (PLWH) (4-6). Overall, cART prolongs life, but PLWH face a variable risk of cognitive impairment.

Based on current diagnostic criteria, HIV-associated cognitive impairment is defined as HIV-associated neurocognitive disorders (HAND). It is classified according to the severity of the disease as asymptomatic neurocognitive impairment (ANI), mild neurocognitive disorder (MND), and HIV-associated dementia (HAD) (7). Given that ANI is the most prevalent category of HAND, the clinical relevance of is identification is obvious (8). Research from Grant et al. indicated that individuals with ANI are approximately three times more likely to experience daily living problems than are PLWH who were initially cognitively normal (9). There is a high degree of clinical similarity between PLWH but no neurocognitive disorders (PWND) and ANI, which makes the diagnosis of ANI particularly challenging. Therefore, the mining of ANI-related features to provide more valuable early diagnostic information for patients with HAND is urgently needed.

HAND diagnosis is primarily implemented via neurocognitive testing, but the universal application of treatment strategies after HIV diagnosis has led many researchers and clinical practitioners to question the 2007 Frascati criteria for overestimating the prevalence and severity of HAND, including ANI (10). A more strict criterion involving increasing the original one standard deviation to 1.5 has been proposed (11). This is considered to be more consistent with the disease’s pattern and onset tendency, and was employed as the criterion in this study (11). In general, diagnostic models constructed from neurocognitive testing have inherent drawbacks, such as being time-consuming, susceptible to environmental interference, and limited by biases related to subjective factors and the level of physician training. However, beyond behavioral symptoms investigated through psychiatric methods, there is no gold standard biomarker for the staging and diagnosis of HAND.

Recently, an increasing number of studies have been conducted on the possible neurobiological mechanisms in patients with HAND via noninvasive imaging methods, particularly resting-state functional magnetic resonance imaging (rs-fMRI) (12,13). This advanced method maps brain function by examining brain signals in a calm state and has indicated that there are similarities between brain regions in resting and task states (14). Interestingly, rs-fMRI can even capture trends in brain activity that are difficult to observe with task fMRI (15). Numerous studies have reported distinct values of rs-fMRI metrics including amplitude of low-frequency fluctuation (ALFF), fractional ALFF (fALFF), and regional homogeneity (ReHo) in those with HAND, with study populations including but not limited to children, adults, healthy controls, and PLWH (16,17). The correlation of abnormal signal values with cognitive test performance has been established, providing a clinical marker to assess functional brain activity and an assessment method for planning neurocognitive interventions in PLWH receiving standard treatment.

Research methods combining machine learning and rs-fMRI have shown great potential in HAND diagnosis. Several studies have used this approach to develop different HAND diagnostic models, and their efficacy in differentiating patients with ANI and those with PWND has been demonstrated (18-20). This approach provides a novel means to identifying neuroimaging biomarkers.

Artificial intelligence was first discussed positively by Turing in 1950, who asked the question “Can machines think?” and devised an experiment known as the “Turing Test”. Several machines have come close to passing the Turing Test (21). John McCarthy first introduced the concept of artificial intelligence at the Dartmouth Conference in 1956, defining it as “the science and engineering of making machines behave as intelligently as human beings” (22). Machine learning is a subfield of artificial intelligence that utilizes known datasets for training, from which machines “learn”, and develops algorithms that can be applied to unknown datasets to perform a variety of tasks, such as diagnostics after training, commonly including supervised and unsupervised learning (23,24). Supervised machine learning relies on labeled datasets, and models learn based on patterns in the data, mapping input features to target outputs that enable the prediction of predefined outcomes. After training, the model is tested using unknown datasets and its predictive performance is evaluated. Logistic regression, support vector machines, decision trees, and k-nearest neighbor algorithms are common supervised learning algorithms. Unlike supervised learning, unsupervised learning does not require labeled data (25,26). For example, k-means clustering, principal component analysis, and other methods critically support data processing, visualization, clustering, dimensionality reduction, and other tasks. In recent years, machine learning has been widely used in the medical field, especially in supervised machine learning. To ensure the validity, reliability and generalizability of the models, these algorithms need to be trained and tested with the help of appropriate datasets and methods so that they can be applied to a wider range of patient groups. In medical image analysis, feature extraction of the image or regions of interest (ROIs) is a critical step, which can be done manually, such as in the case of manual feature extraction in an imaging histology workflow (27). Radiomics is another noninvasive imaging technique that extracts a multitude of features from image data that are inaccessible to the naked eye, converting the images into high-dimensional data that can be used to build diagnostic or predictive models through a variety of characterization algorithms and serve noninvasive predictive biomarkers (28). An increasing number of studies have applied this method to cognitive disorders (29,30). Studies have found the validity of radiomics in the diagnosis of HAND. Qi et al. identified additional features associated with early cognitive impairment in PLWH by combining radiomics and diffusion tensor imaging to improve the accuracy of HAND identification (31).

rs-fMRI focuses on responding to brain activation, and radiomics selects the most representative and meaningful features to build a classification model. Recently, it has been found that the acquisition of radiomics features in rs-fMRI can capture changes in brain activity. Liu et al. proposed a framework for the quantitative diagnosis of attention deficit/hyperactivity disorder (ADHD) based on radiographic features extracted from preprocessed rs-fMRI images, which achieved good performance in classifying individuals with ADHD (32). Wang et al. found that model combining ALFF and radiomics features had high accuracy in diagnosing patients with Alzheimer disease and amnestic mild cognitive impairment (33). Thus, the addition of radiomics enables a greater amount image information to be obtained in separate brain regions and improves diagnostic performance.

We performed a study, which to our knowledge, is the first to integrate radiomics with rs-fMRI for the diagnosis of HAND. We hypothesized that rs-fMRI–based radiomics features have better discriminatory ability than do rs-fMRI signal values and have the potential to act as neuroimaging biomarkers for the diagnosis of early-stage HAND. The specific aims of this study were to construct rs-fMRI-based machine learning models for the early identification of ANI in HAND and to evaluate the performance of the models in classification of ANI and PWND, advancing the development of accurate and robust predictive models. We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-290/rc).

Methods

Patients enrollment

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by the Ethics Committee of Beijing Youan Hospital (No. LL-2023-070-K). Informed consent was obtained from all individual participants. We prospectively analyzed 381 PLWH recruited from October 2020 to October 2024 from Beijing Youan Hospital’s STD/AIDS clinic. Right-handed adult Chinese males diagnosed by an enzyme-linked immunosorbent assay and Western blot analysis were included. Other enrollment criteria included a plasma HIV RNA level of <20 copies/mL, cART stabilization duration of more than 6 months, and the absence of treatment interruptions. The exclusion criteria included the following: (I) tumors, head injury or infection, epilepsy, and other disorders that may cause neurocognitive disorders; (II) significant depressive states, anxiety disorders, schizophrenia, and other psychiatric disorders or the application of antianxiety, antidepressant, and other medications to control mental disorders; (III) habits of drug abuse, alcoholism, or substance abuse; (IV) contraindications to MRI; and (V) a severity of severe vision, hearing, or reading difficulties precluding cooperation with cognitive testing.

Prior to the neuropsychological (NP) testing, the singular Hamilton Depression scale (17-item version) was applied to assess depressive status (34). Each item has multiple options corresponding to different levels of severity of depressive symptoms, ranging from 0 (none) to 2 or 4 (most severe). A higher total score on all items indicates a more severe depressive symptom of the patient. Participants with a score of less than 7 were considered to have no depressive symptoms, those with a score of greater than or equal to 7 and less than 17 were considered to have probable depression, those with a score of greater than or equal to 17 and less than 24 were considered to have definite depression, and those with a score of greater than or equal to 24 were considered to have severe depression. Only patients without depressive symptoms who scored less than 7 were included in this study. In addition, the Activity of Daily Living scale, which consists of two parts—somatic self-care and instrumental activities of daily living—was used to assess whether the ability to perform daily living was impaired (35). In order to ensure the reliability of the assessment results as much as possible, the depression screening, daily living assessment, and NP tests were conducted in a quiet, undisturbed environment. They were conducted by specialized designated physicians, and the test data of each participants were reviewed by two independent assessors to ensure the objectivity and consistency of the data. The NP test (7) covered six cognitive domains: verbal and language, as assessed through animal verbal fluency and action verbal fluency; memory (learning and recall), as assessed via the Hopkins Verbal Learning Test-Revised (36) and the Brief Visuospatial Memory Test-Revised (37); attention/working memory, as assessed via the Wechsler Memory Scale-III (38) and the Paced Auditory Serial Addition Test (39); abstraction and executive function testing, as assessed by Wisconsin Card Sorting Test 64 (40); and fine-motor ability and information processing speed, as assessed by the Grooved Pegboard Test (8) and Trail-Making Test Part A (41), respectively. A Chinese version of the cognitive normative data was obtained by collecting the origin city, residence city, education, age, gender, and cognitive test scores of Chinese adults aged 18–60 years. Average T scores were calculated based on cognitive normality data (42), and when the cognitive item average T score was lower than the normal reference value by 15 points, this was considered a 1.5 standard deviation (SD).

PLWH were classified as PWND or ANI on the basis of cognitive results. With the increasing prevalence of cART, the Frascati criteria for HAND diagnosis have been subjected to greater scrutiny, especially concerning the threshold scores for mild cognitive impairment, which may result in false-positive diagnoses (43). According to Gisslén et al.’s findings, the ANI diagnostic criteria, which consider performance in at least two cognitive domains to be one SD below the normative mean, can lead to overdiagnosis and undue anxiety (11). With a 1.5-SD cutoff , the false-positive rate can be significantly reduced (by approximately 4–8%), thereby enhancing the accuracy of detecting true cognitive abnormalities (11). Consequently, in our study, ANI was defined as performance in at least two cognitive domains falling 1.5 SD below the standardized mean without any deficits in activities of daily living. The cognitive performance of PWND was expected to be better than that of those with ANI. A flowchart of the patient screening process is shown in Figure 1. A total of 152 PLWH cases were included in this study, including 74 individuals with ANI and 78 PWND, and were randomized into two cohorts in a 7:3 ratio (training set: n=106 patients; validation set n=46).

Figure 1 Flowchart of patient selection. ANI, asymptomatic neurocognitive impairment; MR, magnetic resonance; PWND, people living with human immunodeficiency virus but no neurocognitive disorders; rs-fMRI, resting-state functional magnetic resonance imaging; T1WI, T1-weighted imaging.

Imaging data collection

All imaging data were acquired on a 3.0-T MR scanner (Tim Trio, Siemens Healthineers, Erlangen, Germany) equipped with a 32-channel magnetic head coil. Participants wore earplugs and maintained a stable body position and were instructed to keep their eyes e closed but not to sleep. Routine MR images including T2-weighted imaging and T2 fluid-attenuated inversion recovery imaging were acquired for radiological diagnosis. rs-fMRI data were acquired with a gradient single shot echo planar imaging sequence [repetition time (TR) =2,000 ms, echo time (TE) =30 ms, field of view (FOV) =224×224 mm², in-plane matrix =64×64, flip angle =90, slice number =35, slice thickness =3.5 mm]. The scan time was 8 minutes. T1-weighted imaging (T1WI) was obtained with a fast 3D gradient echo sequence (TR =1,900 ms, TE =2.52 ms, matrix =256×256, reversal time =900 ms, flip angle =9, voxel size =1×1×1 mm³, and FOV =250×250 mm²). The scanning time for 3D-T1WI was approximately 6 minutes and 10 seconds. There was no interval scanning, and only 176 scanning layers were employed.

Imaging data prepossessing

Preprocessing for rs-fMRI images was conducted with a MATLAB-based toolkit, Data Processing Assistant for Resting-State fMRI (version 5.3, http://rfmri.org/DPARSF). The preprocessing included the following: (I) discarding the first 10 volumes to avoid the initial magnetic field instability and subject state instability affecting the image quality; (II) slice timing correction to ensure the stability of the blood oxygenation level-dependent (BOLD) signal; (III) excluding those participants with head motion (>2.0 mm of translation or degrees of rotation in any direction) due to the functional signals being potentially corrupted by severe head motions (44); (IV) spatially normalizing the images to the standard Montreal Neurological Institute (NRI) of space and resampling to the voxel size of 3×3×3 mm³; (V) filtering the data to 0.01–0.1 Hz to attenuate respiration and other high-frequency physiological noises; and (6) smoothing with a 4-mm Gaussian kernel [except for regional homogeneity (ReHo) calculation].

In our study, the three preprocessed rs-fMRI metrics—mean amplitude of low-frequency fluctuations (mALFF), mean fractional ALFF (mfALFF,) and mean ReHo (mReHo)—were employed. The ALFF is considered to be the mean Fourier transform amplitude of each voxel time series within a certain frequency range. Meanwhile, the fALFF is the ratio of each voxel’s ALFF signal to the signal power across the whole frequency range (45). The ReHo is the Kendall W between a given voxel’s time series and its closest voxel time series (46). The final goal values (mALFF, mfALFF, and mReHo) were obtained by further normalizing the ReHo, ALFF, and fALFF maps for each voxel by dividing by the average voxel values for the entire brain.

Feature extraction and selection

MarsBaR volume of interest analysis toolbox for SPM2 was applied to acquire 116 ROIs, which were described by the anatomical automatic labeling (AAL) atlas. The utilities module in Data Processing & Analysis of Brain Imaging (v. 6.1 http://rfmri.org/dpabi) toolbox was selected, and the ROI signal extractor was employed to obtain the signal features. Ultimately, a total of 464 signal values features were acquired, including 116 mReHo values, 116 mALFF values, and 116 mfALFF values.

The radiomics features were extracted from Python 3.7 (Python Software Foundation; https://www.python.org) in PyCharm via the “PyRadiomics” package. We resampled the images to a 1×1×1 mm³ voxel size and quantized the gray levels to 25 levels. A total of 32,364 radiomics features were extracted from mALFF, mfALFF, and mReHo maps. These features were derived from 116 ROIs segmented with the AAL atlas, and 93 features were extracted from each brain region including 18 first-order features, 24 gray-level co-occurrence matrix (glcm) features, 14 gray-level dependence matrix (gldm) features, 16 gray-level run-length matrix (glrlm) features, 16 gray-level size-zone matrix (glszm) features, and 5 neighboring gray-tone difference matrix (ngtdm) features.

The extracted features have very high dimensionality, characterized by redundancy, irrelevant data, and duplicate information. We performed feature selection to improve classification performance and the reliability of our results. We first individually normalized each feature in the training set by deleting the mean and scaling to the unit variance and stored the mean and SD to normalize the features in the validation set. We then used the least absolute shrinkage and selection operator (LASSO) with 10-fold crossing validation to select the features with nonzero coefficients for classification.

Development and validation of the models

We used five machine learning classifiers including support vector machine (SVM), extreme gradient boosting (XGBoost), logistic regression (LR), k-nearest neighbors (KNN), and random forest (RF) to build predictive models based on the selected features, including rs-fMRI models, radiomics models, and combined models. In the model training process, we used grid search and 10-fold cross-validation to select the best hyperparameters. We assessed the predictive performance of the models using the area under the receiver operating characteristic curve (AUC).

Validation of key features

Given that features are divided into two categories, rs-fMRI signal values and radiomics features, there is an intersection of the distributed brain regions of the two types of features. The features present in these regions were considered to be representative features. To determine the relationship between these representative features and more fine-grained cognitive performance, correlation coefficients between the representative features and the T scores of the six cognitive domain tests were calculated for all participants. In further analyses, we assessed the differences in the distribution of key features screened according to correlation analyses in ANI and PWND to examine whether these features have reliable diagnostic value. Figure 2 provides a schematic illustration of the workflow of the study.

Figure 2 Schematic of the study procedure. Clinical and imaging information from the enrolled participants was obtained. Signal values of mALFF, mfALFF, and mReHo from the rs-fMRI were acquired through preprocessing, based on which the corresponding radiomics features were further captured. LASSO was applied to select features for constructing classification diagnostic models. Fifteen classification models were constructed and evaluated. Finally, the features were analyzed in a more refined manner, named as representative features and key features, respectively. AAL, anatomical automatic labeling; ANI, asymptomatic neurocognitive impairment; KNN, k-nearest neighbors; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; mALFF, mean amplitude of low-frequency fluctuations; mfALFF, mean fractional amplitude of low-frequency fluctuations; mReHo, mean regional homogeneity; PWND, people living with human immunodeficiency virus without neurocognitive disorders; RF, random forest; ROC, receiver operating characteristic; ROI, region of interest; rs-fMRI, resting-state functional magnetic resonance imaging; SVM, support vector machine; T1WI, T1-weighted imaging; XGBoost, extreme gradient boosting.

Statistical analysis

SPSS version 26.0 (IBM Corp., Armonk, NY, USA), R version 4.2.1 (The R Foundation of Statistical Computing), and Python software version 3.7.0 were used for all analyses. Basic clinical information, demographic characteristics, cognitive data, and features were subjected to independent samples t tests and the Wilcoxon rank-sum test according to whether they satisfied a normal distribution and chi-squared variance. The therapeutic schedule and plasma viral load were analyzed with the chi-squared test. To account for the performance differences between different machine learning models, we performed multiple statistical tests on the different models implemented the Kruskal-Wallis test followed by the Dunnett test. Correlations between the representative features and the six cognitive domains were analyzed via Pearson analysis, with r>0.3 indicating a significant correlation. A two-sided P value <0.05 was considered statistically significant. The intersecting brain regions corresponding to features were produced via BrainNet Viewer (https://www.nitrc.org/projects/bnv/) (47) on a standard MNI space brain surface.

Results

Baseline characteristics

We analyzed the demographic characteristics, clinical features, and NP performance of the participants. The results showed that there were no significant differences between the ANI and the PWND group in terms of age, education, duration of viral suppression, therapeutic schedule, disease course, or in CD4⁺ T-cell count, nadir CD4⁺ T-cell count, and ratios of CD4⁺ T-cell count to CD8⁺ T-cell count (all P values >0.05). However, on NP tests, the ANI group performed worse compared to the PWND group in all six cognitive domains, more significantly in both abstraction/executive and attention/working memory tests (all P values <0.05). These results were consistent in both the training and validation sets. In addition, most participants had undetectable HIV viral loads. All data details and statistical results are summarized in Table 1.

Feature selection and radiomics signature construction

We selected 21 rs-fMRI signal values and 28 radiomics features to construct the classification models. These rs-fMRI signal values included 4 mALFF, 6 mfALFF, and 11 mReHo features. These radiomics features include 9 features derived from the original mALFF image, including 3 first-order features, 4 GLCM features, and 2 GLSZM features;4 features derived from the original mfALFF image, including 1 first-order feature, 2 GLCM features, and 2 GLSZM features; and 15 features derived from the original mReHo image, including 5 first-order features, 5 GLCM features, 2 GLSZM features, 1 GLRLM feature, and 1 NGTDM feature. Tables S1,S2 show the selected rs-fMRI signal values and radiomics features in detail with the AAL template.

Interestingly, we found that 8 rs-fMRI signal value features and 8 radiomics features originated from intersecting brain regions. Specifically, we identified 1 mfALFF feature, 1 mReHo feature, and 1 radiomics feature (mReHo_glcm_Joint Average) from the left orbital part of the inferior frontal gyrus (ORBinf). In left superior parietal gyrus (SPG), 1 mALFF feature, 1 mReHo feature, and 2 radiomics features (mALFF_glcm_Informational Measure of Correlation 2 and mALFF_glszm_Small Area Emphasis) were identified. The right inferior parietal lobule (IPL) contributed 1 mfALFF feature and 1 radiomics feature (mReHo_glcm_Difference Variance). Additionally, 1 mReHo feature and 1 radiomics feature (mReHo_firstorder_Median) were found in the right caudate nucleus (CAU). Finally, right putamen contained 1 mReHo feature and 1 radiomics feature (mReHo_glrlm_Short Run Low Gray Level Emphasis), while the right cerebellum crus (CC) II included 1 ReHo feature and 2 radiomics features (mReHo_firstorder_Energy and mReHo_firstorder_Total Energy). The above-mentioned features were considered to be representative features in this study. Figure 3 visualizes the specific intersecting brain regions and the distribution of representative features within each brain region.

Figure 3 Visual presentation of the results from intersecting brain regions and representative features. (A) The six different colors represent the selected intersecting brain regions, with their anatomical positions indicated in the Anatomical Automatic Labeling template. (B) The 16 representative features and grouped into six intersecting brain areas for type and number presentation. The five categories are coded by color. Caudate_R, right caudate nucleus; Cerebellum_Crus2_R, right cerebellum crus II; Frontal_Inf_Orb_L, left orbital part of the inferior frontal gyrus; mALFF, mean amplitude of low-frequency fluctuations; mfALFF, mean fractional amplitude of low-frequency fluctuations, mReHo, mean regional homogeneity; Parietal_Sup_L, left superior parietal gyrus; Parietal_Inf_R, right inferior parietal lobule; Putamen_R, right putamen.

Assessment and validation of the construction models for predicting ANI

We used 5 classifiers, SVM, XGBoost, LR, KNN, and RF, to construct rs-fMRI model, radiomics models, and combined models (radiomics features and rs-fMRI signal values). Table 2 shows the AUC results of the different models. Detailed performance comparisons between the different classifiers are shown in Table S3 in the supplement.

Among the rs-fMRI models, the AUC of the 5 models in the validation set ranged from 0.742 to 0.856, with the KNN model having the highest AUC of 0.856 (95% CI: 0.751–0.961). In the radiomics models, the AUC in validation for the 5 models was concentrated at 0.862–0.900, with the XGBoost model having the highest AUC at 0.900 (95% CI: 0.815–0.985). Surprisingly, after the rs-fMRI signal value features were combined with radiomics features, the AUC of the combined models in the validation set was improved to differing degrees. Among the models, the RF model performed the best, with an AUC of 0.902, suggesting the advantage of multimodal data integration. Figure 4 shows the AUC of models in the validation set.

Figure 4 Comparison of the ROC curves of the models based on five machine learning classifiers in the validation set. ROC curves of models based on RF (A), LR (B), KNN (C), XGBoost (D), and SVM (E) classifiers in the validation set, respectively. ROC curves of the combined models based on the RF, LR, KNN, XGBoost, and SVM classifiers (F). AUC, area under the receiver operating characteristic curve; KNN, k-nearest neighbors; LR, logistic regression; RF, random forest; ROC, receiver operating characteristic; rs-fMRI, resting-state functional magnetic resonance imaging; SVM, support vector machine; XGBoost, extreme gradient boosting.

The Kruskal-Wallis results indicated that there was no statistical difference between the different machine learning models, indicating that the performance of the machine learning model itself exerted a small impact on the results. This highlights the stable significance of the regularity we found, independent of model choice. Similar performance has been demonstrated in other studies of machine learning models for cognitive disorders (48-50). However, due to the limited sample size and nonmulticenter source of the data, we aim to validate the above-mentioned findings in follow-up work. Nonetheless, we encourage peers and clinicians to consider the critical role of rs-fMRI histology in cognitive assessment and the significance of the joint application of histology and image values.

Validation of key features

As described in the Feature Selection and Radiomics Signature Construction section, we employed 16 representative features of 8 rs-fMRI signal value features and 8 radiomics features. We further analyzed the correlation between these features and the six cognitive domain test T-scores to identify key features (r>0.3). As shown in Figure 5A, in PLWH, the mfALFF signal value features of the left ORBinf and right IPL, as well as the radiomics feature (mReHo_glrlm_Short Run Low Gray Level Emphasis) of the right putamen were significantly positively correlated with the abstraction/executive domain (r=0.511, r=0.301, and r=0.398, respectively; all P values <0.001). In addition, all three key features differed significantly between the groups (all P values <0.001) and could be used to differentiate between ANI and PWND (Figure 5B-5D).

Figure 5 Screening results for key features and differences in the distribution in ANI and PWND. (A) Heatmap depicting the results of correlation analysis for the sixteen features and six cognitive domains. *, P<0.05; **, P<0.01; ***, P<0.001. Blue denotes negative correlations, and red denotes positive correlations, with deeper colors indicating stronger correlations. The features include Radiomics_1, Radiomics_mALFF_59_glcm_Informational Measure of Correlation 2, Radiomics_2, Radiomics_mALFF_59_glszm_Small Area Emphasis, Radiomics_3, Radiomics_mReHo_15_glcm_Joint Average, Radiomics_4, Radiomics_mReHo_62_glcm_Difference Variance, Radiomics_5, Radiomics_mReHo_94_firstorder_Total Energy, Radiomics_6, Radiomics_mReHo_72_firstorder_Median, Radiomics_7, Radiomics_mReHo_74_glrlm_Short Run Low Gray Level Emphasis, and Radiomics_8, Radiomics_mReHo_94_firstorder_Energy. (B-D) Boxplots of the differences between the three key features in ANI and PWND. Pink indicates ANI, and blue indicates PWND. ****, P<0.0001. ANI, asymptomatic neurocognitive impairment; glcm, gray-level co-occurrence matrix; glrlm, gray-level run-length matrix; glszm, gray-level size-zone matrix; mALFF, mean amplitude of low-frequency fluctuations; mfALFF, mean fractional amplitude of low-frequency fluctuations; mReHo, mean regional homogeneity; PWND, people living with human immunodeficiency virus without neurocognitive disorders.

Discussion

In this study, we developed a novel model for the quantitative diagnosis of ANI based on radiomics features extracted from rs-fMRI images that demonstrated good performance. We found that the classification performance of the model based on radiomics features was significantly higher than that of the rs-fMRI signal value features, and the classification performance of the model combining radiomics and signal value features was also significantly better than that of the model constructed with a single feature. In addition, some selected radiomics and rs-fMRI signal value features showed significant associations with the NP tests. Our results suggest that radiomics can effectively capture rs-fMRI information and potentially provide imaging markers of ANI.

This study innovatively introduced radiomics in rs-fMRI and employed high-throughput radiomics features extracted based on independent brain regions and combined features from different brain regions to build machine learning models. We found that the radiomics features could capture the changes in brain activity and demonstrated good discriminative ability for ANI; moreover, the classification performance of the model was further improved after combination with the rs-fMRI signal value features. These findings differ from previous studies based on rs-fMRI for HAND, which indicated changes in connectivity in the default mode network, frontoparietal network, and basal ganglia among individuals with HAND (51,52). These studies focused on the functional connectivity of different brain regions. DSouza et al. analyzed the changes in resting-state connectivity in individuals with HAND symptoms and found connections between the caudate nucleus and most brain regions (18). DSouza et al. also used mutual connectivity analysis to measure nonlinear interactions between the activity of various brain regions in rs-fMRI to classify patients with HAND from healthy controls (19). There are some similarities and differences to our study. In our work, we focused on the characteristic changes in the brain regions, which were primarily characterized by rs-fMRI signal value features and radiomics features. Through an in-depth analysis of rs-fMRI data, we were able to capture subtle changes in the functional connectivity of brain regions in the resting state, thereby clarifying the underlying neurobiological mechanisms. Meanwhile, the radiomics features provided us with high-throughput quantitative metrics to assess changes in brain regions from both structural and functional perspectives, further enhancing our understanding of disease-related brain changes. In addition, six overlapping brain regions were identified between radiomics features and rs-fMRI signal values features, among which was the right caudate. Abnormal functioning of the caudate may trigger cognitive deficits, which in turn negatively affect learning ability, working memory, and executive functions (53). These conclusions may offer avenues for further research into the mechanism of abnormal brain function in patients with ANI.

To the best of our knowledge, this was the first study to use machine learning models based on rs-fMRI to predict ANI. In this study, the XGBoost and RF models had excellent performance in the quantitative diagnosis of ANI. Specifically, the AUC in the training set and the AUC in the validation set of the XGBoost radiomics model were 0.990 and 0.900, respectively, indicating high discriminative ability and stability. XGBoost has excellent performance in handling high-dimensional data, automatically handling missing values and selecting relevant features (54). However, when the radiomics features and rs-fMRI signal value features were combined, the RF model was obviously superior to the other models, with an AUC of 1.000 in the training set and 0.902 in the validation set, while the AUC values of the XGBoost model were 1.000 and 0.856, respectively. This indicates that the RF model had a more stable performance in the validation set and better generalization ability. RF is now widely used in the development of medical predictive models; especially for data sets with a large number of predictions, it has excellent data processing capabilities and predictive performance (55). As an integrated learning algorithm, RF performs prediction by combining multiple decision trees, each of which relies on independently randomly sampled feature vectors, while the distributions of all trees in the forest are kept consistent (56). The RF model has better interpretability due to its decision tree-based feature, which is especially important for the medical field. RF has demonstrated a strong ability to handle highly nonlinearly correlated data with good noise robustness, easy parameter tuning, and the ability to efficiently leverage the advantages of parallel processing (56). Our findings were similar to those of several other studies that used rs-fMRI-based machine learning methods on neuroimaging data to predict cognitive function (57,58). XGBoost has excellent performance, but the interpretability of its model is relatively weak. Moreover, the combined model can better exploit the advantages of both data sources, thus achieving higher accuracy in practical applications. Our results suggest that combining rs-fMRI images with radiomics can obtain more image information from independent brain regions for objective and reliable ANI diagnosis, which to a certain extent, fulfills a clinical need of identifying ANI patients based on independent brain regions on rs-fMRI.

Notably, using 49 features selected via LASSO, we generalized the overlapping performance of the brain regions where the rs-fMRI signal value features and the radiomics features were located. We visualized these six brain regions of concern in the AAL template, which included the left ORBinf, left SPG, right IPL, right CAU, right putamen, and right CC II. Judaš et al. pioneered the isolation of the ORBinf from the remainder of the subfrontal gyrus (59), and subsequent brain imaging in humans has revealed that it plays a crucial role in the perception of semantic content and emotional expression (60). Nguchu et al. found that PLWH with early cognitive dysfunction have reduced functional connectivity between the left inferior frontal gyrus and other brain regions (61), and our results similarly pointed to the importance of this brain region. The SPG and IPL both belong to the parietal lobe, with former being involved in somatosensory and visuospatial integration and the latter being found to correlate with action performance in a neuroimaging study (62). The integrity of the frontoparietal fibers is associated with motor dysfunction (63). The ANI group did worse on the fine motor test than did the PWND group, which may be attributable to injury to the frontoparietal brain regions. Indeed, other research has verified that severe cortical thinning following HIV infection indicates structural changes in the brain (64,65). An autopsy study reported the presence of diffuse plaques formed by Aβ deposition in the frontal cortex of older HIV cases, as well as the accumulation of Aβ in dimeric form and the accumulation of α-synuclein in monomeric and trimeric forms (66). Longitudinal studies have found that HIV infection results in reduced gray-matter volume, cortical atrophy, and abnormal network connectivity in frontoparietal regions. Additionally, this damage continues to progress despite cART treatment and is strongly associated with cognitive decline (67,68).

It has been reported that decades after HIV infection, neurocognitive impairment related to the CAU are present in HIV-positive individuals, with the CAU undergoing a 6% reduction in volume over 10 years and the range of total shrinkage spanning from 1.3% to 12.3% (69). Our previously identified imaging features of the CAU that can be used as biomarkers for the differentiation of preclinical ANI and ANI (70). It should be noted that the striatum, composed of the CAU and putamen, receives afferent fibers from the cerebral cortex (primarily the frontal and parietal lobes) and is also an area of inflammation during acute HIV infection (71). In combination with preceding frontal and parietal lobe abnormalities, inflammatory damage to the striatum induced by viral infection further exacerbates cognitive impairment in fine motor function. The striatum regulates cognition involving executive and working memory, the two functions in which patients with ANI scored far below those of PWND in this study; these were also the two domains where ANI performed the worst among all six assessed cognitive domains. In a previous work, by constructing a model with simian immunodeficiency virus (SIV)-infected rhesus monkeys, our team found that the volume of the striatum shrinks progressively 12–24 months postinfection and is accompanied by a decrease in the gray-matter density of the caudate and chiasma nuclei, which is positively correlated with neurocognitive decline (72). Additionally, a transgenic rat model-based study revealed a direct association between electrophysiological abnormalities of striatal neurons and ion channel dysfunction in HIV infection (73). HIV infection leads to overactivity of medium spiny neurons in the striatum, with concomitant abnormalities in voltage-gated calcium channel function and elevated potassium channel activity, disrupting neuronal electrophysiological homeostasis.

Finally, CC I and II are the lateral expansions and most prominent regions of the cerebellum associated with cognitive and visuomotor functions (74). In our study, we observed signaling abnormalities primarily in the right CC II. However, the degree of brain region refinement presented across different studies varies, and additional investigation into the nuances of these regions should be conducted via detailed brain mapping.

Ultimately, the 16 features we identified were located in intersecting brain regions, and these may be essential for understanding brain function and pathological states. Of these, 8 of the signal value features involved ReHo, ALFF, and fALFF, which are able to reflect the intensity and consistency of neural activity at the level of brain regions and are commonly used as indicators to assess the brain functional separation. Based on the measurement of the temporal similarity in BOLD signals, ReHo reflects the consistency of activity pacing in functional brain regions, whereas ALFF and fALFF represent the intensity of activity in single voxel regions and in the low-frequency range, respectively (46,75,76). The other 8 radiomics features were texture features, including glcm, glszm, first-order, and glrlm which reflect an image’s microstructure based on the gray scale distribution and spatial relationship of pixels or voxels (77). Subsequent correlation analysis revealed a positive association between the abstraction/executive function and the features mReHo_glrlm_Short Run Low Gray Level Emphasis of the right putamen, mfALFF of the left ORBinf and right IPL. Given the association between these three features and cognition, they may serve as potential biomarkers for the differential diagnosis of ANI and PWND. In conclusion, our results indicate that frontoparietal and striatal features are highly valuable in the diagnosis of ANI and that functional abnormalities in these brain regions can be identified via rs-fMRI signal values and high-dimensional radiomics features.

Certain limitations to this study should be addressed. First, the sample size was limited, single-center in nature, and lacking female patients; this restricts the generalizability of our findings, particularly as it related to sex, and external validation is needed to further substantiate our conclusions. Furthermore, the cross-sectional design prevented us from observing the stability of focal brain regions, the selected features, and constructed models over the course of disease evolution. To address these issues, the sample size will be expanded to include patients from multiple institutions and across disease stages and sexes. Moreover, we will examine a larger, heterogeneous sample of patients with HAND over an extended period to develop robust and reliable diagnostic targets for rs-fMRI imaging. Our future work will also be aimed at validating the association between imaging biomarkers and brain pathological changes to improve the credibility and interpretability of the imaging results. Research involving longitudinal cohort studies and the construction of rhesus monkey SIV brain cognitive impairment models has been steadily progressing under funding of the National Natural Science Foundation of China, which has provided us with strong support for subsequent studies.

Conclusions

In this study, the RF combined model constructed based on rs-fMRI had excellent performance in the diagnosis of ANI. The method of binding signal value and radiomics features not only enhanced the diagnostic efficacy of the model but also elucidated the potential structural and functional alterations, which provides a novel perspective for the study of pathological mechanisms of ANI. The findings may contribute to the optimized replacement of the traditional NP test and can be expected to assist in clinical diagnosis. In the clinical management of patients with ANI, it is important to also consider modulating the activity and function of key brain regions, particularly the frontoparietal cortex and striatum. These regions are critically involved in the pathophysiological processes underlying ANI, and targeting them could potentially yield more effective therapeutic outcomes. In the future, the development of early warning prediction products based on the establishment of the imaging diagnostic biomarker system, combined with artificial intelligence technology, may enhance the personalization and precision of ANI diagnosis.

Acknowledgments

The authors gratefully acknowledge the contribution of the study participants.

Footnote

Reporting Checklist: The authors have completed the CLEAR reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-290/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-290/dss

Funding: This work was supported by Beijing Hospital Authority Clinical Medicine Development special funding support (No. ZLRK202333), the National Natural Science Foundation of China (No. 61936013, No. 82271963), the Beijing Municipal Natural Science Foundation (No. 7212051, No. L222097), and the Open Project of Henan Clinical Research Center of Infectious Diseases (AIDS) (No. KFKT202403).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-290/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by the Ethics Committee of Beijing Youan Hospital (No. LL-2023-070-K). Informed consent was obtained from all participants.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Life expectancy of individuals on combination antiretroviral therapy in high-income countries: a collaborative analysis of 14 cohort studies. Lancet 2008;372:293-9. [Crossref] [PubMed]
2. Churchill M, Nath A. Where does HIV hide? A focus on the central nervous system. Curr Opin HIV AIDS 2013;8:165-9. [Crossref] [PubMed]
3. Farhadian S, Patel P, Spudich S. Neurological Complications of HIV Infection. Curr Infect Dis Rep 2017;19:50. [Crossref] [PubMed]
4. Nyondo GG, Njiro BJ, Bwire GM. Cerebrospinal fluid viral escape in HIV patients on antiretroviral therapy: A systematic review of reported cases. Rev Med Virol 2024;34:e2536. [Crossref] [PubMed]
5. Hong S, Banks WA. Role of the immune system in HIV-associated neuroinflammation and neurocognitive implications. Brain Behav Immun 2015;45:1-12. [Crossref] [PubMed]
6. Peluso MJ, Ferretti F, Peterson J, Lee E, Fuchs D, Boschini A, Gisslén M, Angoff N, Price RW, Cinque P, Spudich S. Cerebrospinal fluid HIV escape associated with progressive neurologic dysfunction in patients on antiretroviral therapy with well controlled plasma viral load. AIDS 2012;26:1765-74. [Crossref] [PubMed]
7. Antinori A, Arendt G, Becker JT, Brew BJ, Byrd DA, Cherner M, et al. Updated research nosology for HIV-associated neurocognitive disorders. Neurology 2007;69:1789-99. [Crossref] [PubMed]
8. Heaton RK, Clifford DB, Franklin DR Jr, Woods SP, Ake C, Vaida F, et al. HIV-associated neurocognitive disorders persist in the era of potent antiretroviral therapy: CHARTER Study. Neurology 2010;75:2087-96. [Crossref] [PubMed]
9. Grant I, Franklin DR Jr, Deutsch R, Woods SP, Vaida F, Ellis RJ, et al. Asymptomatic HIV-associated neurocognitive impairment increases risk for symptomatic decline. Neurology 2014;82:2055-62. [Crossref] [PubMed]
10. Nightingale S, Winston A, Letendre S, Michael BD, McArthur JC, Khoo S, Solomon T. Controversies in HIV-associated neurocognitive disorders. Lancet Neurol 2014;13:1139-51. [Crossref] [PubMed]
11. Gisslén M, Price RW, Nilsson S. The definition of HIV-associated neurocognitive disorders: are we overestimating the real prevalence? BMC Infect Dis 2011;11:356. [Crossref] [PubMed]
12. Zhao J, Wu Y, Chen F, Zhao H, Chen J, Jing B, Li H. Distance-specific functional connectivity strength alterations in human immunodeficiency virus asymptomatic neurocognitive impairment patients: a cross-sectional study. Quant Imaging Med Surg 2024;14:1835-43. [Crossref] [PubMed]
13. Jiang X, Hou C, Ma J, Li H. Alterations in local activity and whole-brain functional connectivity in human immunodeficiency virus-associated neurocognitive disorders: a resting-state functional magnetic resonance imaging study. Quant Imaging Med Surg 2025;15:563-80. [Crossref] [PubMed]
14. Biswal B, Yetkin FZ, Haughton VM, Hyde JS. Functional connectivity in the motor cortex of resting human brain using echo-planar MRI. Magn Reson Med 1995;34:537-41. [Crossref] [PubMed]
15. Viviano JD, Buchanan RW, Calarco N, Gold JM, Foussias G, Bhagwat N, Stefanik L, Hawco C, DeRosse P, Argyelan M, Turner J, Chavez S, Kochunov P, Kingsley P, Zhou X, Malhotra AK, Voineskos ANSocial Processes Initiative in Neurobiology of the Schizophrenia. (s) Group. Resting-State Connectivity Biomarkers of Cognitive Performance and Social Function in Individuals With Schizophrenia Spectrum Disorder and Healthy Control Subjects. Biol Psychiatry 2018;84:665-74. [Crossref] [PubMed]
16. Yadav SK, Gupta RK, Hashem S, Bhat AA, Garg RK, Venkatesh V, Gupta PK, Singh AK, Chaturvedi S, Ahmed SN, Azeem MW, Haris M. Changes in resting-state functional brain activity are associated with waning cognitive functions in HIV-infected children. Neuroimage Clin 2018;20:1204-10. [Crossref] [PubMed]
17. Li R, Wang W, Wang Y, Peters S, Zhang X, Li H. Effects of early HIV infection and combination antiretroviral therapy on intrinsic brain activity: a cross-sectional resting-state fMRI study. Neuropsychiatr Dis Treat 2019;15:883-94. [Crossref] [PubMed]
18. DSouza AM. Abidin AZ, Schifitto G, Wismüller A. A multivoxel pattern analysis framework with mutual connectivity analysis investigating changes in resting state connectivity in patients with HIV associated neurocognitve disorder. Magn Reson Imaging 2019;62:121-8. [Crossref] [PubMed]
19. DSouza AM. Abidin AZ, Wismüller A. Investigating Changes in Resting-State Connectivity from Functional MRI Data in Patients with HIV Associated Neurocognitive Disorder Using MCA and Machine Learning. Proc SPIE Int Soc Opt Eng 2017;10137:101371C. [Crossref] [PubMed]
20. Chockanathan U. DSouza AM, Abidin AZ, Schifitto G, Wismüller A. Automated diagnosis of HIV-associated neurocognitive disorders using large-scale Granger causality analysis of resting-state functional MRI. Comput Biol Med 2019;106:24-30. [Crossref] [PubMed]
21. Turing A. Computing machinery and intelligence. Mind 1950;59:433-60.
22. Collins C, Dennehy D, Conboy K, Mikalef P. Artificial intelligence in information systems research: a systematic literature review and research agenda. International Journal of Information Management 2021;60:102383.
23. Sogandi F. Identifying diseases symptoms and general rules using supervised and unsupervised machine learning. Sci Rep 2024;14:17956. [Crossref] [PubMed]
24. Wagner MW, Namdar K, Biswas A, Monah S, Khalvati F, Ertl-Wagner BB. Radiomics, machine learning, and artificial intelligence-what the neuroradiologist needs to know. Neuroradiology 2021;63:1957-67. [Crossref] [PubMed]
25. Pinto A, Pereira S, Meier R, Wiest R, Alves V, Reyes M, Silva CA. Combining unsupervised and supervised learning for predicting the final stroke lesion. Med Image Anal 2021;69:101888. [Crossref] [PubMed]
26. Baur C, Wiestler B, Muehlau M, Zimmer C, Navab N, Albarqouni S. Modeling Healthy Anatomy with Artificial Intelligence for Unsupervised Anomaly Detection in Brain MRI. Radiol Artif Intell 2021;3:e190169. [Crossref] [PubMed]
27. Prinzi F, Currieri T, Gaglio S, Vitabile S. Shallow and deep learning classifiers in medical image analysis. Eur Radiol Exp 2024;8:26. [Crossref] [PubMed]
28. Mayerhoefer ME, Materka A, Langs G, Häggström I, Szczypiński P, Gibbs P, Cook G. Introduction to Radiomics. J Nucl Med 2020;61:488-95. [Crossref] [PubMed]
29. Jian Y, Peng J, Wang W, Hu T, Wang J, Shi H, et al. Prediction of cognitive decline in Parkinson’s disease based on MRI radiomics and clinical features: A multicenter study. CNS Neurosci Ther 2024;30:e14789. [Crossref] [PubMed]
30. Jian Y, Peng J, Wang W, Hu T, Wang J, Shi H, Li X, Chen J, Xu Y, Shao Y, Song Q, Shu Z. Using radiomics-based modelling to predict individual progression from mild cognitive impairment to Alzheimer’s disease. Eur J Nucl Med Mol Imaging 2022;49:2163-73. [Crossref] [PubMed]
31. Qi Y, Wang W, Rao B, Yang X, Yu W, Li JY, Sun ZC, Zhou F, Li YZ, Guo YF, Wang Y, Li HJ. Value of Radiomic Analysis Combined With Diffusion Tensor Imaging in Early Diagnosis of HIV ‐Associated Neurocognitive Disorders. Magnetic Resonance Imaging 2023;58:1882-91. [Crossref] [PubMed]
32. Liu G, Lu W, Qiu J, Shi L. Identifying individuals with attention-deficit/hyperactivity disorder based on multisite resting-state functional magnetic resonance imaging: A radiomics analysis. Hum Brain Mapp 2023;44:3433-45. [Crossref] [PubMed]
33. Wang L, Feng Q, Ge X, Chen F, Yu B, Chen B, Liao Z, Lin B, Lv Y, Ding Z. Textural features reflecting local activity of the hippocampus improve the diagnosis of Alzheimer’s disease and amnestic mild cognitive impairment: A radiomics study based on functional magnetic resonance imaging. Front Neurosci 2022;16:970245. [Crossref] [PubMed]
34. Vindbjerg E, Makransky G, Mortensen EL, Carlsson J. Cross-Cultural Psychometric Properties of the Hamilton Depression Rating Scale. Can J Psychiatry 2019;64:39-46. [Crossref] [PubMed]
35. Tramontano M, Belluscio V, Bergamini E, Allevi G, De Angelis S, Verdecchia G, Formisano R, Vannozzi G, Buzzi MG. Vestibular Rehabilitation Improves Gait Quality and Activities of Daily Living in People with Severe Traumatic Brain Injury: A Randomized Clinical Trial. Sensors (Basel) 2022.
36. Cai Y, Yang T, Yu X, Han X, Chen G, Shi C. The alternate-form reliability study of six variants of the Brief Visual-Spatial Memory Test-Revised and the Hopkins Verbal Learning Test-Revised. Front Public Health 2023;11:1096397. [Crossref] [PubMed]
37. Tam JW, Schmitter-Edgecombe M. The role of processing speed in the Brief Visuospatial Memory Test - revised. Clin Neuropsychol 2013;27:962-72. [Crossref] [PubMed]
38. Liesto S, Sipilä R, Hietanen M, Kalso E. Cognitive function is well preserved in a cohort of breast cancer survivors: Roles of cognitive reserve, resilience, and general health. Breast 2022;65:157-63. [Crossref] [PubMed]
39. Berrigan LI, LeFevre JA, Rees LM, Berard JA, Francis A, Freedman MS, Walker LAS. The symbol digit modalities test and the paced auditory serial addition test involve more than processing speed. Mult Scler Relat Disord 2022;68:104229. [Crossref] [PubMed]
40. Miles S, Howlett CA, Berryman C, Nedeljkovic M, Moseley GL, Phillipou A. Considerations for using the Wisconsin Card Sorting Test to assess cognitive flexibility. Behav Res Methods 2021;53:2083-91. [Crossref] [PubMed]
41. Kubo K, Hama S, Furui A, Mizuguchi T, Soh Z, Yanagawa A, Kandori A, Sakai H, Morisako Y, Orino Y, Hamai M, Fujita K, Yamawaki S, Tsuji T. Cognitive screening test for rehabilitation using spatiotemporal data extracted from a digital trail making test part-A. Heliyon 2024;10:e33135. [Crossref] [PubMed]
42. Shi C, Kang L, Yao S, Ma Y, Li T, Liang Y, Cheng Z, Xu Y, Shi J, Xu X, Zhang C, Franklin DR, Heaton RK, Jin H, Yu X. The MATRICS Consensus Cognitive Battery (MCCB): Co-norming and standardization in China. Schizophr Res 2015;169:109-15. [Crossref] [PubMed]
43. Meyer AL. Need to Revise Frascati Criteria for HIV-Associated Neurocognitive Disorders to Improve Relevance for Diverse Global Populations. Neurol Clin Pract 2022;12:328-30. [Crossref] [PubMed]
44. Li ZY, Si LH, Shen B, Yang X. Altered brain network functional connectivity patterns in patients with vestibular migraine diagnosed according to the diagnostic criteria of the Bárány Society and the International Headache Society. J Neurol 2022;269:3026-36. [Crossref] [PubMed]
45. Zou QH, Zhu CZ, Yang Y, Zuo XN, Long XY, Cao QJ, Wang YF, Zang YF. An improved approach to detection of amplitude of low-frequency fluctuation (ALFF) for resting-state fMRI: fractional ALFF. J Neurosci Methods 2008;172:137-41. [Crossref] [PubMed]
46. Jiang L, Zuo XN. Regional Homogeneity: A Multimodal, Multiscale Neuroimaging Marker of the Human Connectome. Neuroscientist 2016;22:486-505. [Crossref] [PubMed]
47. Xia M, Wang J, He Y. BrainNet Viewer: a network visualization tool for human brain connectomics. PLoS One 2013;8:e68910. [Crossref] [PubMed]
48. James C, Ranson JM, Everson R, Llewellyn DJ. Performance of Machine Learning Algorithms for Predicting Progression to Dementia in Memory Clinic Patients. JAMA Netw Open 2021;4:e2136553. [Crossref] [PubMed]
49. Hu M, Shu X, Yu G, Wu X, Välimäki M, Feng H. A Risk Prediction Model Based on Machine Learning for Cognitive Impairment Among Chinese Community-Dwelling Elderly People With Normal Cognition: Development and Validation Study. J Med Internet Res 2021;23:e20298. [Crossref] [PubMed]
50. Liu S, Jie C, Zheng W, Cui J, Wang Z. Investigation of Underlying Association Between Whole Brain Regions and Alzheimer’s Disease: A Research Based on an Artificial Intelligence Model. Front Aging Neurosci 2022;14:872530. [Crossref] [PubMed]
51. Thomas JB, Brier MR, Snyder AZ, Vaida FF, Ances BM. Pathways to neurodegeneration: effects of HIV and aging on resting-state functional connectivity. Neurology 2013;80:1186-93. [Crossref] [PubMed]
52. du Plessis L, Paul RH, Hoare J, Stein DJ, Taylor PA, Meintjes EM, Joska JA. Resting-state functional magnetic resonance imaging in clade C HIV: within-group association with neurocognitive function. J Neurovirol 2017;23:875-85. [Crossref] [PubMed]
53. Ellis R, Langford D, Masliah E. HIV and antiretroviral therapy in the brain: neuronal injury and repair. Nat Rev Neurosci 2007;8:33-44. [Crossref] [PubMed]
54. Lin X, Gao F, Lin H, Yao W, Wang Y. XGBoost-based nomogram for predicting lymph node metastasis in endometrial carcinoma. Am J Cancer Res 2024;14:5769-83. [Crossref] [PubMed]
55. Speiser JL. A random forest method with feature selection for developing medical prediction models with clustered and longitudinal data. J Biomed Inform 2021;117:103763. [Crossref] [PubMed]
56. Becker T, Rousseau AJ, Geubbelmans M, Burzykowski T, Valkenborg D. Decision trees and random forests. Am J Orthod Dentofacial Orthop 2023;164:894-7. [Crossref] [PubMed]
57. Khatri U, Kwon GR. Alzheimer’s Disease Diagnosis and Biomarker Analysis Using Resting-State Functional MRI Functional Brain Network With Multi-Measures Features and Hippocampal Subfield and Amygdala Volume of Structural MRI. Front Aging Neurosci 2022;14:818871. [Crossref] [PubMed]
58. Zhang T, Liao Q, Zhang D, Zhang C, Yan J, Ngetich R, Zhang J, Jin Z, Li L. Predicting MCI to AD Conversation Using Integrated sMRI and rs-fMRI: Machine Learning and Graph Theory Approach. Front Aging Neurosci 2021;13:688926. [Crossref] [PubMed]
59. Judaš M, Cepanec M. Oskar Vogt: The first myeloarchitectonic map of the human frontal cortex. Translat Neurosci 2010;1:72-94.
60. Rodd JM, Vitello S, Woollams AM, Adank P. Localising semantic and syntactic processing in spoken and written language comprehension: an Activation Likelihood Estimation meta-analysis. Brain Lang 2015;141:89-102. [Crossref] [PubMed]
61. Nguchu BA, Zhao J, Wang Y, Li Y, Wei Y, Uwisengeyimana JD, Wang X, Qiu B, Li H. Atypical Resting-State Functional Connectivity Dynamics Correlate With Early Cognitive Dysfunction in HIV Infection. Front Neurol 2020;11:606592. [Crossref] [PubMed]
62. Wang J, Yang Y, Fan L, Xu J, Li C, Liu Y, Fox PT, Eickhoff SB, Yu C, Jiang T. Convergent functional architecture of the superior parietal lobule unraveled with multimodal neuroimaging approaches. Hum Brain Mapp 2015;36:238-57. [Crossref] [PubMed]
63. Hamilton AF, Grafton ST. Action outcomes are represented in human inferior frontoparietal cortex. Cereb Cortex 2008;18:1160-8. [Crossref] [PubMed]
64. Yu X, Gao L, Wang H, Yin Z, Fang J, Chen J, Li Q, Xu H, Gui X. Neuroanatomical Changes Underlying Vertical HIV Infection in Adolescents. Front Immunol 2019;10:814. [Crossref] [PubMed]
65. Shin NY, Hong J, Choi JY, Lee SK, Lim SM, Yoon U. Retrosplenial cortical thinning as a possible major contributor for cognitive impairment in HIV patients. Eur Radiol 2017;27:4721-9. [Crossref] [PubMed]
66. Mackiewicz MM, Overk C, Achim CL, Masliah E. Pathogenesis of age-related HIV neurodegeneration. J Neurovirol 2019;25:622-33. [Crossref] [PubMed]
67. Xu F, Ma J, Wang W, Li H. A longitudinal study of the brain structure network changes in HIV patients with ANI: combined VBM with SCN. Front Neurol 2024;15:1388616. [Crossref] [PubMed]
68. Sanford R, Ances BM, Meyerhoff DJ, Price RW, Fuchs D, Zetterberg H, Spudich S, Collins DL. Longitudinal Trajectories of Brain Volume and Cortical Thickness in Treated and Untreated Primary Human Immunodeficiency Virus Infection. Clin Infect Dis 2018;67:1697-704. [Crossref] [PubMed]
69. Cilliers K, Muller CJF. Effect of human immunodeficiency virus on the brain: A review. Anat Rec (Hoboken) 2021;304:1389-99. [Crossref] [PubMed]
70. Qi Y, Xu M, Wang W, Wang YY, Liu JJ, Ren HX, Liu MM, Li RL, Li HJ. Early prediction of putamen imaging features in HIV-associated neurocognitive impairment syndrome. BMC Neurol 2021;21:106. [Crossref] [PubMed]
71. Sailasuta N, Ross W, Ananworanich J, Chalermchai T, DeGruttola V, Lerdlum S, Pothisri M, Busovaca E, Ratto-Kim S, Jagodzinski L, Spudich S, Michael N, Kim JH, Valcour V. RV254/SEARCH 010 protocol teams. Change in brain magnetic resonance spectroscopy after treatment during acute HIV infection. PLoS One 2012;7:e49272. [Crossref] [PubMed]
72. Song B, Li Y, Liu J, Mi H, Liu D, Wang W, Sun J, Wang Y, Li H. A longitudinal study of brain volume changes in rhesus macaque model infected with SIV. J Neurovirol 2020;26:581-9. [Crossref] [PubMed]
73. Khodr CE, Chen L, Al-Harthi L, Hu XT. HIV-Induced Hyperactivity of Striatal Neurons Is Associated with Dysfunction of Voltage-Gated Calcium and Potassium Channels at Middle Age. Membranes (Basel) 2022.
74. Balsters JH, Cussans E, Diedrichsen J, Phillips KA, Preuss TM, Rilling JK, Ramnani N. Evolution of the cerebellar cortex: the selective expansion of prefrontal-projecting cerebellar lobules. Neuroimage 2010;49:2045-52. [Crossref] [PubMed]
75. Yang H, Long XY, Yang Y, Yan H, Zhu CZ, Zhou XP, Zang YF, Gong QY. Amplitude of low frequency fluctuation within visual areas revealed by resting-state functional MRI. Neuroimage 2007;36:144-52. [Crossref] [PubMed]
76. Yang L, Yan Y, Li Y, Hu X, Lu J, Chan P, Yan T, Han Y. Frequency-dependent changes in fractional amplitude of low-frequency oscillations in Alzheimer’s disease: a resting-state fMRI study. Brain Imaging Behav 2020;14:2187-201. [Crossref] [PubMed]
77. van Timmeren JE, Cester D, Tanadini-Lang S, Alkadhi H, Baessler B. Radiomics in medical imaging-"how-to" guide and critical reflection. Insights Imaging 2020;11:91. [Crossref] [PubMed]