Individual-level cortical morphological network analysis in idiopathic normal pressure hydrocephalus: diagnostic and prognostic insights

Abstract

Background

Idiopathic normal-pressure hydrocephalus (iNPH) is a neurodegenerative disorder characterized by treatable cognitive impairment, remains poorly understood in terms of its underlying pathological mechanisms. Cortical morphological similarity network, which quantify synchronized morphological changes across brain regions, offer novel insights into inter-individual neuroanatomical variability. This study investigates individual-level cortical morphological network patterns in iNPH, explores their diagnostic utility and prognostic value for postoperative outcomes.

Methods

We enrolled 56 confirmed iNPH patients, 50 Alzheimer’s disease (AD) patients, and 60 healthy controls (HC). Cortical morphological similarity networks were constructed using a morphometric inverse divergence network (MIND) framework, integrating five key cortical features: cortical thickness, mean curvature, sulcal depth, surface area, and cortical volume. Graph theory analysis was employed to quantify global and nodal network properties. Partial correlations with MMSE scores assessed network-cognition relationships. A LASSO-regularized support vector machine (SVM) classifier differentiated iNPH, AD, and HC groups using regional MIND similarity (MINDs) features. Finally, preoperative MRI-derived MINDs were integrated into a LASSO-regularized support vector regression (SVR) model to predict postoperative cognitive and gait improvements following shunt surgery.

Results

Both iNPH and AD exhibited disrupted MIND network topology versus HC, including lower clustering coefficient, global efficiency, and local efficiency (all p < 0.05). Distinct spatial patterns emerged: iNPH showed localized lower values in cingulate subregions (degree centrality, node efficiency, MINDs), whereas AD demonstrated widespread alterations in fusiform, insular, and temporoparietal cortices. MMSE-associated MINDs in iNPH localized to frontostriatal circuits, contrasting with diffuse associations in AD. The multimodal classifier combining ventricular enlargement, regional brain volume, and MINDs achieved 87.00% accuracy (macro-AUC = 0.96) in three-group discrimination. Moreover, preoperative MINDs effectively predicted postoperative improvements in cognition and gait, with correlation coefficients of 0.941 and 0.889, respectively, between predicted and actual scores.

Conclusions

The MIND-based morphological similarity network reveals coordinated cortical morphological alterations in iNPH and highlights its heterogeneity compared to AD. These findings offer potential biomarkers to differentiate iNPH from AD. Furthermore, the predictive efficacy of MIND-based features for postoperative outcomes underscores their utility as non-invasive preoperative tools for evaluating shunt surgery effectiveness.

Background

Idiopathic normal pressure hydrocephalus (iNPH) is a neurological disorder characterized by normal cerebrospinal fluid (CSF) pressure within the skull, enlarged ventricles, cognitive decline, gait disturbance, and urinary incontinence [1]. Predominantly affecting older adults, iNPH remains underrecognized in clinical practice due to its insidious progression and substantial symptomatic overlap with age-related neurodegenerative pathologies, particularly Alzheimer’s disease (AD) and Parkinsonian syndromes [2]. Even experienced neuroradiologists struggle to differentiate iNPH from AD based on T1-weighted MR images, with reported accuracy rates ranging from 68 to 78% [3]. Conventional structural MRI markers, such as disproportionately enlarged subarachnoid space hydrocephalus (DESH), the Evans index (EI), and callosal angle (CA) exhibit limited sensitivity and specificity, particularly in early stages or comorbid neurodegenerative conditions [4]. Additionally, even advanced volumetric measures of hippocampal atrophy or midbrain morphometry, though useful in distinguishing AD or progressive supranuclear palsy (PSP), often fail to resolve diagnostic uncertainty due to shared atrophic patterns caused by CSF stasis in iNPH [5,6,7]. This diagnostic imprecision may result in shunt-eligible iNPH patients being misdiagnosed, while others undergo unnecessary invasive procedures. Therefore, the development of novel imaging markers for iNPH remains a critical research challenge.

Recent advances in neuroimaging have shifted the conceptualization of iNPH from a localized ventricular pathology to a disorder of large-scale brain network disruption [8]. Diffusion tensor imaging (DTI) studies reveal microstructural alterations in periventricular and frontal white matter tracts—including the corticospinal tract, corpus callosum, and superior longitudinal fasciculus—that correlate with gait disturbances and cognitive deficits [9, 10]. Functional MRI (fMRI) and electroencephalography (EEG) further implicate dysregulated interactions within the default mode network (DMN), salience network, and frontoparietal circuits, potentially explaining the heterogeneity of clinical symptoms and variable shunt responsiveness [11,12,13]. These findings suggest that iNPH disrupts the integration ability of large-scale networks, potentially contributing to the heterogeneity in its clinical presentation and shunt surgery efficacy.

Beyond traditional connectivity metrics derived from DTI-based tractography, emerging paradigms emphasize the role of interregional morphometric similarity networks (MSNs). MSNs quantify structural connection by assessing correlations between regions by taking a vector (or distribution) of one or more structural MRI geometric indices measured locally in each brain region [14,15,16,17]. Grounded in the principle of homophily—whereby cytoarchitectonically similar regions exhibit preferential connectivity—MSNs provide insights into neurodevelopmental trajectories and alterations in cortical organization. This approach has proven valuable in mapping neurodevelopmental trajectories [18, 19], neurodegenerative progression including AD [20, 21] and mild cognitive impairment (MCI) [22]; as well as neuropsychiatric disorders such as schizophrenia [23, 24] and depression [25, 26]. For instance, MSN alterations in AD involve disrupted similarity in the insula, hippocampus, and temporo-frontal regions, reflecting structure reconfigurations or synchronized degeneration patterns [20, 21]. In multiple sclerosis (MS), MSNs demonstrate reorganization independent of global atrophy pattern, highlighting their sensitivity to network-level pathology [27]. In iNPH, multiple mechanisms such as mechanical compression caused by ventricular dilatation and chronic hypoperfusion cause cortical deformation, and its atrophy pattern is significantly different from typical neurodegenerative diseases such as AD [6]. Preliminary work by Yin et al. [28] reveals the topological reconstruction of the iNPH gray matter covariant network, which is manifested by enhanced modularity and decentralized hub nodes and may be a unique compensatory mechanism caused by cerebrospinal fluid dynamics disorders [28]. However, critical limitations persist: (1) The group average network masks individual variation; (2) Single imaging features (e.g., gray matter volume) fails to capture the multidimensional nature of cortical alterations; and (3) The relationship between MSN alterations and neurologic dysfunction (e.g., cognitive decline) in iNPH remains poorly understood.

To address these limitations, we employed Morphometric Inverse Divergence (MIND), a robust method that constructs individualized MSNs by quantifying multivariate morphological similarity using high-dimensional features (cortical thickness, curvature, sulcal depth, surface area, and volume). Unlike conventional approaches, MIND leverages vertex-level data to model regional feature distributions, enhancing biological validity and technical reliability [29]. By integrating graph theory and machine learning, we aim to: (1) identify aberrant MIND network topology in iNPH; (2) elucidate associations between network disruption and cognitive impairment; and (3) evaluate MIND-derived biomarkers for differentiating iNPH from AD and predicting post-shunt outcomes, which holds the potential to improve early diagnosis and enhance patient outcomes (Fig. 1).

Fig. 1

The flow chart of this study. The individual-level morphological similarity network (MSN) is constructed using the Morphometric INverse Divergence (MIND) method, which integrates five high-dimensional structural features: cortical thickness (CT), mean curvature (MC), sulcal depth (SD), surface area (SA), and cortical volume (CV). Subsequently, graph theory analysis is applied to examine the topological properties of the MIND network. Additionally, correlation analyses are conducted to assess the relationship between regional MIND similarity features and cognitive impairment scores in iNPH group and AD group. The discriminative power of regional MIND similarity (MINDs) in distinguishing iNPH, AD, and HC groups is evaluated using Lasso-SVM. Finally, a Lasso-SVR regression model, based on preoperative MRI-derived features, is developed to predict improvements in cognitive and gait functions following shunt surgery. D-K, Desikan-Killiany atlas; HC, Healthy controls; iNPH, idiopathic normal pressure hydrocephalus; AD, Alzheimer’s disease

Methods

Participants

We enrolled individuals who were medically identified with iNPH at Huadong Hospital affiliated with Fudan University from August 2021 to October 2023. The inclusion criteria for iNPH patients were in strict accordance with the guidelines (Third Edition) for iNPH management, with all patients meeting the criteria for possible, probable, and definite iNPH [30]. These criteria included (1) individuals who were 60 years of age or older; (2) the manifestation of clinical signs that align with at least one of the following conditions: an unsteady gait, a decline in cognitive function, or urinary dysfunction, which gradually worsened over a period exceeding 6 months; (3) routine MRI demonstrating an enlargement of the lateral ventricle (with an Evans index > 0.3), a reduction in the subarachnoid space above the high-convexity regions (DESH), and no significant cerebrospinal fluid obstruction; and (4) a noticeable alleviation of symptoms postcerebrospinal fluid tap test (CSF-TT) and subsequent lumboperitoneal (LP) shunt surgery, which met the criteria for a definite diagnosis of iNPH. The exclusion criteria were those with recent heavy alcohol consumption, a history of severe mental illness hospitalization, or other established etiologies of dementia, including but not limited to psychiatric, metabolic, or neoplastic conditions. Furthermore, individuals with secondary normal pressure hydrocephalus due to conditions such as cerebral infarction, head trauma, intracranial bleeding, or meningitis were not included in the study.

For comparison, HC and subjects diagnosed with AD were selected at random from the hospital’s patient pool and were age-matched with the iNPH group. The diagnosis of AD was in line with the NINCDS-ADRDA criteria [31]. The HC group was defined by the absence of gait instability, cognitive decline, or urinary issues; normal findings on routine MRI scans of the head; and no evidence of systemic or psychological disorders.

The study ultimately included a total of 171 subjects, which included those with a confirmed diagnosis of iNPH (n = 61), AD (n = 50) and HC (n = 60). Ethical approval for the study was granted by the Medical Ethics Committee at Huadong Hospital (reference number: 2017 K027). The informed consent of all subjects was waived.

Neurological examinations

Comprehensive neurological assessments were administered to all individuals with a diagnosis of iNPH. These evaluations were performed by a pair of experienced neurosurgeons, both prior to the CSF-TT and at 3-month intervals after shunt surgery. The battery of tests included the Idiopathic Normal Pressure Hydrocephalus Grading Scales (iNPHGS), which is designed to assess the severity of iNPH-related symptoms; the Mini Mental State Examination (MMSE), which is used to evaluate cognitive function (range: 0–30); and the Timed Up and Go (TUG) test, which assesses motor coordination and gait stability. Higher MMSE scores and lower TUG test scores represent better performance.

MRI acquisition

All participants had 3D high-resolution T1-weighted images scanned on a Prisma 3.0 T scanner (Siemens). Images were acquired via an MPRAGE sequence with the following imaging parameters: TR = 1800 ms, TE = 2.37 ms, field of view = 250 mm × 250 mm, number of excitations = 1, sagittal slice thickness = 0.85 mm, gap = 50%, and total slices = 208.

Image processing and feature extraction

Assessment of ventricular dilatation

Two experienced neuroradiologists conducted the assessment of ventricular enlargement via 3D-T1 MRI. The evaluation encompassed a suite of established indices: the commonly used Evans index (EI), the z-Evans index (zEI), the brain-to-ventricle ratio (BVR), the callosal angle (CA), the distance between the callosum and ventricles (CVD), the height of the callosum (CH), the distance between the callosal and commissural regions (CCD), and the width of the cella media (CMW). Elaborate descriptions of these indices can be found in Supplementary Fig. S2.

Assessment of regional volume

The structural MRI data were processed via FreeSurfer software (version 7.2.0, https://surfer.nmr.mgh.harvard.edu/fswiki), which employs the recon-all command. This standard procedure involves skull stripping, B1 bias field correction, differentiation between gray and white matter, segmentation of hemispheric and subcortical structures, and reconstruction of cortical surface models (gray-white boundary surface and pial surface) (Fig. 2) [32]. Following a visual quality assessment, five participants with subpar image quality in the iNPH cohort were excluded from the analysis. For each remaining participant, the total intracranial volume (TIV) was calculated, serving as a reference for normalizing subsequent morphological assessments. Additionally, the volumes of gray matter (GMV), white matter (WMV), and cerebrospinal fluid (CSF) were quantified. Our focus extended to brain areas typically implicated in iNPH pathology, encompassing the bilateral lateral ventricles, third ventricles, fourth ventricles, corpus callosum, choroid plexus, and hippocampus. To ensure accuracy in our volume measurements, two experienced radiologists manually fine-tuned the segmentation outcomes generated by FreeSurfer in native space.

Fig. 2

Examples of automated segmentation of cortical and subcortical structure using FreeSurfer on 3D T1-weighted images in axial, coronal, and sagital planes

Assessment of cortical morphometric features

Following the generation of cortical surface reconstruction, a comprehensive set of morphological metrics was derived, including measurements of cortical thickness (CT), mean curvature (MC), sulcal depth (SD), surface area (SA), and cortical volume (CV). The resulting morphometric maps were aligned with the fsaverage surface template, a standardized anatomical template with 163,842 vertices per hemisphere. This standardized approach allowed for the precise characterization of each vertex by these five morphological attributes, all of which were extracted from the structural MR datasets.

Construction of MIND networks

The MIND-based morphological similarity network consistes of nodes (brain regions) and edges (the similarity in the distribution of morphological indicators across regions). Nodes were defined based on the Desikan-Killiany (D-K) atlas, which was further subdivided into 308 spatially contiguous cortical regions, each approximately uniform in size (~ 500 mm²), to minimize size-related variability [33]. A detailed description of the D-K308 atlas can be found at https://github.com/RafaelRomeroGarcia/subParcellation. The D-K308 template was then aligned to the native cortical space of each participant via the mri_surf2 surf tool, enabling comprehensive quantification of MRI-derived metrics for all vertices within each cortical region.

For the definition of edges, we employed a novel methodology known as MIND to quantify the morphological distribution similarity between paired brain regions, as outlined by Sebenius et al. [29]. The MIND similarity metric lies in estimating multivariate KL divergence by using a k-nearest neighbor approach from the observed vertex-level data, thus eliminating the need for complex density estimation and improving computational efficiency. Specifically, we first extracted five morphological features: CT, MC, SD, SA, and CV, for all vertices within each brain region. These features are chosen because they provide critical information about the brain’s cortical structure and reflect various aspects of its geometry and morphology. Each of these features is computed for every vertex on the cortical surface, capturing the fine-grained characteristics of the brain’s surface geometry. These values were then subjected to Z-score normalization to standardize the scale of the morphological features across different brain regions. MIND metric computes the structural similarity between paired brain regions by comparing the distribution of the five morphological features within each region. To ensure interpretability and comparability, the MIND values are constrained within the range [0, 1] through an inverse transformation. Specifically, a MIND value of 0 indicates the complete absence of similarity, while a value of 1 signifies perfect similarity between two regions. A detailed introduction to the MIND algorithm was provided in the Supplementary Materials. We final constructed an MIND network matrix of dimension 308 × 308 for each subject.

Network property analysis

For each of the individual MIND network (308 × 308 weighted matrices), a set of topological properties was assessed in MATLAB via the DPABINet toolbox (https://rfmri.org/DPABI). The following graph-theoretic metrics were calculated, comprising seven global topological properties categorized into two dimensions: (1) classical small-world parameters including clustering coefficient (Cp), characteristic path length (Lp), standardized clustering coefficient (γ), standardized characteristic path length (λ), and small-worldness (σ); (2) efficiency metrics encompassing global efficiency (Eglob) and local efficiency (Eloc). Additionally, two nodal topological properties were analyzed: nodal degree centrality (DC) and nodal efficiency (NE). The detailed formulas and interpretations of these measures can be found elsewhere [34]. The sparsity (S) threshold range was 0.01 < S < 0.50 with an interval of 0.01. For MIND network at each sparsity level, we calculated the area under the curve (AUC) for each topological property, which is highly sensitive to the topology of brain disease abnormalities [35].

Regional MIND similarity strength refers to the average weight of connections (edges) from a given node within the MIND network. This metric reflects the overall strength or intensity of morphological similarities emanating from each brain region. Specifically, it is computed as the mean value of the MIND similarity between a node and all other 307 cortical areas, as expressed in the following equation:

$$MINDs = \frac{{\sum\nolimits_{{\left\{ {i \ne j} \right\}}} {MIND\left( {i,j} \right)} }}{n - 1}$$

where $n$ is the total number of brain regions. $MIND(i,j)$ represents the MIND similarity between region $i$ and region $j$.

Classification analysis

In this study, we developed a three-class classification model to distinguish iNPH patients, AD patients, and HC via the least absolute shrinkage and selection operator (LASSO) algorithm and the OneVsRest classifier from the scikit-learn library in Python. MRI-derived features such as ventricular morphology measurements (VM), brain volume (VOL), and regional MINDs, either alone or in combination, were used as model inputs. Prior to implementing LASSO for feature selection, we employed the f_classif function within the feature_selection module of scikit-learn to conduct an ANOVA, which served as a preliminary screening mechanism for features. Features with a p value less than 0.05 across the three groups were selected, forming a candidate feature set aimed at mitigating the risk of overfitting due to high-dimensional data. For the OneVsRestClassifier, we utilized the support vector machine (SVM) classifier as a foundational classifier. To optimize the hyperparameters of both the LASSO and SVM classifiers, we employed a grid search with tenfold cross-validation. The search space for LASSO’s L1 regularization parameter, alpha, ranged from 10^–6 to 10². For the SVM, the hyperparameters optimized included the kernel type (‘rbf’ and ‘linear’) and the regularization parameter C [range: (0.001, 0.01, 0.1, 1, 10)].

To ensure an unbiased evaluation of the model’s performance, we conducted leave-one-out cross validation and calculated macro-averaged accuracy, precision, recall, and F1 scores on the basis of the confusion matrix. The receiver operating characteristic (ROC) curve was also plotted to visualize the classification performance. In addition, we also employed SHapley additive exPlanations (SHAP) analysis to compute the average SHAP values for each feature within the best model, providing insights into the impact of individual features on the model’s predictions. This analysis also facilitated the visualization of key features, enhancing our understanding of the model’s decision-making process.

Postsurgical prognosis analysis

Given the strong correlation between postoperative neurological function scores and preoperative baseline scores, we quantified the change in neurological function scores relative to preoperative scores, which was calculated as the ratio of the difference between postoperative and preoperative clinical scores to the preoperative score ($\Delta s$), which was calculated as (postoperative score-preoperative score)/preoperative score. We used features from preoperative MR images, including ventricular morphology measurements (VM), brain volume (VOL), and regional MINDs, both individually and in combination, as model inputs. Age and sex were used as covariates. Univariate correlation analysis and cascading Lasso feature selection identified the optimal feature subset. The support vector regressor (SVR) was employed to build the prediction model. We optimized hyperparameters for Lasso and SVR via a grid search with tenfold cross-validation. Lasso tuning matched our classification analysis, whereas the SVR hyperparameters included a linear kernel and a range of regularization parameters C [range: (0.001, 0.01, 0.1, 1, 10)].

Predictive ability was assessed via leave-one-out cross validation, which provides individual patient outcome predictions. Model performance was gauged via regression metrics, namely, the coefficient of determination (R²), mean square error (MSE), and mean absolute error (MAE), which collectively evaluate the mode’s performance in predicting post-shunt clinical score changes. Additionally, we applied the predicted change rate to the baseline clinical scores to estimate the postoperative scores. We then calculated the Pearson correlation coefficient between these predicted scores and the actual postoperative scores to quantify the model’s predictive accuracy.

Statistical analysis

SPSS version 26.0 was employed for the statistical analysis. The X² test was utilized to examine sex distribution differences. For other continuous variables, ANOVA or the Kruskal–Wallis test was used to test for differences among the three groups, followed by Tukey’s post hoc analysis. For pre- and postshunt surgery score comparisons in the iNPH group, the Wilcoxon matched pairs signed-rank test was employed. Manual ventricular dilatation measurements’ interobserver reliability was determined via the intraclass correlation coefficient (ICC).

Independent sample t tests were conducted to analyze MIND network global and nodal property differences between groups, adjusting for age, sex, and TIV as covariates. Partial correlation analysis, controlling for the same covariates, explored associations between regional MIND similarity and MMSE scores in the iNPH and AD groups. Significant brain regions were mapped onto the Yeo 7 network to examine their corresponding functional subnetworks [36]. The above statistical analyses were all subjected to 1000 permutation tests by randomly shuffling the group assignments to ensure the reliability of the results. FDR correction was used to correct the results for multiple comparisons as appropriate, and P < 0.05 after correction was considered statistically significant.

Results

Study population

Table 1 shows the demographics, clinical scores, and MRI measurements of ventricular dilatation in the three groups. Age and sex varied significantly across groups (p < 0.05). The MMSE scores of the iNPH group were notably lower than those of the AD group (p = 0.009). The distributions of imaging measurements for ventricular dilatation and regional brain volumes across the three groups are shown in Supplementary Fig. S3. Notably, there was partial overlap in the imaging measurements of ventricular dilatation and normalized regional volumes among the three groups.

Table 1 Demographics, clinical data, and measurements of ventricular dilation

Global properties of MIND network

Figure 3 showed the group mean MIND similarity for each group (top) and the differences of regional dissimilarity in each condition relative to the control group (bottom).

Fig. 3

The group mean MIND similarity for each group (top) and the differences of regional dissimilarity in each condition relative to the control group (bottom). HC, healthy controls; iNPH, idiopathic normal pressure hydrocephalus; AD, Alzheimer’s disease; iNPH-HC, regional differences in group mean MIND similarity between iNPH group and HC group; AD-HC, regional differences in group mean MIND similarity between AD group and HC group; iNPH-AD, regional differences in group mean MIND similarity between the iNPH group and the AD group;

All the participants’ MIND networks presented greater gamma (γ) and similar lambda (λ) values than did the random networks did, indicating that the MIND brain networks of all the participants presented small-world properties (γ > 1, λ≈1, σ = γ/λ > 1), as shown in Supplementary Fig. S4. The comparison of global topological properties among the iNPH, AD and HC groups were shown in Fig. 4, revealing the following important findings: (1) the Cp was significantly lower in both the iNPH and AD groups (p < 0.05, permutation test, uncorrected); (2) the Lp was greater in the iNPH group with no significant difference (p > 0.05, permutation test, uncorrected), but was higher in the AD group (p < 0.001, permutation test, uncorrected); (3) both the local efficiency (Eloc) and global efficiency (Eglob), as did the Lambda (λ) parameter, were significantly lower in the affected groups (p < 0.001, permutation test, uncorrected). For the comparison of iNPH group and AD group, Cp, Eloc, and Eglob were found to be greater, and Lp was lower in the iNPH group.

Fig. 4

Comparison of global topological properties among the iNPH, AD and HC groups. Compared with those in the HC cohort, the iNPH and AD groups exhibited lower values of Cp, Eglob, Eloc, and lambda (λ), while a greater Lp value was observed. Abbreviations: HC, healthy controls; iNPH, idiopathic normal pressure hydrocephalus; AD, Alzheimer’s disease; Cp, clustering coefficient; Lp, shortest path length; Eglob, global efficiency; Eloc, local efficiency; small-world parameters:γ, Gamma;λ, Lambda; σ, Sigmma; * indicates a significant difference between groups, p < 0.05;** indicates a significant difference between groups, p < 0.01;*** indicates a significant difference between groups, p < 0.001. NS indicates no statistical significance

Nodal properties of MIND network

As shown in Fig. 5 and detailed in Supplementary Table S1, we observed significant variations in nodal properties and regional MINDs within distinct brain regions when comparing iNPH and AD groups to the HC group. The iNPH group exhibited pronounced lower in degree centrality (DC), node efficiency (NE), and MINDs in the bilateral caudal anterior cingulate, bilateral isthmus cingulate, and bilateral posterior cingulate regions, with these measures being relatively higher than those observed in the AD group. In contrast, the AD group experienced a more extensive alteration in nodal properties, including bilateral fusiform, bilateral insula, bilateral superior temporal, and bilateral supramarginal gyrus, all of which had notably weaker MIND similarity in comparison to the iNPH group. Despite shared patterns of nodal properties between the iNPH and AD groups in some brain areas, the AD group has exhibited a more extensive range of affected areas.

Fig. 5

Nodal region with significant alterations in the nodal properties of the MIND network among iNPH, AD, and HC subjects (P < 0.05, permutation test, FDR corrected). All brain regions were defined by DK-308; see Supplementary Table S2 for the full names of the nodes in the figure. The blue spheres indicate brain regions with lower nodal properties, while the red spheres represent brain regions with greater nodal properties. The size of the spheres corresponds to the significance of group differences in node properties, with larger spheres indicating more significant differences. DC, degree centrality; NE, node efficiency; MINDs, MIND similarity; L, left hemisphere; R, right hemisphere

Correlations with cognitive impairment

Figure 6 showed the brain regions whose MIND similarity was associated with MMSE scores in iNPH-MINDs and AD-MINDs. Significant correlations between the MIND similarity of 20 brain region nodes and MMSE scores in the iNPH cohort (Fig. 6a, d) were primarily distributed across four critical functional networks: the visual network (VN) (including rh_lateraloccipital, lh_lingual, rh_pericalcarine, rh_lateraloccipital), the ventral attention network (VAN) (including lh_caudalanteriorcingulate, rh_rostralmiddle frontal), the somatomotor network (SMN) (including rh_lateralorbitofrontal, lh_frontal pole), and the default mode network (DMN) (including lh_precuneus, rh_parstriangularis, lh_rostralmiddlefrontal, lh_caudalmiddlefrontal, lh_inferiorparietal). In the AD cohort, MIND similarity in 14 brain region nodes correlated with MMSE scores (Fig. 6b, e), which were predominantly located in the VAN (including rh_supramarginal, bilateral caudal anterior cingulate), the SMN (including lh_superior parietal, lh_transversetemporal, rh_postcentral), and the DMN (including lh_superior frontal, rh_rostralanterior cingulateulate, lh_superiortemporal, lh_medialorbitofrontal). Notably, the MIND similarity of lh_caudalanteriorcingulate in the VAN (iNPH: r = 2.351, p < 0.05; AD: r = 3.093, p < 0.05) and the lh_superiortemporal strength in the DMN (iNPH: r = 2.055, p < 0.05; AD: r = 2.051, p < 0.05) exhibited significant positive correlations with MMSE scores in both the iNPH and AD cohorts.

Fig. 6

a and b illustrate the color bars representing regional nodes with significant correlations between MIND similarity (MINDs) and Mini-Mental State Examination (MMSE) scores in iNPH and AD groups, respectively. These nodes are assigned to the Yeo-7 functional network atlases. The color bars represent the correlation coefficients between regional MINDs and MMSE scores within each patient group. c provides a reference visualization of the seven Yeo functional networks rendered on the lateral surface of the left hemisphere of the brain. d In the iNPH group, 20 nodal regions exhibit significant correlations between MINDs and MMSE scores. These regions are displayed on the lateral surfaces of both hemispheres and in a dorsal view of the brain. Refer to (a) and (c), they are primarily distributed across four critical functional networks: the visual network (VN) encompassing nodes such as rh_laterialoccipital, lh_lingual, rh_pericalcarine, and rh_lateraloccipital; the ventral attention network (VAN) including lh_caudalanteriorcingulate and rh_rostralmiddlefrontal; the somatomotor network (SMN) represented by rh_lateralorbitofrontal and lh_frontal pole; and the default mode network (DMN) incorporating nodes like lh_precuneus, rh_parstriangularis, lh_rostralmiddlefrontal, lh_caudalmiddlefrontal, and lh_inferiorparietal. e In the AD group, 14 nodal regions demonstrate significant correlations between MINDs and MMSE scores. Refer to (a) and (c), they primarily belong to three functional networks: the ventral attention network (VAN), including rh_supramarginal and bilateral caudalanteriorcingulate; the somatomotor network (SMN) encompassing lh_superior marginal, lh_transversetemporal, and rh_postcentral; and the default mode network (DMN) featuring lh_superiorfrontal, rh_rostral_acingulate, lh_superiortemporal, and lh_medialorbitofrontal. The red circles indicate brain regions with significantly positive correlations between MINDs and MMSE scores, while the blue circles represent regions where MINDs exhibit significant negative correlations with MMSE scores

Classification results

We developed three-class classification models to differentiate iNPH patients, AD patients, and HC via individual features, including ventricular expansion indicators (VM), standardized brain volume (VOL), and regional MINDs of 308 ROIs, as well as a comprehensive combination of these feature sets. Table 2 presents the macro average performance metrics for each model. The model that integrated all three feature sets achieved the highest classification performance, with a macro average accuracy of 87.00% and a macro AUC of 0.96. The confusion matrix indicates the proportion of each class correctly identified (Fig. 7a–g). Compared with models based on single features, the model incorporating all three feature sets improved the likelihood of correctly classifying AD patients and HC while reducing the misclassification of iNPH patients as AD patients.

Table 2 Performance in distinguishing iNPH, AD and HC (macro-average)

Fig. 7

The plots of the confusion matrix, ROC curves and feature importance quantified by the SHAP value. a–g Confusion matrices and ROC curves of three-class classification models to differentiate iNPH patients, AD patients, and HC via individual features, including ventricular expansion indicators (VM), standardized brain volume (VOL), and the node MIND similarity (MINDs) of 308 ROIs, as well as a comprehensive combination of these feature sets. The value in the confusion matrix indicates the proportion of each class correctly identified. The model that integrated all three feature sets (VM_VOL_MINDs) achieved the highest classification performance, with an accuracy of correct prediction for iNPH, AD and HC of 0.92, 0.83, and 0.86, respectively, and a macro AUC of 0.96. h–j Ranking of feature importance quantified by the SHAP value of the model based solely on VM, VOL and the integrated feature set (VM, VOL and MINDs), which reflects the contribution of features to the predicted outcome. In the best model, which is based on the integration of all three feature sets, the MINDs make a larger contribution to the predictive output. Among them, the rh_posterioculate brain region had the most significant influence

The ROC curves and AUC values revealed that when ventricular expansion metric (VM) and brain volume (VOL) features, or their combination, were considered, the diagnostic capability of the model, as reflected by the AUC values, was ranked as follows: HC > iNPH > AD (Fig. 7a–g). However, when MINDs were included in the analysis, the model’s discriminative ability for the AD group improved, with the AUC ranking as HC > AD > iNPH. The model that combined all three feature sets demonstrated the highest diagnostic capabilities for iNPH patients, AD patients, and HC, with AUC values of 0.95, 0.97, and 0.97, respectively.

The SHAP analysis quantified the importance of features. As depicted in Fig. 7h–j, the larger the absolute average SHAP value of a feature is, the more significant its impact on the model’s output. In the model based solely on ventricular expansion indicators (VM), the importance ranking of features was BVR > CMW > CA > z-EI > CVD (Fig. 7h). This finding indicates that when considering only ventricular expansion indicators, BVR made the most significant predictive contribution. In the model based on single-brain volume features, as shown in Fig. 7i, the volume feature of Left_Inf_Lat_Vent (left inferior lateral ventricle) had the most significant impact on the model’s output, followed by the 3rd_Ventricle (third ventricle) and CC_Central (central corpus callosum) features. Notably, in the model that integrated all three feature sets, the MINDs made a larger contribution to the predictive output. Among these, the rh_posterioculating brain region had the most significant influence (Fig. 7j).

Prediction of shunt outcome

Both the MMSE score and the TUG score before and after shunt surgery were strongly correlated, with r = 0.828 (p < 0.001, Fig. 8a) and r = 0.754 (p < 0.001, Fig. 9a). These findings indicate that the baseline neurological function of patients plays a crucial role in postoperative recovery. With the rate of change between postoperative and preoperative scores as the prediction target, we tested the predictive performance of improvements in cognitive function (dMMSE) and gait function (dTUG) based on a single feature and a combination of three features, as shown in Figs. 8c and 9c. The model that integrated all three feature sets demonstrated the best predictive performance for cognitive improvement (R² = 0.638, MAE = 0.081, MSE = 0.009), with the model’s predictions highly correlated with actual postoperative scores (r = 0.941, p < 0.001) (Fig. 8b). Similarly, the model that integrated all three feature sets also showed the best performance in predicting the rate of change in the TUG score after surgery (R² = 0.616, MAE = 0.068, MSE = 0.008), with the model’s predictions highly correlated with the actual postoperative scores (r = 0.889, p < 0.001) (Fig. 9b). Interestingly, the standalone MINDs had predictive capabilities for both the dMMSE and dTUG (dMMSE: R² = 0.621, MAE = 0.078, MSE = 0.01; dTUG: R² = 0.569, MAE = 0.074, MSE = 0.009), whereas the standalone VM and VOL features did not exhibit predictive power.

Fig. 8

The predictive performance of regional MIND similarity (MINDs) on cognitive improvement post-shunt surgery. a The post-shunt MMSE score exhibited a strong correlation with the preoperative baseline MMSE score (r = 0.828, p < 0.001). b The predicted post-shunt MMSE score based on the regression model demonstrated a high correlation with the actual post-shunt MMSE score (r = 0.941, p < 0.001). c The predictive performance of cognitive function improvement (dMMSE) was evaluated using four distinct input features models:VM-based, the VOL-based, MINDs-based, and an integrated model incorporating VM, VOL, and MINDs features (VM_VOL_MINDs). The combined model exhibited superior predictive performance with the highest R-squared value (R² = 0.638), alongside minimal mean absolute error (MAE = 0.081) and mean squared error (MSE = 0.009). d This figure illustrates the feature importance weights from the VM_VOL_MINDs regression model used for dTUG predicting. Each dot and line represents the estimated weight or coefficient of a specific predictor. Positive weights signify that an increase in the corresponding regional MINDs is associated with higher post-shunt MMSE scores, while negative weights indicate a decrease in MINDs corresponds to lower MMSE scores. Key predictors include cerebrospinal fluid volume (CSF) and specific brain regions such as the left isthmus cingulate, left lateral occipital, and right rostral middle frontal cortex

Fig. 9

Fig. 9 The predictive performance of regional MIND similarity (MINDs) on gait improvement post-shunt surgery. a The post-shunt TUG score exhibited a strong correlation with the preoperative baseline TUG score (r = 0.754, p < 0.001). b The predicted post-shunt TUG score based on the regression model demonstrated a high correlation with the actual post-shunt TUG score (r = 0.889, p < 0.001). c The predictive performance of gait function improvement (dTUG) was evaluated using four distinct input features models:VM-based, the VOL-based, MINDs-based, and an integrated model incorporating VM, VOL, and MINDs features (VM_VOL_MINDs). The combined model exhibited superior predictive performance with the highest R-squared value (R² = 0.616), alongside minimal mean absolute error (MAE = 0.068) and mean squared error (MSE = 0.008). d This figure illustrates the feature importance weights from the VM_VOL_MINDs regression model used for dTUG predicting. Each dot and line represents the estimated weight or coefficient of a specific predictor. Positive weights signify that an increase in the corresponding regional MINDs is associated with higher post-shunt TUG scores, while negative weights indicate a decrease in MINDs corresponds to lower TUG scores. Key predictors include callosal angle (CA) angle, as well as the lh_inferior temporal, lh_lateral occipital, and lh_supramarginal regions

The weights of the predictors in the regression model reflect the contribution of the features to the predicted outcome. As shown in Fig. 8d, the features selected for predicting dMMSE primarily involved cerebrospinal fluid volume and regions such as the left isthmus cingulate, left lateral occipital, and right rostral middle frontal regions. These brain regions also overlap with the abovementioned brain regions found to be associated with preoperative MMSE scores (e.g., rh_rostralmiddle frontal.part10, rh_lateraloccipital.part5, lh_superiortemporal.part3). Among these, the morphological similarity enhancement of the lh_precuneus with other brain regions was a strong predictor of MMSE score changes. The predictive features for the dTUG included the CA angle, as well as the lh_inferior temporal, lh_lateral occipital, and lh_supramarginal regions (Fig. 9d).

Discussion

By calculating the MIND values across various regions of the whole brain, we established individual-level MSN, introducing a new paradigm to identify the brain regions most significantly affected by iNPH. The observed alterations in the topological properties of MSN suggested a reduction in both global and local regional interactions associated with iNPH. The distinct neuroanatomical correlates of cognitive function observed in iNPH and AD patients underscored the heterogeneity of brain structural deformation in cognitive decline. By utilizing the Lasso-SVM method with multistructural features, we trained a three-class classification model that accurately distinguished iNPH patients, AD patients, and HC. Finally, our Lasso-SVR-based predictive model indicates that MSN features can independently predict improvements in cognitive and gait function after shunt surgery. These findings provide an objective reference for screening patients who may benefit from shunt surgery.

Altered patterns of MIND networks

The primary characteristic of the human brain network is the small-world property, characterized by high clustering coefficients and short path lengths. Although our results indicate that both iNPH and AD patients, as well as HC individuals, exhibit small-world topological properties, the clustering coefficients (Cps) of the MIND networks in the iNPH and AD groups were significantly lower than those in the HC group, whereas the characteristic path length (Lp) was greater, suggesting a trend toward randomization in the network structure of the affected groups. In alignment with previous studies based on functional or white matter structural networks [37, 38], the reduction in global and local efficiency within the MIND networks of the affected groups indicates a weakening of functional segregation and integration within the brain networks. Although the perspective of physiology to explain the connection between brain areas is quite complex, MSNs have demonstrated conserved structural-transcriptional coupling across cortical regions, with neurodevelopmental plasticity governed by astrocyte- and oligodendrocyte progenitor-driven gene expression [19]. The formation of the brain by congenital development and acquired learning and memory plasticity results in brain areas with similar morphology distributions, which help coordinate cognitive and perceptual tasks [39, 40]. In addition, morphological similarity relationships across individual brain regions may characterize the structural features of the constituent circuits between individual neurons or neuronal clusters [41]. However, in disease states, neuronal loss and synaptic degeneration disrupt these interregional connections, particularly in vulnerable hub regions such as cortical sulci, leading to diminished network efficiency and impaired functional integration. This dual pathology suggests that MSN alterations in iNPH might stem from disruptions in both specific neurodevelopmental plasticity and acquired axonal disconnection.

At the local brain region level, we observed that, compared with HC group, iNPH presented significantly lower degree centrality, nodal efficiency, and average MIND similarity coefficients in the bilateral cingulate regions. The cingulate cortex has extensive cortico-subcortical connections and partially overlaps with the default mode network (DMN). Previous BOLD studies have confirmed compensatory changes in the DMN in iNPH patients [42]. Additionally, DTI and sMRI studies have shown that significant lateral ventricular expansion in iNPH causes compression and stretching of the corpus callosum and affects the entire cingulate cortex [15, 43]. Consequently, disruption of the cingulate cortex may be a key region affected by iNPH. Interestingly, we also found that MSN similarity in the bilateral cuneus, bilateral lateral occipital, and left lingual regions was significantly greater in the iNPH group than in the HC group. These regions, which are all part of the VN network, play crucial roles in visual processing and integration. The heightened MSN similarity observed in these regions suggests that there may be structural reorganization occurring within the VN in iNPH patients, potentially as a result of disease-related changes in neural function and connectivity. For example, our previous functional network-based study revealed increased connectivity between the VAN and VN in iNPH patients, suggesting a possible compensatory mechanism [44]. This adaptive response likely involves the VAN, which is responsible for directing attention toward behaviorally relevant stimuli, strengthening its connection with the visual network (VN). Such heightened coupling aims to mitigate impaired visuospatial processing caused by structural distortions associated with iNPH. These findings contribute to our understanding of adaptive changes in brain networks during iNPH conditions.

Compared to the HC group, AD patients exhibited more pronounced nodal alterations, particularly in the bilateral fusiform region (part of VN), bilateral insula (part of SMN), bilateral superior temporal region (part of DMN), and bilateral supramarginal area (part of VAN). SPECT studies have demonstrated that cerebral hypoperfusion in AD patients corresponds to these regions of atrophy [45]. Likewise, Hořínek et al. identified significant reductions in fractional anisotropy (FA) within key regions, including the fusiform gyrus, bilateral parieto-temporal areas, and supramarginal regions [46]. The overlap of above findings with our observations highlights a strong association between disrupted connectivity and underlying structural changes, such as reduced gray matter volume, neuronal loss, or white matter fiber degeneration. These alterations likely drive coordinated structural adaptations across interconnected brain regions, reflecting a compensatory or pathological response to network-level impairments. While there are similarities in nodal properties between iNPH and AD patients in certain brain regions, the AD group exhibited a distinct pattern of brain tissue damage that was more extensive and profound. Unlike the focal atrophy caused by ventricular enlargement in iNPH, AD is characterized by a pattern of diffuse brain atrophy. The pathological progression of AD begins with the deposition of tau proteins and subsequent brain tissue shrinkage in the hippocampal pathways, which then spreads to involve the temporal, parietal, and frontal lobes, affecting the brain more widely [47, 48].

Cognition-related MIND changes

Previous studies have described iNPH cognitive impairment as frontal subcortical dementia, characterized by executive function and short-term memory dysfunction [49]. Consistent with this research, our study found that the synchrony of morphological deformations in brain regions associated with the frontostriatal circuit, including the left caudal anterior cingulate cortex (part of the DMN) and the right rostral middle frontal gyrus (part of the VAN), were significantly correlated with MMSE scores. These findings underscore the critical role of the frontostriatal circuit in cognitive dysfunction in iNPH. Additionally, our study revealed that MIND similarity strength of brain regions such as the right supramarginal gyrus and the left cuneus was positively correlated with the MMSE score. These brain regions are involved in the DMN (as shown in Fig. 6a, d). This aligns with prior evidence linking DMN disruptions in iNPH to executive dysfunction and verbal memory deficits [50, 51]. Interestingly, while frontal subcortical changes are often emphasized in iNPH, our results also identified a non-frontal dominated pattern of cognitive impairment. Specifically, multiple nodes in the occipital lobe, which is responsible for visual information processing and storage, exhibited significant negative correlations with MMSE scores. These nodes include the right lateral occipital cortex (part of the visual network, VN), the left lingual gyrus (part of the VN), and the right pericalcarine cortex (part of the VN). This finding aligns with previous reports by Paulo’s, who emphasized visuospatial impairment as a hallmark of iNPH [52]. This nonfrontal type-dominated change may be a compensatory mechanism of the disease.

In contrast, AD is characterized by a pervasive cognitive dysfunction that affects a broad spectrum of cognitive faculties, rather than being confined to specific functions. A correlation analysis of the node MIND similarity across the entire brain in AD patients revealed significant associations between MMSE scores and several key brain regions, including the right supramarginal gyrus (part of the VAN), left superior frontal gyrus (part of the DMN), left superior parietal lobule (part of the SMN), left transverse temporal gyrus (part of the SMN), and bilateral caudal anterior cingulate gyrus (part of the VAN). Consistent with study by Arlt’s, who have reported significant associations between the volume of the hippocampus and entorhinal cortex and performance on memory tasks, the left temporal cortex and language tasks, and the frontal, parietal, and occipital lobes and executive functions as well as cognitive speed [53]. These findings underscore the widespread impact of AD on multiple cognitive networks [54]. Collectively, these findings emphasize the distinct neuroanatomical correlates of cognitive function in iNPH and AD patients, highlighting the heterogeneity of brain networks involved in cognitive decline. This heterogeneity underscores the complexity of brain networks involved in cognitive decline, which would provide valuable insights for differential diagnosis and therapeutic strategies targeting these neurodegenerative diseases.

Classification performance

The reliance solely on ventricular expansion metrics or brain volume characteristics fails to differentiate all iNPH patients from AD patients and healthy controls, which is consistent with previous observations [48, 55]. In our study, we further noted that even when multiple ventricular expansion metrics and brain volume features were combined, the model’s accuracy in distinguishing iNPH patients from AD patients and HC was only 77%, with an AUC of 0.92. However, the integration of MIND features resulted in a 10% increase in macroaveraged accuracy and an AUC of 0.96. This improvement underscores the significant enhancement in the model’s classification capabilities caused by the inclusion of MIND features, making the distinction between iNPH, AD, and HC more discernible. Furthermore, the enhancement in model performance, as indicated by the confusion matrix and ROC curves, is reflected primarily in the improved identification of the HC and AD groups. This improvement may be attributed to the additional disease-relevant information provided by the morphological similarity networks, which is not fully captured by traditional ventricle measurement indices or brain volume characteristics. Consequently, this finding underscores the limitation of relying solely on conventional brain region-specific measures in the differential diagnosis of iNPH from other brain disorders, potentially restricting their clinical specificity and sensitivity. This conclusion is also supported by previous findings [6, 56]. By introducing MIND network, our study captured the complex interactions between various brain regions across the entire brain, revealing subtle yet critical morphological changes. SHAP analysis further confirmed the significant role of MIND features, such as the bilateral posterior cingulate, bilateral lateral occipital, and superior temporal regions, in the prediction process. These brain regions also appeared in the node attribute analysis results. Therefore, incorporating MIND network provides a novel perspective for distinguishing iNPH patients from AD patients and HC, enhancing the accuracy and confidence of clinical disease differentiation.

Prediction of shunt outcome

Several studies have explored potential prognostic factors, including age, comorbidities, severity, duration of symptoms, imaging findings, and so on, following shunt surgery. However, there has been no consistent and reproducible reporting of parameter differences between responders and nonresponders to iNPH shunt surgery [57]. A retrospective single-center study by Virhammar et al. in 2014 suggested that patients with DESH, a smaller corpus callosum angle, and larger temporal angles had better shunt outcomes [58]. However, a meta-analysis by Thavarajasingam et al. in 2023 revealed that only the corpus callosum angle and periventricular white matter changes were effective radiological factors for diagnosing shunt responders among iNPH patients [59]. Nonetheless, these factors were not sufficient as standalone predictors. In this study, we found that the corpus callosum angle combined with MIND features played a significant role in predicting postoperative gait function outcomes (Fig. 9d). A smaller CA correlated with a more significant decrease in the TUG score from pre- to postoperative, indicating greater improvement in gait function. In contrast, the combination of CSF volume combined with MINDs significantly contributes to the prediction of postoperative cognitive outcomes. While Dan and colleagues demonstrated that volumetric analysis effectively predicts gait and cognitive outcomes following shunt surgery in iNPH patients [16], our research emphasized that subtle cortical morphological similarity features may offer superior predictive accuracy for cognitive improvement compared with macroscopic volumetric characteristics. The strong correlation between the predicted results and the actual results in this study once again confirms the feasibility of using preoperative MRI to evaluate the potential outcomes of shunt surgery in patients.

Notably, the precuneus emerged as a strong predictor for both changes in MMSE scores and TUG scores postoperatively. The precuneus is a critical component of the default mode network and is involved in motor planning and executive functions. This finding aligns with previous research showing that gait phenotypes among iNPH patients are linked to global cognition as assessed with the MMSE [60]. Our results underscore the importance of the precuneus in maintaining normal cognitive and gait functions and may provide new targets for interventions aimed at individuals with cognitive and gait disorders.

Limitations

Several limitations need to be addressed in future studies. First, it was a single-center study with relatively few participants. Although the consistency of data collection strongly supports the reliability of the analytical results, the reproducibility and generalizability of the proposed model need to be further validated and improved in multicenter and larger cohort studies. Second, potential confounding factors such as diabetes, hypertension, severity and duration of symptoms were not taken into account in this study. These factors could influence the MIND network and, consequently, might diminish the study’s statistical power. Another possible source of confounding factors in this research is the presence of patients with both iNPH and AD or PD. Although such cases were excluded from the current analysis, they are frequently encountered in clinical practice. Further investigation is needed to determine whether dual-diagnosis patients can be identified via these methods.

Conclusions

Using the multivariate morphological distribution similarity estimation, this study revealed the coordinated alteration pattern of brain morphology in iNPH patients from the individual-level brain morphological networks, promoting the development of research from isolated regional studies to whole-brain large-scale effects. These findings provided new biomarkers for distinguishing iNPH from AD, especially the morphological changes occurred in the bilateral posterior cingulate, bilateral lateral occipital, and superior temporal regions. Furthermore, the effectiveness of MIND network in predicting the postoperative outcomes of shunt surgery highlights its important significance in the preoperative non-invasive assessment of shunt effects. Further studies would combine morphological, anatomical, and functional networks to further elucidate the role of specific cortical circuits in related dysfunction patterns, thereby facilitating more targeted and effective treatment strategies.