Multimodal brain network topology and enhanced computer-aided diagnosis in Parkinson’s Disease: a systematic review and meta-analysis

Abstract

Parkinson’s disease (PD) is increasingly recognized as a brain network-disconnection syndrome. However, there is little consistent evidence on multimodal global topological alterations and their diagnostic value. We systematically searched PubMed, Embase and Web of Science up to March 2025 for articles reporting brain network topology in PD, to which we applied a multilevel random-effects meta-analyses with robust variance estimation to account for statistical dependencies. Our case-control meta-analysis included 80 studies (42 fMRI, 25 dMRI, 10 EEG, 4 sMRI, 3 others) involving 3736 PD patients and 2384 healthy controls. Compared to controls, PD patients showed lower structural and functional network segregation, especially when cognitively impaired. Structural network integration was also lower in PD, such deficits appearing to correlate with disease progression. Drug and network construction strategies were identified as potential moderating factors. Our diagnostic meta-analysis of 10 studies yielded a pooled diagnostic odds ratio of 16.4 and a pooled area under the curve of 0.86, with better diagnostic performance observed in studies using combined network metrics. These results support the clinical relevance of topological metrics in PD as potential biomarkers for disease characterization, prognosis and patient stratification, and underscore the importance of methodological harmonization and prospective validation in future research.

Introduction

Parkinson’s disease (PD), an increasing prevalent neurodegenerative disease, poses a growing socioeconomic burden^1,2. Currently, diagnosis is predominantly clinical, and it remains challenging to identify individuals with atypical symptoms or in preclinical and prodromal phases. Robust and reliable biomarkers for early diagnosis, prognosis, and disease management are urgently needed³. Advances in psychoradiology⁴ offer promise for identifying such biomarkers through noninvasive characterization of brain structure⁵ and function.

Brain connectivity can be measured noninvasively in vivo by various neuroimaging methods⁶: functional connectivity is constructed from the statistical dependences between regional neurophysiological signals acquired by functional MRI (fMRI), positron emission tomography (PET), electroencephalography (EEG), and magnetic encephalography (MEG); structural connectivity is determined by fiber tracking using diffusion MRI (dMRI)⁷; and structural similarity is estimated by morphological metrics of correlation/divergence in structural MRI (sMRI)⁸. Graph theoretical analysis (GTA) provides a powerful tool to quantify the global topological organization of brain networks (the ‘connectome’)⁹, whose alterations underlie many neuropsychiatric disorders^10,11.

PD is increasingly conceptualized as a disconnection syndrome¹². Although many studies have reported brain network abnormalities in PD, findings are inconsistent^13–15. In such situations a meta-analysis can be useful for identifying robust, consistent patterns of alterations. In an earlier meta-analysis¹⁶ we were able to define some robust patterns of network alterations in PD, but this considered only a single modality (dMRI) and failed to account explicitly for within-study statistical dependencies. To address these limitations, here we used advanced multilevel random-effects models to account for statistical dependencies¹⁷, conducting a multimodal GTA meta-analysis for a comprehensive assessment of brain dysconnectivity across both structural and functional dimensions¹⁸.Given the diagnostic promise of GTA metrics¹⁹, combining them with computational techniques may enable automated early PD diagnosis, although their diagnostic utility will require comprehensive quantitative validation.

This systematic review and meta-analysis aimed to identify consistent multimodal global brain network alterations in PD and to assess the diagnostic accuracy of GTA metrics. The study included both case-control and diagnostic meta-analyses. In the case-control meta-analysis, recognizing that distinct symptom subtypes may involve different network alterations^13,20,21, we also conducted a meta-analysis focused on cognitive impairment—the most commonly reported PD subtype. Furthermore, we systematically reviewed studies investigating the role of global network topology in disease progression and treatment response.

Results

Characteristics of the included studies in case-control meta-analysis

Figure 1 shows the flow chart for study identification, screening, and selection. Overall, 136 articles met the eligibility criteria for systematic review, and Table S1–5 details their methods and main findings. Eighty studies comprising 3736 PD patients and 2384 healthy controls (HC) fulfilled the inclusion criteria for quantitative meta-analysis: 68 PD datasets from 42 fMRI studies comprised 2183 PD patients (mean [SD] age, 61.1 [5.2] years; mean male, 57.9%); 39 PD datasets from 25 dMRI studies included 1416 PD patients (mean [SD] age, 63.1 [3.5] years; mean male, 58.3%); 4 sMRI studies included data from 182 PD patients (mean [SD] age, 56.3 [2.3] years; mean male, 47.9%); and 10 EEG studies enrolled 292 PD patients (mean [SD] age, 64.9 [4.8] years; mean male, 64.4%): demographic and clinical characteristics are summarized in Table 1. Details of demographic and clinical characteristics are presented in Table S6–9, and the GTA methods used in each study are described in Table S10–12.

Fig. 1. PRISMA flow diagram.

The figure depicts the literature search and selection criteria. Abbreviations: HC healthy controls, PD Parkinson’s disease.

Table 1.

Summary of demographic and clinical characteristics in the included multimodal neuroimaging studies

Characteristic	Metrics	dMRI	fMRI	sMRI	EEG
No. of studies	Numbers	25	42	4	10
No. of datasets	Numbers	39	68	5	12
No. of PD	Numbers	1416	2183	182	292
Age	Mean (SD)	63.1 (3.5)	61.1 (5.2)	56.3 (2.3)	64.9 (4.8)
	Missing datasets	0	0	0	2
Gender (Male%)	Mean (SD)	58.3 (10.7)	57.9 (11.9)	47.9 (10.8)	64.4 (18.2)
	Missing datasets	1	0	0	0
Education	Mean (SD)	11.3 (2.2)	11.2 (2.5)	9.6 (1.4)	10.7 (1.3)
	Missing datasets	9	17	1	8
Duration	Mean (SD)	5.0 (2.6)	5.2 (2.7)	3.4 (1.8)	5.8 (1.5)
	Missing datasets	1	4	0	3
UPDRS-III	Mean (SD)	25.0 (7.2)	24.1 (8.2)	20.6 (3.8)	26.1 (7.6)
	Missing datasets	0	6	0	3
H&Y	Mean (SD)	2.1 (0.5)	1.0 (0.3)	2.0 (0.4)	1.7 (0.4)
	Missing datasets	7	9	1	9
MMSE	Mean (SD)	27.6 (1.4)	27.8 (1.40)	28.1 (0.5)	25.9 (1.7)
	Missing datasets	14	31	1	8
Drug	On-state	15	31	1	6
	Drug-naïve/ off-state	22	37	3	2
	Missing datasets	2	0	1	4
LEDD	Mean (SD)	563 (125)	563 (271)	293 (127)	655 (173)
	Missing datasets	20	18	0	7

PD Parkinson’s disease, UPDRS unified Parkinson’s Disease rating scale, H&Y Hoehn and Yahr, MMSE mini-mental state examination, LEDD levodopa-equivalent daily dose, dMRI diffusion MRI, fMRI functional MRI, sMRI structural MRI, EEG electroencephalography, SD standard deviation.

Primary results in case-control meta-analysis

The different modalities detected somewhat different patterns of alteration in PD: dMRI revealed deficits in both network segregation (lower clustering coefficient and local efficiency) and integration (lower global efficiency and higher characteristic path length); fMRI showed mainly changes in segregation (lower clustering coefficient and modularity); and sMRI and EEG reported no consistent abnormalities (Fig. 2). These findings represent partially overlapping but modality-specific patterns of network disruption, rather than cross-modality convergence on a single underlying biology. Table 2 summarizes the results of the pooled meta-analysis for global topological abnormalities in PD, and Forest plots are presented in Figs. S1–15.

Fig. 2. Orchard plots of differences in graph metrics.

Orchard plots showing the mean differences in A clustering coefficient, B local efficiency, C characteristic path length, D global efficiency, E normalized clustering coefficient, F normalized characteristic path length, G small-worldness, and H modularity between Parkinson’s disease and healthy controls across multiple modalities. Dots represent effect sizes from individual studies, with colors indicating imaging modalities (dMRI red, fMRI blue, sMRI purple, EEG green); the size of each dot reflects the corresponding sample size. Black circles and horizontal error bars represent the pooled effect sizes and standard errors. Three studies with extreme effect sizes (2.51 for clustering coefficient in fMRI, 4.78 for characteristic path length in dMRI, and 2.91 for global efficiency in fMRI) are omitted for clarity (see forest plots in Figs. S3, S6, and S11). Significance levels are indicated as follows: P < 0.05 (*), P < 0.01 (**), P < 0.001 (***). Abbreviations: dMRI diffusion MRI, fMRI functional MRI, sMRI structural MRI, EEG electroencephalography.

Table 2.

Meta-analysis effect size and heterogeneity in multimodal studies

	Metrics	No. of studies	No. of effect sizes	g	95% CI	SE	t	P from t	Q	P from Q	I²%	Tau
Modalities
dMRI	Clustering coefficient	20	30	-0.328	-0.513 to -0.143	0.088	-3.735	0.002	108.310	< 0.001	75.109	0.421
	Local efficiency	15	26	-0.272	-0.454 to -0.090	0.085	-3.209	0.007	55.906	< 0.001	54.884	0.288
	Characteristic path length	20	30	0.396	0.193 to 0.599	0.096	4.121	0.001	112.343	< 0.001	78.691	0.472
	Global efficiency	21	34	-0.445	-0.615 to -0.276	0.081	-5.496	< 0.001	89.490	< 0.001	64.917	0.348
	Normalized clustering coefficient	9	15	0.245	0.037 to 0.452	0.089	2.736	0.026	24.368	0.041	41.166	0.203
	Normalized characteristic path length	9	15	0.174	-0.065 to 0.413	0.103	1.685	0.131	22.913	0.062	50.210	0.243
	Small-worldness	15	25	0.160	-0.027 to 0.348	0.087	1.840	0.088	46.820	0.004	54.810	0.273
	Modularity	3	4	0.125	-0.448 to 0.697	0.117	1.067	0.410	3.341	0.342	5.550	0.057
	Assortativity	2	3	0.161	-0.753 to 1.075	0.072	2.239	0.267	2.513	0.285	19.055	0.099
	Strength	4	5	-0.413	-1.147 to 0.320	0.227	-1.820	0.169	10.821	0.029	66.454	0.367
	Degree	2	4	-0.379	-1.021 to 0.262	0.051	-7.502	0.084	0.981	0.806	< 0.001	< 0.001
	Density	3	4	-0.039	-1.225 to 1.148	0.270	-0.143	0.900	11.108	0.011	75.164	0.406
fMRI	Clustering coefficient	29	46	-0.351	-0.577 to -0.124	0.111	-3.175	0.004	262.520	< 0.001	89.880	0.577
	Local efficiency	24	39	-0.217	-0.450 to 0.016	0.113	-1.931	0.066	206.980	< 0.001	84.569	0.523
	Characteristic path length	28	44	-0.001	-0.195 to 0.194	0.095	-0.005	0.996	219.115	< 0.001	83.297	0.451
	Global efficiency	26	46	0.026	-0.240 to 0.291	0.129	0.198	0.844	311.667	< 0.001	90.273	0.643
	Normalized clustering coefficient	12	18	0.075	-0.095 to 0.245	0.076	0.983	0.349	21.292	0.213	30.154	0.163
	Normalized characteristic path length	12	18	0.005	-0.240 to 0.249	0.111	0.045	0.965	44.569	< 0.001	62.505	0.317
	Small-worldness	14	22	0.029	-0.225 to 0.284	0.117	0.250	0.806	78.683	< 0.001	77.750	0.398
	Modularity	6	9	0.217	0.023 to 0.410	0.070	3.087	0.036	46.551	< 0.001	79.331	0.248
	Assortativity	4	6	-0.179	-1.096 to 0.739	0.266	-0.671	0.556	27.613	< 0.001	84.450	0.656
	Strength	6	10	-0.425	-0.881 to 0.030	0.175	-2.425	0.062	26.822	0.001	71.139	0.374
sMRI	Clustering coefficient	4	5	0.026	-0.788 to 0.841	0.256	0.103	0.925	19.144	0.001	79.262	0.467
	Local efficiency	4	5	0.096	-0.799 to 0.991	0.280	0.343	0.754	24.057	< 0.001	83.099	0.532
	Characteristic path length	3	4	-0.375	-0.902 to 0.152	0.119	-3.164	0.091	3.078	0.380	3.115	0.045
	Global efficiency	3	4	0.356	-0.409 to 1.122	0.170	2.102	0.177	6.910	0.075	55.456	0.280
	Normalized clustering coefficient	2	3	0.424	-1.916 to 2.764	0.184	2.303	0.261	2.198	0.333	29.375	0.167
	Normalized characteristic path length	3	4	-0.367	-0.790 to 0.055	0.098	-3.764	0.064	2.251	0.522	< 0.001	< 0.001
	Small-worldness	2	3	0.467	-2.170 to 3.103	0.208	2.248	0.266	2.341	0.310	40.484	0.214
EEG	Clustering coefficient	6	7	0.349	-0.581 to 1.279	0.361	0.968	0.378	41.118	< 0.001	87.688	0.843
	Local efficiency	3	4	0.123	-1.798 to 2.044	0.446	0.275	0.809	19.532	< 0.001	85.209	0.724
	Characteristic path length	6	8	0.158	-0.885 to 1.202	0.405	0.391	0.712	60.681	< 0.001	88.273	0.964
	Global efficiency	5	6	0.082	-0.947 to 1.112	0.365	0.226	0.833	28.000	< 0.001	84.286	0.783
Cognition in fMRI
Normal cognition	Clustering coefficient	4	4	-0.623	-1.015 to -0.232	0.121	-5.132	0.015	1.721	0.632	< 0.001	< 0.001
	Local efficiency	3	3	-0.782	-1.080 to -0.484	0.068	-11.450	0.008	0.318	0.853	< 0.001	< 0.001
	Characteristic path length	5	5	0.006	-0.858 to 0.870	0.311	0.019	0.986	23.304	< 0.001	83.716	0.635
	Global efficiency	6	6	0.012	-0.673 to 0.697	0.266	0.046	0.965	26.132	< 0.001	81.671	0.589
Cognitive impairment	Clustering coefficient	4	6	-0.993	-1.962 to -0.024	0.294	-3.374	0.047	20.022	0.001	76.703	0.544
	Local efficiency	3	5	-1.293	-2.630 to 0.044	0.304	-4.258	0.053	10.665	0.031	67.007	0.453
	Characteristic path length	5	8	-0.307	-1.296 to 0.682	0.356	-0.863	0.437	41.746	< 0.001	87.980	0.761
	Global efficiency	6	11	0.475	-0.555 to 1.505	0.400	1.187	0.289	95.823	< 0.001	91.433	0.965
Cognitive impairment vs normal cognition	Clustering coefficient	5	7	-0.222	-1.151 to 0.707	0.332	-0.670	0.540	76.511	< 0.001	89.797	0.701
	Local efficiency	3	5	-0.508	-1.024 to 0.007	0.084	-6.051	0.051	5.242	0.263	21.553	0.152
	Characteristic path length	6	9	-0.197	-0.616 to 0.221	0.158	-1.247	0.272	26.047	0.001	65.712	0.311
	Global efficiency	7	12	0.268	-0.299 to 0.836	0.231	1.160	0.291	77.623	< 0.001	84.883	0.560

dMRI diffusion MRI, fMRI functional MRI, sMRI structural MRI, EEG electroencephalography. The table shows results with sample size ≥3; findings for sample size <3 are presented in Table S13.

Taking the modalities in turn, dMRI finds a breakdown in both local specialization and global integration in PD vs HC. Specifically, clustering coefficient (g = -0.328, P = 0.002), local efficiency (g = -0.272, P = 0.007) and global efficiency (g = -0.445, P < 0.001) were significantly lower in PD, while characteristic path length (g = 0.396, P = 0.001) and normalized clustering coefficient (g = 0.245, P = 0.026) were significantly higher in PD.

By fMRI, PD mainly shows lower network segregation, while measures of network integration remain largely unchanged. Specifically, PD vs HC showed a significantly lower clustering coefficient (g = -0.351, P = 0.004) and a trend falling short of formal statistical significance toward lower local efficiency (g = -0.217, P = 0.066), though global efficiency (P = 0.844) and characteristic path length (P = 0.996) showed no significant differences. Modularity was significantly higher in PD (g = 0.217, P = 0.036).

By sMRI and EEG there were no significant case-control differences in GTA metrics. For meta-analyses including three studies or fewer (e.g., PET), differences in some GTA metrics between groups should be interpreted with caution pending further validation: other main results are presented in Table S13.

Quality assessment, publication bias, and sensitivity analyses in case-control meta-analysis

The quality assessment by the modified Newcastle-Ottawa Scale provided in Table S14, ranged between low (QA ≤ 4; k = 5), moderate and high (QA ≥ 5; k = 75) with average scores of 6.9 for dMRI studies, 6.6 for fMRI, 7.5 for sMRI, and 5.0 for EEG. Overall, the studies were of moderate to high quality, although lower in EEG studies vs other modalities. Egger’s regression tests indicated no significant publication bias in most of the meta-analyses (Table S15). Visual inspection of funnel plots revealed no publication bias overall, although a few small-sample effects differed substantially from the pooled estimates (Figs. S16–22). Sensitivity analyses were performed by removing outliers/influential effect sizes. A total of 15 meta-analyses were found to contain influential effect sizes, many of which corresponded to points that deviated from the funnel plot symmetry (Table S16). After excluding these influential effect sizes, the pooled results remained largely unchanged, except for modularity in the fMRI studies, where the significant effect in the primary meta-analysis became non-significant, suggesting that the statistical significance had been primarily driven by the influential effect size. Furthermore, sensitivity analyses confirmed the robustness of the results after excluding task-based studies, studies only reporting multiple sparsity thresholds, and under different data extraction strategies, such as using binary instead of weighted data, or functional instead of anatomical atlases (Table S17).

Meta-regression with continuous and categorical variables in case-control meta-analysis

Figure 3A–E shows significant meta-regression results with continuous independent variables. In dMRI studies, local efficiency was positively associated with levodopa-equivalent daily dose (LEDD) (B = 0.002, P = 0.041), and small-worldness increased with the number of nodes (B = 0.017, P = 0.034). In fMRI studies, global efficiency was negatively correlated with education level (B = -0.117, P = 0.015), network strength with male gender (B = -0.043, P = 0.039), and normalized characteristic path length with Unified PD Rating Scale III (UPDRS-III) scores (B = -0.031, P = 0.040). Table S18 gives detailed results of the meta-regression with continuous variables. None of these results survived multiple comparison corrections at false discovery rate (FDR) < 0.05.

Fig. 3. Meta-regression of clinical and methodological factors.

Significant results from meta-regressions with continuous variables (A–E) and categorical factors (F–I). Each colored dot represents an effect size from an individual study. In the continuous variable plots (A–E), the black line indicates the meta-regression line, with standard error band in gray. In the categorical factor plots (F–I), black circles and horizontal error bars represent the pooled effect sizes and standard errors. Abbreviations: dMRI diffusion MRI, fMRI functional MRI, LEDD levodopa-equivalent daily dose, UPDRS-III Unified Parkinson’s Disease Rating Scale III scores.

Figure 3F–I shows significant moderating effects of categorical variables reflecting scanning parameters, brain parcellations, thresholding approaches, and medication status. These were significant only in fMRI studies: studies using 3 T vs 1.5 T scanners showed lower clustering coefficients (B = -0.803, P = 0.010); studies using functional parcellation schemes vs macroanatomical atlases showed lower clustering coefficients, approaching marginal statistical significance (B = 0.501, P = 0.052); studies using proportional thresholding vs non-proportional thresholds showed lower local efficiency (B = -0.640, P = 0.033); and medication status showed a moderate effect on normalized characteristic path length (B = 0.589, P = 0.022). Table S19 gives detailed results of the meta-regression with categorical variables. However, most correlations except for that between scanning parameters and clustering coefficients did not survive multiple comparison corrections at FDR < 0.05.

Comparisons of cognitive normal and impairment in case-control meta-analysis

Compared to HC, both cognitively normal and cognitively impaired PD patients showed reduced network segregation, as reflected by lower clustering coefficient (Cognitive normal: g = -0.623, P = 0.015; Cognitive impairment: g = -0.993, P = 0.047) and local efficiency (Cognitive normal: g = -0.782, P = 0.008; Cognitive impairment: g = -1.293, P = 0.053) (Table 2). Patients with cognitive impairment had significantly lower local efficiency vs those with normal cognition (g = -0.508, P = 0.051). In contrast, there were no significant differences in network integration (characteristic path length and global efficiency) between PD subtypes (with and without cognitive impairment) or between each subtype and HC.

Systematic review of disease progression and treatment response

Although findings varied across imaging modalities and clinical subtypes, several consistent patterns emerged (Table S20). Patients with more severe or rapidly progressing forms of PD, including those at high risk of developing dementia or exhibiting diffuse-malignant subtype, showed lower global efficiency and higher characteristic path length at baseline, reflecting impaired network integration^14,22. These baseline alterations were also associated with longitudinal worsening in motor and cognitive outcomes, including increased global composite outcome, increased postural instability gait difficulty score, and progressive cognitive decline^14,22,23. Notably, associations between graph metrics and cognitive trajectories might be moderated by disease severity, as significant correlations were observed mainly in patients with worse clinical profiles, whereas those with milder symptoms showed no such relationships.

Several studies examined the relationship between global brain network topology and treatment response in PD, including pharmacological interventions, cognitive training, and deep brain stimulation (DBS) (Table S21). Patients who responded better to dopaminergic therapy or who exhibited levodopa-induced dyskinesias tended to show higher global or local efficiency and lower characteristic path length, suggesting relatively preserved or compensatory network integration^24,25. In the context of DBS, progressive changes in global network metrics were observed in patients selected as surgical candidates, whereas those not considered suitable for DBS showed no such progression, suggesting that topological features may hold potential in distinguishing patients likely to benefit from surgical intervention²⁶. However, findings across treatment strategies and patient subgroups remain inconsistent. While some studies reported network reorganization associated with stimulation or medication state, others found no significant alterations in global metrics. Non-dopaminergic interventions, including atomoxetine, citalopram, and multi-domain cognitive training, also failed to produce consistent changes in topological properties. Overall, although certain network metrics were associated with treatment response in specific contexts, no uniform pattern emerged across all therapeutic approaches or populations.

Primary results in diagnostic meta-analysis

A total of 10 articles met the criteria for inclusion in the diagnostic meta-analysis. Their detailed characteristics are shown in Table S22. A total of 43 contingency tables were included from studies on the classification of PD vs HC using GTA metrics. Table 3 summarizes the pooled performance estimates of GTA metrics in the diagnosis of PD.

Table 3.

Summary estimates and meta-regression of pooled performance of GTA metrics in diagnosing PD from HC

Parameter	No. of tables	AUC (95% CI)	Sensitivity		P	Specificity		P	LR + (95% CI)	LR- (95% CI)
			SE (95% CI)	I2 (95% CI)		SP (95% CI)	I2 (95% CI)
Overall	43	0.86 (0.83–0.89)	0.77 (0.70–0.82)	89.15 (86.62–91.67)		0.81 (0.76–0.86)	88.86 (88.25–91.47)		4.1 (2.9–5.8)	0.29 (0.21–0.39)
Algorithm					0.60			0.36
AI algorithm	35	0.88 (0.85–0.90)	0.78 (0.71–0.84)	89.41 (86.69–92.13)		0.83 (0.77–0.88)	89.74 (87.13–92.35)		4.7 (3.1–7.1)	0.26 (0.18–0.38)
Without AI algorithm	8	0.77 (0.73–0.81)	0.68 (0.56–0.78)	74.96 (57.41–92.51)		0.73 (0.64–0.81)	55.21 (19.69–90.73)		2.6 (1.9–3.4)	0.44 (0.32–0.60)
Imaging technique					0.32			0.52
Functional imaging	33	0.91 (0.88–0.93)	0.82 (0.76–0.87)	88.27 (85.08–91.47)		0.86 (0.81–0.89)	83.83 (79.00–88.86)		5.7 (4.1–8.0)	0.21 (0.15–0.30)
Structural imaging	10	0.58 (0.53–0.62)	0.54 (0.44–0.63)	52.52 (18.44–86.60)		0.60 (0.46–0.72)	75.91 (61.05–90.78)		1.3 (0.8–2.1)	0.77 (0.52–1.14)
Feature selection					< 0.01			< 0.01
GTA Only	36	0.83 (0.80–0.86)	0.74 (0.66–0.81)	80.89 (75.17–86.61)		0.79 (0.71–0.85)	80.63 (74.81–86.45)		3.5 (2.4–5.1)	0.33 (0.23–0.47)
GTA Plus	7	0.95 (0.92–0.96)	0.86 (0.80–0.91)	97.53 (96.58–98.49)		0.91 (0.88–0.93)	93.17 (89.55–96.78)		9.4 (7.0–12.5)	0.15 (0.10–0.23)
Threshold					0.02			< 0.01
Non–proportion	20	0.89 (0.86–0.92)	0.77 (0.68–0.85)	90.97 (88.02–93.92)		0.86 (0.80–0.90)	83.69 (77.38–90.00)		5.6 (3.5–8.7)	0.26 (0.17–0.40)
Proportion	22	0.81 (0.77–0.84)	0.75 (0.65–0.83)	84.01 (78.16–89.87)		0.74 (0.65–0.82)	82.55 (75.99–89.10)		2.9 (1.9–4.6)	0.34 (0.21–0.54)

PD Parkinson’s Disease, HC healthy control, SE sensitivity, SP specificity, AUC area under curve, LR + positive likelihood ratio, LR −  negative likelihood ratio, CI confidence interval, GTA graph theory analysis, AI artificial intelligence, GTA Only only GTA metrics used as classification features, GTA Plus both GTA metrics and other network metrics used as classification features.

P value for heterogeneity between subgroups with meta-regression analysis.

In the bivariate meta-analysis, the pooled sensitivity (SE), specificity (SP) and area under the curve (AUC) were 0.77 (95% CI: 0.70-0.82), 0.81 (95% CI: 0.76-0.86) and 0.86 (95% CI: 0.83-0.89) respectively (Fig. 4 and Fig. S23). High heterogeneity was observed, with an I² of 89.15% (95% CI: 86.62–91.67%) for SE and 88.86% (95% CI: 88.25–91.47%) for SP. No significant publication bias was observed (Fig. S24). To address the dependency among multiple outcomes within studies, a three-level meta-analysis was performed, yielding a pooled diagnostic odds ratio (DOR) of 16.4 (95% CI: 6.1–44.8, prediction interval: 0.40–666.5, P < 0.001). Significant heterogeneity was identified (Q (42) = 524.409, P < 0.0001), including significant between-study variance (τ² = 1.530, I² = 44.202%, n = 10 studies) and within-study variance (τ² = 1.081, I² = 51.751%, k = 43 tables).

Fig. 4. Summary receiver operating characteristic (SROC) curves of studies to classify Parkinson’s disease vs healthy controls.

SROC curves of studies included in the meta-analysis (10 studies with 43 tables). Red dots represent data points from contingency tables. The blue line shows the SROC curve, with the blue diamond marking the Summary Operating Point (pooled sensitivity and specificity). The gray area represents the 95% Prediction Contour, and the orange area represents the 95% Confidence Contour. Abbreviations: SROC summary receiver operating characteristic, SENS summary sensitivity, SPEC summary specificity, AUC area under the curve.

Meta-regression/subgroup analysis for SE, SP and AUC in diagnostic meta-analysis

In meta-regression analysis, studies using GTA Plus (i.e., both GTA metrics and other network metrics) vs GTA Only metrics for distinguishing PD from HC showed higher SE and SP (P < 0.01), and studies using non-proportional thresholds vs proportional thresholds showed higher SE and SP (P < 0.05) (Table 3). The forest plots, SROC curves, and publication bias regression plots for subgroup analysis are shown in Figs. S25–28, Figs. S29–32, and Figs. S33, respectively. Below are the specific details of the subgroup analysis results:

Algorithm type

35 contingency tables from eight studies utilized the AI algorithm, while 8 tables from 2 studies employed other algorithms. The pooled SE for AI algorithm was 0.78 (95% CI: 0.71–0.84), and for other algorithms was 0.68 (95% CI: 0.56–0.78); pooled SP was 0.83 (95% CI: 0.77–0.88) for AI algorithm and 0.73 (95% CI: 0.64–0.81) for other algorithms. AUC was 0.88 (95% CI: 0.85–0.90) for AI algorithm and 0.77 (95% CI: 0.73–0.81) for other algorithms. No significant publication bias was observed.

Imaging technique

33 contingency tables from 8 studies used functional imaging, while 10 tables from 2 studies used structural imaging. The pooled SE for functional imaging was 0.82 (95% CI: 0.76–0.87), and for structural imaging was 0.54 (95% CI: 0.44–0.63); pooled SP was 0.86 (95% CI: 0.81–0.89) for functional imaging and 0.60 (95% CI: 0.46–0.72) for structural imaging. AUC was 0.91 (95% CI: 0.88–0.93) for functional imaging and 0.58 (95% CI: 0.53–0.62) for structural imaging. No significant publication bias was observed.

Feature selection

36 contingency tables from 8 studies used only GTA metrics as input features (GTA Only), while 7 contingency tables from 3 studies used GTA metrics combined with other imaging metrics (GTA Plus). The pooled SE for GTA Only was 0.74 (95% CI: 0.66–0.81), and for GTA Plus was 0.86 (95% CI: 0.80–0.91); pooled SP was 0.79 (95% CI: 0.71–0.85) for GTA Only and 0.91 (95% CI: 0.88–0.93) for GTA Plus. AUC was 0.83 (95% CI: 0.80–0.86) for GTA Only and 0.95 (95% CI: 0.92–0.96) for GTA Plus. No significant publication bias was observed.

Threshold

20 contingency tables from 4 studies used non–proportion threshold, while 22 contingency tables from 5 studies used proportion threshold. The pooled SE for non–proportion threshold was 0.77 (95% CI: 0.68–0.85), and for proportion threshold was 0.75 (95% CI: 0.65–0.83); pooled SP was 0.86 (95% CI: 0.80–0.90) for non–proportion threshold and 0.74 (95% CI: 0.65–0.82) for proportion threshold. AUC was 0.89 (95% CI: 0.86–0.92) for non–proportion threshold and 0.81 (95% CI: 0.77–0.84) for proportion threshold. Significant publication bias was observed only in the proportion subgroup (P < 0.01).

Three–level moderator analysis for DOR in diagnostic meta–analysis

The between–study heterogeneity in DOR values was primarily influenced by feature selection (Table S23). Significant moderator effects were observed in the feature selection moderator analysis (QM (1) = 6.763, p = 0.009; between–study variance τ² = 1.385 n = 10; within–study variance τ² = 1.239, k = 43). Using mixed network metrics as input features (β = 1.997, P = 0.009) significantly improved the DOR compared to using only GTA metrics.

No significant moderator effects were observed in the algorithm selection moderator analysis (QM (1) = 1.198, P = 0.274; between–study variance τ² = 1.848 n = 10; within-study variance τ² = 1.515, k = 43) and threshold moderator analysis (QM (1) = 0.082, P = 0.775; between-study variance τ² = 1.861, n = 9; within-study variance τ² = 1.515, k = 42).

Controlling for feature selection, no significant moderator effects were observed in the imaging techniques analysis (QM (1) = 3.762, P = 0.052; between-study variance τ² = 0.805, n = 8; within-study variance τ² = 1.635, k = 36).

Exploratory result of highest and lowest performance in diagnostic meta-analysis

In exploratory analysis, when selecting the contingency table with the best or worst performance, the pooled SE, SP, and AUC were 0.88 (95% CI: 0.76–0.94) vs 0.73 (95% CI: 0.58–0.84), 0.92 (95% CI: 0.89–0.95) vs 0.79 (95% CI: 0.66–0.88), 0.96 (95% CI: 0.93–0.97) vs 0.83 (95% CI: 0.79–0.86) (Figs. S34, 35). No significant publication bias was observed (Fig. S36).

Quality assessment in diagnostic meta-analysis

The quality of the included studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies-AI (QUADAS-AI) tool. The detailed results of the assessment are shown in Table S24. Most studies had a high risk of bias in patient selection and the index test because they did not use samples from public databases and lacked adequate external validation.

Discussion

We found that network segregation and integration were consistently lower in PD vs HC in dMRI studies, while fMRI studies showed selective impairments in segregation only. Meta-regressions revealed that GTA metrics were influenced by both clinical/demographic factors, such as gender and medication status, and methodological factors, including scanning parameters and network construction approaches. Network segregation was lower in both cognitively normal and cognitively impaired PD patients, with more pronounced abnormality in the latter. Our systematic review further highlighted associations between GTA metrics and disease progression and treatment response, suggesting potential for use in prognosis and stratification. Finally, results from the diagnostic meta-analysis indicated that topological properties may serve as reliable biomarkers in distinguishing PD vs HC.

Our results suggest that structural networks are more severely disrupted than functional networks in PD. Notably, this pattern is not unique to PD but has also been reported in other neurodegenerative and neuropsychiatric disorders. For example, Liu and colleagues demonstrated that patients with Alzheimer’s disease (AD) show lower segregation and integration in structural networks, whereas functional networks showed only lower segregation; also, in preclinical AD, structural disorganization is already evident while functional topology remains relatively preserved²⁷. Similarly, patients with focal epilepsy show lower structural integration but largely intact functional topology²⁸. Patients with schizophrenia show consistently disrupted structural segregation and integration, with functional networks showing minimal alteration²⁹. The convergence across these conditions suggests that structural networks are generally more detectably vulnerable to pathological insults than functional networks.

This difference likely reflects the distinct neurobiological signals captured by each modality: dMRI assesses the physical integrity of white matter tracts; by contrast fMRI measures statistical dependencies in dynamic blood-oxygen-level-dependent signals, which do not represent physical connections but are a good proxy of neuronal activation³⁰. While the structural network provides a fundamental scaffold that constrains functional interactions, their relationship is neither linear nor regionally uniform^31,32. Functional connectivity can emerge through indirect, polysynaptic or dynamically modulated routes, and may persist even in the absence of direct anatomical pathways, as evident in studies of callosal agenesis³³ and split-brain patients³⁴. It is therefore clear that functional networks can flexibly reorganize through indirect, dynamic or context-dependent interactions, whereas structural networks are inherently constrained by the relatively fixed architecture of white-matter tracts. Thus, the larger alterations observed in dMRI-based networks likely reflect the irreversible nature of structural degeneration (e.g., axonal loss, demyelination), while fMRI-based connectivity may be relatively preserved by compensatory reconfiguration or the recruitment of alternative polysynaptic pathways that maintain large-scale functional communication, despite underlying anatomical disruption. Our results may suggest that compensatory changes occur primarily through long-range connections, rather than mechanisms of network segregation. This is consistent with the findings of Shine and colleagues³⁵, who investigated network topology in PD patients both on and off dopamine replacement therapy. They reported greater network integration in the off-state compared with the on-state, with higher integration associated with preserved motor function and positively linked to cognitive and brain reserve.

In our analyses various disorder factors—including dopaminergic medication, cognitive status, gender and others—have a significant influence brain network topology in PD (Fig. 3). However, as very few of these relationships survive multiple comparison corrections, and given the inherent limitations of meta-regression, such as inability to control for multiple confounding factors, they should be taken as exploratory and hypothesis-generating. With that caveat, the correlation between higher LEDD and higher local efficiency (Fig. 3A) may be worth noting, for the suggestion it offers that dopaminergic therapy may partially restore network segregation; this would also be consistent with our systematic review on treatment response. Dopaminergic depletion plays a central role in the pathogenesis and progression of PD, reflected in recently proposed biological staging frameworks³⁶. The lower functional segregation we see in PD likely reflects the widespread effects of dopamine loss on large-scale brain organization. The idea that dopamine supplementation might facilitate partial recovery of modular network structure is perhaps compatible with the recent demonstration that levodopa drives PD networks toward a more segregated and modular configuration³⁷.

Male patients showed greater reductions in network strength than females, aligning with the higher prevalence of PD and faster cognition decline in males^38–40. Our findings suggest that topological disorganization of brain networks may represent an inherent characteristic of PD, regardless of cognitive status, but that the severity of network disruption increases with the emergence of cognitive impairment. Lower clustering coefficient and local efficiency in the cognitively impaired subgroup may reflect a breakdown of local communication and reduced capacity for functional specialization, consistent with other evidence linking segregation deficits to executive dysfunction and AD^27,41. These results support the growing view that cognitive impairment in PD is underpinned by network-level disorganization, and emphasize the importance of stratifying patients by cognitive status in future network analyses. While UPDRS-III scores and education were related to specific network metrics, these associations did not correspond to significant case–control differences, indicating limited robustness of these effects. All the studies included in the current meta-analysis were of cross-sectional design, which cannot explicitly elucidate the causal relationships between brain network alterations and disorder; longitudinal follow-up studies would help clarify these issues.

Methodologically, several acquisition and analytic parameters significantly shaped the reported outcomes. fMRI studies using 3 T scanners detected much larger reductions in clustering coefficients than 1.5 T scanners, highlighting the importance of signal-to-noise ratio in resolving subtle connectivity alterations. This suggests that higher field MRI may reveal larger topological effect sizes^42,43. Functional parcellations showed larger reductions in clustering coefficient than those based on macroanatomical templates, suggesting that functional atlases may be more sensitive to functional connectivity alterations. However, consensus on the optimal parcellation has yet to be established^44,45. Studies applying proportional thresholding reported lower local efficiency than those using other thresholds, suggesting that network sparsity choices may affect the sensitivity to group differences. These findings highlight the need for better standardization in imaging protocols, network construction and reporting practices to enhance reproducibility and interpretability in future GTA-based studies of PD³⁰.

Current evidence in PD suggests that alterations in global brain network topology, especially lower network integration, may not only reflect current network disruption but might also predict future motor and cognitive decline, supporting their use as progression markers in clinical trials^14,22,23,46. However, inconsistencies across studies—due to varied imaging modalities, analytic methods, small samples, and patient heterogeneity—limit generalizability. The heterogeneous findings across studies investigating treatment-related changes highlight the complexity of linking therapeutic interventions to network-level mechanisms in PD. Given the clinical heterogeneity and variable treatment responses, stratifying PD patients by neuroimaging markers could help identify those likely to benefit from specific therapies^3,47. Preliminary evidence from Albano and colleagues suggests that network topology may support such stratification, as global topological alterations over time are observed in DBS candidates but not in non-candidates²⁶, pointing to its potential utility in identifying DBS responders. More broadly, mixed results across intervention types, particularly in studies involving DBS by Bočková and colleagues⁴⁸, pharmacological agents^49,50, and cognitive training⁵¹, may be attributed to methodological differences, variability in patient selection, and inconsistent response definitions.

Our meta-analysis demonstrated that the GTA metric exhibits promising but preliminary diagnostic potential for PD, with a pooled AUC of 0.86 (95% CI: 0.83–0.89) and a pooled DOR of 16.36 (95% CI: 6.07–44.8). However, this estimate should be interpreted cautiously given the limited external validation and methodological heterogeneity across studies. Significant heterogeneity was primarily attributable to feature selection, as “GTA Plus” studies (combining GTA metrics with other network features) showed significantly higher both AUC and DOR vs “GTA Only”. This supports the integration of multimodal features for enhanced PD diagnosis. Although the meta-regression analysis showed that thresholds significantly affect SE and SP, the same impact was not observed for DOR. While the meta-regression and moderator analysis did not reveal statistically significant differences, studies using AI algorithms showed higher DOR, AUC, SE and SP vs non-AI studies, highlighting the potential advantages of automated AI-based diagnostic methods for PD. Future diagnostic models should leverage collaborative human-AI frameworks to optimize accuracy by combining imaging with non-imaging data (e.g., demographics, clinical history)^52,53. Given the current reliance on internal cross-validation, external validation using stratified datasets (training, tuning, validation sets) will be essential.

This study has several limitations. First, although our meta-analysis integrates multimodal evidence on global brain network alterations in PD, we observed substantial study-level heterogeneity in some analyses (I² > 75%), particularly in fMRI studies. We conducted a series of meta-regressions examining potential sources of this heterogeneity, finding it partly explained by disorder factors (e.g., medication status, symptom severity, gender, years of education) and methodological differences (e.g., MRI field strength, brain parcellation atlas, thresholding approaches). Egger’s regression tests found that publication bias was largely absent in the main analyses. However, in view of the heterogeneity of results, the constraints placed by incomplete reporting on our ability to control for confounding factors, and the resulting low statistical power, these findings should be interpreted conservatively. Second, our focus on global topological metrics, while facilitating comparison across studies, limits spatial specificity; local metrics (e.g., nodal efficiency, hub disruption) could provide more detailed insights, but quantitative synthesis is hindered by variability in node definitions and selective reporting. Third, the limited number of studies using certain modalities, such as MEG or PET, prevented robust conclusions for these techniques, and may have constrained the multimodal perspective. Fourth, the study was not pre-registered, which arguably limits transparency and increases potential bias; however, to enhance reproducibility we have provided extensive supplementary materials including detailed search strategies, data extraction, and analysis scripts. Finally, quantitative analysis of symptom subtypes was restricted to cognition-related changes in fMRI due to data scarcity; other clinical domains, such as motor or psychiatric symptoms, could not be assessed.

Our findings suggest that PD involves distributed alterations of brain network topology, particularly lower segregation and impaired integration. These results have implications for the neuroimaging and PD communities. Methodologically, the sensitivity of GTA measures to acquisition and analytic choices underscores the need for harmonized protocols and transparent reporting of atlas selection, preprocessing and thresholding strategies. Greater openness in sharing analytic code and methodological details could further improve reproducibility and allow more consistent meta-analytic synthesis. Clinically, GTA metrics show potential as imaging biomarkers, particularly when integrated with multimodal imaging or clinical features, but establishing their utility will require careful external validation across sites and patient subtypes. Future research should prioritize larger, longitudinal and multi-center studies, more systematic analyses of nodal and hub-level properties, and the integration of structural, functional and molecular modalities. This will strengthen the translational value of brain network approaches in understanding disease mechanisms and in developing reliable imaging-based tools for diagnosis and stratification in PD.

This systematic review and meta-analysis provide robust evidence for global brain network topology alterations in PD, particularly reduced network segregation as revealed by dMRI and fMRI, which were more pronounced in patients with cognitive impairment. Additional impairments of network integration in PD were revealed by dMRI, which appeared to correlate with disease progression. These alterations were further influenced by clinical and methodological factors, including medication status, gender, and network construction strategies. The diagnostic meta-analysis demonstrated potential utility for distinguishing PD patients from HC. Together, these findings highlight the potential of GTA measures as biomarkers for disease characterization, prognosis, and patient stratification in PD, while emphasizing the need for methodological standardization and prospective validation in future research.

Methods

This systematic review and meta-analysis were conducted according to the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)⁵⁴.

Search strategy, study selection and graph theoretical metrics interpretations

We searched PubMed, Embase and Web of Science up to March 2025 for articles examining the topological properties of the whole brain network in PD. Table S25 provides details of the search terms. Additional articles were identified from the reference lists of eligible articles. Screening and inclusion were conducted independently by two reviewers (C.Z. and W.L.), discrepancies being resolved by discussion or by consulting a third senior investigator (X.S.). Articles included fulfilled the following criteria: 1) original research investigating network alterations at the whole-brain level applying GTA methods to multimodal neuroimaging; 2) reporting global topological parameters; 3) comparing PD with HC; and 4) reporting sufficient statistics for calculation of effect sizes. Studies eligible for diagnostic meta-analysis reported GTA metrics for PD diagnosis and provided outcomes, such as SE and SP, used to calculate 2 × 2 contingency tables.

The main GTA metrics analyzed in this study were measures of network segregation, integration and small-worldness. Network segregation reflects the brain’s capacity for specialized processing within densely interconnected groups of regions, known as modules or clusters; it is commonly quantified by metrics, such as the clustering coefficient, local efficiency, transitivity and modularity. Network integration, indexed by the characteristic path length and global efficiency, reflects the capacity for efficient communication across distant brain regions. Small-worldness describes the coexistence of high local clustering and short global path length, a hallmark of efficient network organization. Less frequently reported metrics include assortativity, strength, degree and density, which characterize complementary aspects of brain network organization: assortativity reflects the tendency of nodes to connect with other nodes of similar degree, which has implications for network resilience and vulnerability to targeted attacks; network degree quantifies the mean number of connections per node, and network strength represents the average connection weight across all nodes, describing the overall level of connectivity in the brain network; density reflects the proportion of existing edges relative to all possible connections in the network, indicating how densely connected it is. These metrics are widely used to assess topological alterations in structural and functional brain networks⁵⁵.

Data extraction and quality assessment in case-control meta-analysis

The data were independently extracted by two reviewers (C.Z. and W.L.), and any discrepancies were resolved by consulting a third reviewer (X.S.), following the procedure used in the diagnostic meta-analysis. Extracted data consisted of first author, publication year, demographic data of patients with PD and HC, clinical manifestations (disease duration, medication status, UPDRS-III scores, Hoehn and Yahr stage (H&Y), LEDD, and Mini-Mental State Examination (MMSE) scores); methodology of data acquisition, data preprocessing and network construction (including MRI field strength, parcellation procedure, edge definition, network framework and threshold); and global graph theoretical parameter outcomes. When studies reported Movement Disorder Society-UPDRS scores, we converted them to UPDRS-III scores using established formulas that require the H&Y stage⁵⁶; where the H&Y stage was not provided, we applied a simplified conversion method⁵⁷. If the age at onset was not reported, we estimated it by subtracting the disease duration from the age. When important data was not reported, the corresponding authors were contacted via email. When outcomes were reported only as graphs, data were manually extracted using WebPlotDigitizer software⁵⁸. When mean and standard deviation for mean differences were not reported, they were approximated from the sample size, median and interquartile range⁵⁹. Datasets with < 10 participants per group were excluded to minimize bias, as small studies with skewed outcome data can produce misleading findings⁶⁰.

When outcomes were reported in patients with both off- and on-state medication status, we used only off-state^49,61. When results were reported for PD subgroups and for combined PD patients, we used the latter²³. In longitudinal studies, we used only baseline results. If both low-order (temporal synchronization) and high-order functional connectivity were reported, we used the former for their more widespread application. Some studies provided data from different methods to validate result robustness; since the outcomes were similar, primary reported data were extracted. When both binary and weighted networks were reported, we used only weighted data⁶². We conducted sensitivity analyses to evaluate the impact of data selection. For studies offering multiple comparisons across different EEG frequency bands, we extracted theta band data, aligning with the majority of included studies that report metrics in this band, and consistent with a meta-analysis on focal epilepsy brain networks²⁸. When multiple studies used overlapping cohorts (e.g. PPMI), only the dataset with the largest sample size within each modality was used for the quantitative meta-analysis. For an MEG study reporting baseline data in delta and alpha2 bands, we used the average of these⁶³. We assessed study quality and risk of bias of included studies using a modified version of the Newcastle-Ottawa Scale (Table S14).

Data analysis in case-control meta-analysis

The effect sizes of each metric were calculated using Hedges’ small sample correction. Many studies reported comparisons of different PD subgroups vs the same HC group, resulting in the statistical dependency known as the unit-of-analysis error⁶⁰. To address this, we used multilevel random-effects meta-analysis with robust variance estimation^64–66, encompassing three sources of variance: sampling variance (level 1), variance between effect sizes from the same study (level 2), and variance between studies (level 3). Unlike traditional meta-analytic approaches, the dependent effect sizes are nested within studies (level 2) before they are pooled across studies (level 3); this enhances statistical power and maximizes the information obtained¹⁷. Where a few studies reported multiple proportional thresholds for GTA metrics, we employed a fixed-effect model to pool these effect sizes, ensuring that each dataset contributed a single, consolidated effect size⁶⁷. We assessed the heterogeneity of effect sizes using I² statistics: I² shows the fraction of total variance in true effects for each level. All statistical analyses were conducted using the “metafor” package in R (version 4.2.1)⁶⁸.

Graph metrics were analyzed using meta-regression for categorical factors, such as medication status (off-state/drug naïve versus on-state), data acquisition, preprocessing, and network construction methods, and continuous-variable meta-regression to evaluate moderating effects of mean patient age, percentage of males, disease duration, years of education, LEDD, UPDRS-III scores, MMSE scores, and H&Y stage. Categorical regression was performed only when each subgroup contained >3 effect sizes, and continuous regression only when >10 effect sizes were available⁶⁷. To address the issue of multiple comparisons, we adopted an FDR correction method at a significance value of 0.05. Publication bias was assessed using funnel plots and modified Egger’s regression test⁶⁹. To evaluate the robustness of our results a series of sensitivity analyses were conducted related to outliers/influential effect sizes⁷⁰, removing task studies, removing multiple sparsity studies, and differing extracted data selection.

Data extraction and quality assessment in diagnostic meta-analysis

In contrast to case-control meta-analysis comparing network metrics between PD patients and HC, the diagnostic meta-analysis specifically evaluates the ability of GTA metrics to distinguish PD patients from HC. The diagnostic performance of GTA metrics was assessed by a meta-analysis of studies incorporating contingency tables. The collected information included the sample size, imaging techniques, algorithms, feature selection and diagnostic performance data. Binary diagnostic accuracy data were extracted, and contingency tables were constructed. Reported diagnostic accuracy data, including SE, SP, AUC, true positive (TP), false positive (FP), true negative (TN), and false negative (FN), were directly extracted into contingency tables and used to calculate SE, SP and accuracy if not directly reported. If the samples of multiple studies were derived from the same open database, the study with the largest sample size was included. If a study used only connectivity strength (instead of conventional GTA metrics) as feature inputs, it was excluded. If a study used different algorithms, imaging techniques or features for classification, each contingency table was analyzed independently^53,71. An additional exploratory analysis of the included studies was conducted to determine the highest and lowest diagnostic performance of GTA metrics^53,71. Contingency tables with the highest and lowest performance, as defined by the accuracy value, were selected from each study and subjected to meta-analysis. The risk of bias and applicability of all included studies were assessed using the Quality Assessment of Diagnostic Accuracy Studies-AI criteria⁷², which can be found in Table S26.

Data analysis in diagnostic meta-analysis

A bivariate random-effects model, accounting for the intrinsic correlation between sensitivity and specificity, was applied using the “midas” module in STATA software (version 18.0) to estimate pooled values of AUC, SE, and SP. Summary receiver operating characteristic (SROC) curves, forest plots, and pooled estimates of SE and SP were derived from the contingency tables, with an anticipated high level of heterogeneity⁷³. The SROC figures included the combined curve, along with 95% confidence intervals and 95% prediction intervals around the mean estimates of SE, SP and AUC. Publication bias was evaluated using funnel plots and regression analysis. Heterogeneity was quantified using the I² statistic. Statistical significance was defined as a P < 0.05.

As computer-aided diagnosis analysis often requires comparing the effectiveness of different input features and algorithm combinations, many studies have reported multiple contingency tables. To address the hierarchical structure of the data, with multiple outcomes nested within studies, a three-level meta-analysis was performed. The log-transformed DOR was modeled, incorporating random effects at both the “study “ and “table” levels to examine variability within and between studies. To avoid issues with zero-counting, a continuity correction of 0.5 was applied to the contingency table where any cells contained zero. Between-study variance (τ²) was estimated by the DerSimonian-Laird method, while overall heterogeneity was evaluated using Cochran’s Q, I², and τ².

To identify the sources of extreme heterogeneity, we performed subgroup analyses and moderator analyses based on the following four factors:

• Algorithm type (AI algorithms vs other algorithms): The ‘AI algorithms’ group includes studies that use machine learning algorithms, such as support vector machines, random forests, or deep learning algorithms, to differentiate between PD and HC. The ‘other algorithms’ group includes studies that use ROC analysis to calculate performance metrics for distinguishing PD vs HC.

• Imaging technique (functional imaging vs structural imaging techniques): The former includes studies that use functional imaging techniques, such as fMRI, PET, or EEG to construct brain networks and extract GTA metrics for distinguishing PD vs HC. The latter includes studies that use structural imaging techniques, such as dMRI or sMRI to construct brain networks and extract GTA metrics to distinguish PD vs HC.

• Feature selection: ‘GTA Only’ subgroup, using only GTA metrics as feature inputs vs ‘GTA Plus’ subgroup, combining GTA metrics with other network features as inputs.

• Threshold (proportion vs non-proportion): The ‘proportion’ group comprised studies that calculated GTA metrics using proportion thresholds, the ‘non-proportion’ group studies that calculated GTA metrics using non-proportion thresholds.

Author contributions

X.L.S. and Q.Y.G. designed the study. C.Z., W.X.L., X.L.S., and H.L. contributed to the literature search, data collection and interpretation. C.Z. contributed to statistical analysis of case-control meta-analysis. W.X.L. contributed to statistical analysis of diagnostic meta-analysis. C.Z. and W.X.L. drafted the manuscript. L.C., N.L., Y.Y., L.L., C.L., G.J.K., S.L., X.L.S., and Q.Y.G. critically revised the manuscript. All authors read and approved the final manuscript.

Data availability

All the data included in this study are available within the paper and its supplementary information files. The codes used in this paper are available on GitHub: https://github.com/chao9791/PD-brain-network-topological-properties-alterations.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-025-02301-x.