Neural networks involved in painful diabetic neuropathy: A systematic review

2.1 Search strategy

A first comprehensive systematic search was performed via OVID in Embase, PsycINFO, and MEDLINE to identify eligible studies from 2010 to the date of the last search (May 16th, 2024). In MEDLINE and EMBASE, we used a combination of terms for three concepts (i.e., neuroimaging AND neuropathic pain/diabetic neuropathy NOT animals). The search strategy for not including animal studies in MEDLINE and EMBASE was based on the Cochrane Handbook’s highly sensitive strategy to identify animal studies [18,19,20]. In PsycINFO, we used a combination of terms for two concepts (i.e., neuroimaging AND neuropathic pain/diabetic neuropathy “Search terms by database can be sent upon request”). Synonyms and keywords differed slightly between the databases, due to variations in the architecture of their search engines. It was ensured that each search was maximally inclusive by exploring all major headings and manually including or excluding each of the respective subheadings. The identified studies were imported into EndNote for duplicate removal.

2.2 Selection process

The process of selecting the eligible studies was conducted using the Covidence software [21]. Four authors (J.A., J.S., C.F., and U.K.) screened all titles and abstracts, with each unique record being screened by two independent authors. A fifth author (E.D.) was involved in case of disagreement. Titles and abstracts were included or excluded based on the criteria defined below. Only publications written in English in a peer-reviewed journal were included. The included studies in the title/abstract screening were subject to full-text screening. Here four independent authors (J.A., J.S., C.F., and E.D.) screened the studies, and discrepancies were discussed between them and with the wider team with the relevant expertise (U.K. and S.A.).

2.3 Updated searches

Two updated searches were conducted to incorporate the latest literature, enhancing the comprehensiveness and relevance of this systematic review. The updated searches were conducted on December 5th, 2023, and May 16th, 2024. The search strategy used for the updated searches mirrored the original search conducted on October 19th, 2022.

2.4 Inclusion and exclusion criteria

2.5 National institutes of health (NIH) quality assessment tools

Assessing methodological quality (risk of bias) is a crucial step when conducting a systematic review [22]. Among the various tools available, the NIH Quality Assessment Tools (NIH-QATs) [23] were chosen for the current review. The NIH-QATs were preferred over alternatives like the Newcastle-Ottawa Scale (NOS) [24] and the Cochrane Risk of Bias Tool 2.0 (RoB 2.0) [25] for several reasons. The NIH-QATs provide a flexible and comprehensive framework suitable for evaluating different study designs, including case-control, observational cohort, and cross-sectional studies – all essential to the current review. Additionally, the NIH-QATs are specifically designed to help reviewers focus on key concepts critical for assessing methodological quality, and their user-friendly format promotes consistent application [23]. While NOS and RoB 2.0 are valuable in specific contexts, they have a narrower scope and require more advanced expertise for effective use [22,26]. The NIH-QATs were used by four independent authors (J.A., C.F., J.S., and F.P.) for the quality assessment of the included studies. Discrepancies were discussed between the authors. The NIH-QATs include the NIH-QAT for Case-Control Studies, Observational Cohort and Cross-Sectional Studies, and Controlled Intervention Studies. The NIH-QAT for Case-Control Studies was used for most of the included studies. The risk of bias part of the tool comprises 12 questions; however, two of the questions were removed because they were only relevant for rating studies with a longitudinal design (question 7: “If less than 100% of eligible cases and/or controls were selected for the study, were the cases and/or controls randomly selected from those eligible?” and question 9: “Were the investigators able to confirm that the exposure/risk occurred before the development of the condition that defined a participant as a case?”). For the remaining questions, exposure or the independent variable was treated as diabetes or DPN, depending on the study. For question 14, which related to the control of confounding variables, we considered the following variables as the variables that should be controlled or matched across the study groups: (1) age; (2) gender; (3) disease duration; (4) body mass index (BMI); and (5) haemoglobin A1C (HBA_1C) (variables 3 and 5 only in diabetes and DPN groups) [23]. This was based on the literature on demographic effects on MR sequences and a recent review [27,28]. The remaining included studies were scored with the NIH-QAT for Observational Cohort and Cross-Sectional Studies [23]. Each study was also individually assessed according to quality outcomes for the neuroimaging method used in the studies. This included looking at external validity and precision, sample size, representativeness of the sample included, and variables controlled for.

2.6 Data extraction

Relevant data from the included studies was extracted in Covidence by J.A. and C.F. and checked by U.K., S.A., or E.D. The data included the title of the study, year of publication, study aim, as well as the age, gender, diabetes duration, HbA_1c, type of diabetes, BMI, and microvascular complications of the target population. We also extracted data related to the neuroimaging technique used in the included studies, such as the analysis type and results (Region of interest [ROI], MNI-coordinates for peak voxels, voxel sizes [K-value or volume] and p-values). The data are presented in Tables 1 and 2.

2.7 Data synthesis

Given the heterogeneity of the applied research methods in terms of measures and study design, we anticipated that a meta-analytical approach would not be feasible. This was confirmed after conducting the searches and extracting the data. A narrative synthesis summarizing the characteristics and findings of the studies included in the review, along with a discussion of the risk of bias and quality, are, therefore, presented below.

3.1 Study selection

Figure 1 summarizes the number of studies in each stage of the selection process. We identified a total of 11,735 studies in the databases. Specifically, we identified in EMBASE: 5,819 studies, in PsycINFO: 2,447 studies, and in MEDLINE: 3,469 studies. After 3,306 duplicates were removed, 8,429 studies remained. Of these, 8,258 studies were excluded during the title/abstract screening because they did not fulfill the inclusion criteria, leaving 171 studies for the full-text screening. The 8,258 studies were mainly excluded because they: (1) involved conditions other than DPN, (2) were case studies, or (3) were animal studies. Of the 171 studies screened on full-text level, 153 studies were excluded for the reasons detailed in Figure 1. The remaining 18 studies were included in this systematic review and critically appraised.

Figure 1

PRISMA diagram.

3.2 Risk of bias assessment

The results for the risk of bias assessments are shown in supplementary material (Table S1 for case-control studies and Table S2 for observational studies). For case-control studies (15 out of the 18 included), all the studies presented the research question (Q1). The target population was clearly defined in 94% of the studies (Q2), with one study not distinguishing between type 1 diabetes mellitus (T1DM) and type 2 diabetes mellitus (T2DM). However, only one study justified the sample size (Q3). For Q4, 87% had controls from a similar population. Questions 5, 6, and 8 demonstrated better quality with all the studies receiving a yes for these questions. For Q10, 33% did not clearly define the exposure. Regarding blinding (Q11), 80% of the studies did not document blinding. For Q12, the response was assessed as a “yes” if the study had matched the study groups regarding the variables, used them as covariates, or controlled for them in the statistical analyses. 60% of the studies failed to satisfy at least one of these criteria.

Of the observational cohort and cross-sectional studies, one study received a score 6/12, the second one received a score 7/12, and the last one received a score 8/12 [29,30,31]. As mentioned above, methodological shortcomings included the sample size, statistical analyses, blinding questions, and levels of exposure.

3.3 Study characteristics

Table 1 details the characteristics of the included studies. We found that four studies [29,32,33,34] investigated individuals with both T1DM and T2DM, six studies [35,36,37,38,39,40] investigated only T1DM, and six studies [31,41,42,43,44,45] investigated only T2DM. Two studies [30,46] did not specify the type of diabetes. All the studies had a cross-sectional design, and 13 studies [33,35–46] included a healthy control (HC) group. Moreover, 15 studies [29,31,33–39,41–46] used structural NRI or fMRI, 2 [30,40] used EEG, and 1 [32] used SPECT. In addition, one study [40] used an EEG registration with a tonic pain paradigm in participants with P-DPN compared to HCs.

3.4 P-DPN vs NP-DPN

Overall, the results can be categorized into (1) studies comparing P-DPN with NP-DPN and (2) studies comparing P-DPN or NP-DPN with individuals with diabetes without DPN (no DPN) or HCs. The included studies employ a variety of imaging methods and analytical approaches. The results section will first present studies investigating the difference between P-DPN and NP-DPN, detailing functional brain changes in these groups. An examination of structural brain changes will follow this. Finally, the result section will provide an overview of studies comparing P-DPN or NP-DPN with no DPN or HCs.

3.5 P-DPN or NP-DPN vs no DPN or HCs

While our primary focus is on the differences between P-DPN and NP-DPN in terms of brain changes, the comparisons to individuals without DPN and HCs are also of interest. The results of the included studies investigating P-DPN or NP-DPN vs no DPN or HCs are given in Table S3 of supplementary material.

4.1 Structural alterations

The thalamus, crucial for relaying nociceptive information from spinal neurons to cortical areas [47], exhibited grey matter volume reductions in P-DPN [35]. These reductions correlated directly with neuropathy severity, underscoring thalamic involvement in understanding P-DPN’s symptoms and its potential contribution to pain [33,34,45]. Although most included studies have studied functional S1 alterations, a key area for sensory information processing, one study showed that P-DPN relative to NP-DPN was related to lower grey matter volume in this and the M1 regions [33]. Interestingly, a study demonstrated that people with P-DNP who did not respond to lidocaine treatment had lower S1 volume compared to responders and HCs [29]. Additionally, in the P-DPN phenotypes of insensate and sensate, the insensate group had the lowest somatosensory cortical thickness, particularly in the precentral gyrus thickness [39]. While more studies on structural studies on S1 alterations driven by P-DPN are needed to better understand the influence of these changes, the available literature emphasizes that structural changes in S1, influence pain perception and modulation in P-DPN and that specific alterations in brain anatomy are associated with variations in pain perception or sensory deficits.

Grey matter volume reductions extend beyond these areas to the hippocampus and insula, further linking neuropathy severity to brain structural alterations [34].

4.2 Functional alterations

The thalamus showed heightened thalamocortical connectivity in P-DPN, indicating increased involvement in pain processing and evaluation compared to NP-DPN [33,34,45]. In comparisons of NP-DPN or P-DPN with HCs or people with diabetes without DPN, P-DPN demonstrated reduced connectivity between the postcentral and precentral gyri and between the thalamus and motor areas [36]. Another study also showed lower connectivity between the left thalamus and supplementary MI/superior frontal gyrus in the group with P-DPN compared with the group without DPN [37].

In S1, evidence of increased energy consumption [31] and higher connectivity in EEG-recorded theta and beta bands in P-DPN [30] were shown. Increased theta activity likely indicates heightened emotional and cognitive attention to pain, while beta activity may relate to motor-related aspects [48,49]. Differences in fALFF further support this detected across multiple brain regions associated with cognitive, somatosensory, and emotional functions, including the temporal, occipital, and parietal lobes, and various regions within the central gyrus and supplementary motor area [41]. Overall, P-DPN appears to be characterized by widespread brain activity, extending beyond the traditional pain-processing centers.

P-DPN exhibited higher CBF in the ACC than NP-DPN. A structure that plays a crucial role in the emotional aspects of pain [42,43,44], and underscores the greater emotional burden in P-DPN [29,32,33]. The ventrolateral periaqueductal grey, a midbrain region [50], demonstrated altered FC and enhanced pain-induced activation [34], highlighting the complex network of brain regions activated during painful states, involving basal automatic mid-brain to complex higher-order cortical controlled processing.

Moreover, reduced white matter connectivity within regions of the DMN, suggests that P-DPN affects not only pain perception but also cognitive and emotional responses, contributing to distress and impaired task performance [51]. Higher connectivity in P-DPN was observed in electrode pars associated with the DMN, indicating a hyperactive state in chronic pain [30]. This heightened activity may contribute to the amplified perception of pain and emotional distress characteristic of P-DPN, reflecting an impaired ability to regulate and filter painful information [52]. Moreover, peripheral nociceptor hyperexcitability, a key factor in DPN, has been suggested to cause ectopic nerve firing and increase sensitization of second-order neurons, ultimately impeding the descending modulation of pain [53].

4.3 Neuroimaging in P-DPN: Comparative insights and therapeutic potential

P-DPN exhibits both shared and unique neuroimaging characteristics compared to other chronic pain conditions. Chronic pain commonly affects large-scale brain networks, leading to structural and functional changes in regions such as the insular cortex, frontoparietal network, and prefrontal cortex [54,55]. Studies on conditions like chronic back pain and fibromyalgia highlight the prefrontal cortex’s role in cognitive-emotional pain modulation [56]. P-DPN, however, demonstrates distinct alterations, including reduced grey matter volume in the anterior thalamus and increased volume in the postcentral gyrus compared to individuals with diabetes but without neuropathy [33]. These findings suggest both shared and condition-specific neural mechanisms in chronic pain.

Such neuroimaging insights may guide targeted neuromodulation. Identified regions, such as the thalamus, S1, ACC, and insula, could be key targets for precision pain management. rTMS targeting the MI has shown efficacy for neuropathic pain [57], while transcutaneous electrical nerve stimulation may benefit peripheral neuropathic pain. However, evidence for tDCS and deep brain stimulation remains inconclusive, and while invasive MI stimulation has shown promise, its role in P-DPN is unclear. Spinal cord stimulation has potential but lacks rigorous sham-controlled trials.

Future research should personalize neuromodulation strategies based on individual neuroimaging findings. Given variability in brain alterations among patients [58], a precision medicine approach could optimize treatment selection. Well-controlled trials incorporating patient preferences, cost, and risk-benefit considerations are essential to advance clinical application.

In summary, integrating neuroimaging with neuromodulation may enhance personalized treatments for P-DPN and other chronic pain conditions. Despite challenges in translating imaging biomarkers into clinical use [59], refining network-based interventions holds promise for improving patient outcomes.

4.4 Summary

Together, these structural and functional changes provide a comprehensive view of how P-DPN is associated with alterations in brain networks beyond those linked to pain processing, potentially influencing emotional and cognitive processing in ways that differentiate it from NP-DPN. Functional changes in key pain-processing regions, such as the thalamus and S1, were observed. Although FC is bidirectional between brain regions, we hypothesize that alterations in primary sensory areas like the thalamus and S1 likely drive changes in secondary processing regions, such as the insula and ACC. Given the roles of the thalamus and S1 in relaying and processing nociceptive input, it is plausible that disruptions in these areas affect how pain is later modulated and emotionally evaluated in higher-order regions. This supports the notion that, in P-DPN, altered sensory processing contributes to the heightened emotional and cognitive perception of pain, rather than secondary regions driving changes in primary sensory areas. Understanding this directional relationship offers valuable insights into how pain information is processed in P-DPN and can inform future therapeutic strategies to target primary sensory circuits to modulate pain perception.

4.5 Strengths and limitations of imaging techniques

Various imaging techniques have been employed to study the neural correlates of pain in P-DPN, each with unique strengths and limitations. Understanding these methodologies is crucial for interpreting the findings of the neuroimaging studies and their contributions to the literature.

4.6 Methodological strengths and limitations

This review fills a gap in the scientific literature and adheres to PRISMA guidelines. However, it also has methodological limitations. First, for feasibility reasons, the search was restricted to the three most used databases, and date limits were applied to focus on recent studies, potentially introducing selection bias. Second, grey literature was not included, which may lead to publication bias. Third, the included studies were heterogeneous in design, with three being observational and employing cross-sectional or case-control designs. This heterogeneity complicates the ability to determine whether observed associations reflect true cause-and-effect relationships or are merely coincidental, influenced by confounding variables. Our quality appraisals revealed several key areas for improvement. Notably, most studies had small sample sizes and lacked power calculations, with only three of 18 using blinded assessors. Eleven of 18 studies omitted adjustment for potential confounding variables [29,30]. While comparing P-DPN to no DPN provides insight into the neural networks associated with P-DPN, comparisons between P-DPN and NP-DPN would be preferable. The observed differences between P-DPN and no DPN may arise from the presence of DPN rather than the specific pain characteristic of P-DPN. Similarly, comparing P-DPN to HCs introduces additional variables, complicating the identification of neural networks uniquely associated with P-DPN.