Participants older than 18 at the time of screening were prospectively enrolled at the MS Center of the Sant'Andrea Hospital in Rome between March 2018 and August 2019. Participants with a diagnosis of MS according to most recent revised McDonald criteria 2017 (18) regardless of disease phenotype were recruited together with their family members and hospital staff to act as control group. The ethical committee board of Sapienza University of Rome at Sant'Andrea Hospital provided approval for the project. Informed, written consent has been obtained in all participants.

Exclusion criteria for both groups were: (1) diagnosis of metabolic diseases; (2) mental or psychiatric diseases; (3) pregnancy or breastfeeding; (4) food allergies and food intolerances; (5) ongoing vegetarian or vegan diet; (6) any condition preventing participants to provide adequate answers to the administered questionnaires; (7) ability to walk without support/aid. For latter exclusion criteria, PwMS with and Expanded Disability Status Scale (EDSS) (19) higher than 5.5 were not recruited.

The following information were collected in PwMS and HC: demographic data, pharmacological therapies, smoking (number of cigarettes smoked per day and duration of habit, i.e., smoking pack/year), telephone and email contacts.

We also evaluated the following anthropometric indicators: weight and height to calculate Body Mass Index (BMI kg/m²), waist (WC) and hip circumferences, waist-hip ratio (WHR).

For BMI, we used World Health Organization (WHO) cut-points to asses underweight (<18.5 kg/m²), normal weight (18.5–24.9 kg/m²), overweight (25.0–29.9 kg/m²), and obese (>30 kg/m²). For men WC≥94 cm and for women WC≥80 cm are cut-points used to assess cardiovascular risk in the general population (20). Fat distribution is further defined as WHR above 0.90 for males and above 0.85 for females (21).

Subsequently, the following questionnaires were administered:

For each PwMS, the respective level of disability as measured by EDSS (and fatigue in terms of Fatigue Severity Scale (FSS) (24) were evaluated the same day of questionnaire completion.

PwMS were dichotomized into two subgroups according to their EDSS score: no disability (EDSS ≤1.5; i.e., no or minimal signs in more than one functional system) and disability (EDSS ≥2).

To identify what demographic or MeDi items were most influential for the eating habits of PwMS and how they were linked to disability and fatigue, we performed a network analysis. We created a structure depicting the connections, namely edges, among all the variables of interest, namely nodes, such as age, sex, and MeDi items (fruit, vegetables, legumes, cereals, fish, meat, dairy, olive oil, alcohol). Thanks to this representation, we were able to identify the hubs, that are the most connected variables, thus the variables with the highest amount of connections with the other variables. First, to generate the edges connecting the nodes of the network, we built a mutual information matrix, representing the relation among age, sex, and MeDi items for both patients and HC (17). The mutual information matrix was selected to describe the relation among age, sex and MeDi items, as it provided information on the mutual dependence between couples of variables, without any a priori assumption of linear association among them. Since the investigated variables were all categorical except for age, the continuous variable age was grouped by ranges of years and devised in 5 classes (from class 0 if age <20 years to class 4 if age>49 years and the other classes covering 10 years each) to build the mutual information matrix.

Then, we investigated the difference of age, sex and MeDi items association between PwMS and HC We obtained 1,000 random matrices by reshuffling, 1,000 times, values taken from the information matrices of PwMS and HC, and placing them into two matrices, that had the same dimensionality of the mutual information matrices but casual entries, and that were subtracted one from the other. Then, we averaged these 1,000 random matrices and compared the result with the difference between the mutual information matrix of HC and that of PwMS, by means of a Z-score. Z-scores were considered significant at alpha = 0.05 if Z>|1.96|.

Moreover, on the mutual information matrices of both PwMS and HCs, we calculated the hubs. Analytically, in a network, hubs are the variables with the degree larger than the average plus a standard deviation. The degree of an item is the number of its neighbors, e.g., of the number of links incident upon it. Network analysis was implemented in the open-source R environment (https://www.r-project.org/. For a flowchart of the network analysis see Figure 1.

All values are presented as mean (standard deviation) or median [range], as appropriate. Normal distribution assumption was checked for continuous variables by using Shapiro-Wilk tests.

Unpaired two-sample t-test and X² test were used to compare demographic, clinical, anthropometric indicators, adherence to MeDi, frequency of food intake and IPAQ scores between PwMS MS and HC, as well as between PwMS with different levels of disability, as appropriate.

To test the association between hub nodes from network analysis and disability, a multivariate logistic regression was conducted. Specifically, we included hubs (obtained by network analysis) as independent variable and scores of either disability or fatigue as dependent variable. Classes of disability were devised as no disability (class 0, EDSS < = 1.5), or disability (class 1, EDSS>1.5). Similarly, classes of fatigue were devised as no fatigue (class 0, FSS lower than the median value) or fatigue (class 1, FSS higher than the median value). Initially, we performed a sensitivity analysis to evaluate the minimum regression coefficient (beta) relative to sample size, considering alpha = 0.05 and power = 0.90. In multiple regression, collinearity exists if one variable is linearly predicted from the other and it may affect the reliability of the calculation of regression coefficients. Therefore, we checked for collinearity via Spearman correlation and removed correlated variables (p < 0.05 not corrected for multiple comparisons to be conservative at the most). However, for completeness, also association between dependent variables and removed hubs was calculated as bivariate logistic regression. Significance was reached if p < 0.05.

Finally, as the network analysis identified as hubs only items pertaining to the “positive foods,” and did not include lifestyle factors other than diet, we conducted a post-hoc analysis, investigating the relationship between “negative foods,” lifestyle factors (smoking, exercise) and clinical outcomes (EDSS and FSS) via hierarchical regression.

Analyses were conducted with the Statistical Package for Social Sciences 25.0 (SPSS, Chicago, IL, USA). Network statistical analyses were implemented in the open-source R environment (https://www.r-project.org/. Sensitivity analysis was performed with G-power (https://stats.idre.ucla.edu/other/gpower/.)

From an initial screening of 571 PwMS we enrolled 424 (74%) participants. Among the excluded patients there were 52 PwMS with metabolic diseases (9%), 47 with food allergies or food intolerances (8%), 19 people following a vegetarian or vegan diet (4%), 13 women in a gestational or breastfeeding state (2%), 12 (2%) have not been included due to incomplete data and 4 (1%) patients had psychiatric comorbidities. Two hundred and eighty four participants were screened as HC, of which 165 people (58%) were found to be suitable for study inclusion. Among the excluded participants there were 30 people with metabolic diseases (10%); 48 with food allergies or intolerances (17%); 18 people following a vegetarian or vegan diet (6%), 9 women in a gestational or breastfeeding state (3%), 7 have not been included due to incomplete data (3%) and 7 patients had psychiatric comorbidities (3%). For a flowchart of the study participants please see Figure 2.

Table 1 summarizes the main demographic characteristics and anthropometric measurements of the two examined groups. PwMS showed higher values of hip circumference compared to HC (p < 0.014), while there was a trend for waist circumference (p < 0.069) and no differences in terms of WHR and BMI. Among study participants only men in the HC group had BMI values belonging to the overweight range. No differences in terms of frequencies among the two groups were detected (Figure 3). Male HC subjects had WHR values above the cut-off, while all the other participants were normal weighted with normal WHR values. All subjects in our study were within the WC cut-off values.

Table 2 shows the main clinical characteristics and anthropometric measurements of the PwMS subgroups. PwMS in the disability group were older, had longer disease duration as well as higher FSS scores compared to PwMS in the no disability group. Waist circumferences and WHR were significantly different among the two subgroups.

The majority of participants in each group adhered sufficiently to MeDi (60% in the group of PwMS and 56% in HC group). In the PwMS group, only one participant had an inadequate diet, while 38% poorly adequate and 2% completely adequate diet. Among HC, 2% had inadequate adherence to MeDi, 41% of participants showed poor adherence, while 1% followed a completely adequate diet. There were no significant differences in terms of adequacy to MeDi between PwMS and HC. Participants less adherent to MeDi showed higher smoking pack/years values in the whole sample (p < 0.001) as well in the PwMS (p < 0.001).

When we explored differences in terms of food intake frequency, we observed a tendency to a healthier diet in the group of PwMS, as they consumed more frequently fish and less frequently alcohol than HC (p < 0.001) (Figure 4).

PwMS showed overall lower IPAQ scores (2227.21 ± 3587.26) compared to HC (3701.33 ± 8338.35) [p = 0.03]. Figure 5 shows the percentage of participants for each category according to groups.

In the group of PwMS there was a greater proportion of participants with a long-time smoking habit compared to HC (4.2 ± 8.6 vs. 1.0 ± 3.1, p < 0.01).

There were no differences in terms of adherence to MeDi between the two PwMS subgroups. However, when we compared the frequency of food intake, a greater proportion of participants in the no disability subgroup consumed fish more frequently than patients in the disability subgroup (p < 0.01).

PwMS in the no disability subgroup engaged in physical activities more often and more intensely compared to those with disability (2667.8 ± 4263.6 vs. 1814.7 ± 2757.7, p = 0.014).

PwMS in the no disability group had lower values of smoking pack/years compared to those with disability (2.9 ± 6.4 vs. 5.5 ± 10.1, p < 0.002).

We found no difference between mutual information matrices of PwMS and HC (Figure 6). When exploring each group separately, the network of PwMS identified fruit, vegetables, cereal and fish as hubs (Figure 7A), while HC's network identified meat and alcohol as hubs (Figure 7B).

Beta = |0.32| was the minimum beta evaluable with a sample of 424 PwMS. Hubs were all inter-correlated, as shown in Table 3. Significant positive Spearman correlations show monotonic association between variables, thus that the increase of one is paralleled by the increase of the other, thus eating habits of PwMS include a large intake of all the “positive foods” identified as hubs.

We found no difference between mutual information matrices of PwMS subgroups classified by EDSS or FSS.

Because of the collinearity, we did not perform a multiple regression including all the hubs in the same model, but rather a bivariate logistic regression for each hub to predict either EDSS or FSS groups. EDSS was predicted by the intake of vegetables (beta = −0.36, p < 0.03) and fish (beta = −0.34, p < 0.02). FSS was not predicted by any hub, e.g., fruit, vegetables, cereal and fish.

When testing the relationship between “negative foods” diet and disability, the model including smoking pack/year, IPAQ score and intake of “negative foods” predicted 6% of the variance in EDSS (p < 0.001), with smoking pack/year, exercise and diet being independent contributors (beta = 0.190, p < 0.001; beta = −0.138, p < 0.004; beta = 0.098, p < 0.039, respectively).

The model including smoking pack/year and IPAQ score predicted 4% of the variance in FSS (p < 0.001), with smoking pack/year and exercise being independent contributors (beta = 0.135, p < 0.005; beta = −0.169, p < 0.001, respectively).

We found no significant differences in MeDi adequacy between PwMS and HC, in line with previous findings (25). However, although most of the participants in both groups were sufficiently adherent to MeDi, a very low percentage of participants were completely adherent to it (2% PwMS and 1% HC). Even though MeDi is the most used food model in Mediterranean countries, such as Italy, even the general population is not completely adherent to it, with younger age groups and smokers being less adherent (26). Accordingly, in our work participants less adherent to MeDi showed higher values of smoking pack/years and particularly, in the group of PwMS there was a higher percentage of participants with long-term smoking habits compared to HC. Generally, the category of smokers seems to have a less healthy lifestyle and incorrect eating habits than non-smokers (27).

When we explored differences in terms of food intake frequency, we observed a tendency toward a healthier diet in the group of PwMS, as they consumed more frequently fish and less frequently alcohol than HC. Despite the healthier diet, in our sample of PwMS higher values of waist and hip circumferences compared to HC were detected. WC is used to measure abdominal adiposity and high values are common among PwMS (28). While in our study population the BMI was in the range of accepted values and most of the PwMS fell in the normal weight category, there is evidence showing that an increase in circumferences values is associated with greater disability even in normal weight subjects (28). This is confirmed in our sample, where patients with higher level of disability showed significantly higher WC and WHR values in comparison with the no disability group. An excess of visceral or abdominal adiposity is one of the characteristics of the metabolic syndrome often related to other disorders such as diabetes, hyperlipidemia and hypertension (28). Metabolic and vascular comorbidities affect both neuroperformance and brain and gray matter volumes in MS, contributing to neurodegeneration and long-term disability (29).

Even though in our work we did not collect serological nutrional biomarker and such interpretations remain speculative, several evidences support the correlation between an excess of adipose tissue and the severity of MS. Firstly, an overweight or obese subject has a state of chronic inflammation characterized by an altered production of cytokines, such as IL-6, TNF-α, leptin, and a downregulation of anti-inflammatory molecules (30). Additionally, overweight negatively impacts disease course by modulating monocyte cell number through ceramide-induced DNA methylation of anti-proliferative genes (31). This increased state of inflammation of overweight/obese subjects is also observed within the Central Nervous System (CNS). In particular, a recent study has shown elevated levels of proinflammatory molecules (IL-6 and leptin) and reduced levels of the anti-inflammatory cytokine IL-13 in the cerebrospinal fluid (CSF) of obese MS patients, and patients with higher BMI (BMI > 30 kg/m²) also had significantly higher EDSS values (30).

Excess of adipose tissue in PwMS might also affect daily clinical practice, as it can interfere with the response to drugs, possibly due to altered drug pharmacokinetics (32). Indeed, a study carried out on a population of adult patients, under interferon-beta (IFNβ) therapy, showed that overweight and obese patients have increased disease activity, as assessed by NEDA status (composite score of no evidence of disease activity). This result indicates that overweight and obesity may have an impact on IFNβ-treatment response (33). Based on these data, regular exercise by reducing visceral adiposity may play a key role in reducing inflammation. Finally, in line with previous evidence (34), PwMS were less physically active and showed overall lower IPAQ scores compared to HC.

The food network analysis identified fruits, vegetables, cereals and fish as hubs in PwMS with meat and alcohol being hubs in HC, suggesting, in line with our analysis of food intake frequency, that PwMS tend to have a healthier diet than HC. It is possible that PwMS, having a chronic disease, are urged to pay more attention to what they eat than the general population. Indeed, a previous study has shown that patients, following an initial clinical diagnosis of CNS demyelination, tend to change their diet, increasing the amount of fruit and/or vegetables and following a low-fat diet (35). Additionally, a recent work has shown a trend toward reduced levels of storage lipids (fatty acids, cholesterol esters, triglycerides, and diglycerides) in the plasma of PwMS in comparison with HC (36). Additionally, in line with previous findings highlighting the potential effects of MeDi on MS course and disability (16) in our study participants higher consumption of vegetables and fish was inversely correlated to global motor disability, as assessed via EDSS. This relationship hints to a beneficial role of food hubs on disability outcomes but might also disclose an inverse causality.

Foods identified as hub of PwMS (fruits, vegetables, cereals and fish) are rich in fiber, vitamins, ω-3 polyunsaturated fats or antioxidant molecules and possibly exert a protective effect on the disease (14, 37). In particular, antioxidant molecules such as vitamins C and E, plant polyphenol and carotenoids might contrast free radicals, with beneficial effects on the inflammatory response (37). Another mechanism by which hub foods that could influence the disease outcomes is the modulation of the microbiota (38). Specifically, foods with omega-3 polyunsaturated fatty acids (PUFA) and fiber can positively modify microbiota through the proliferation of bacteria with anti-inflammatory action. Both short-chain fatty acids (SCFA) obtained from the bacterial fermentation of dietary fibers, and PUFAs, contained in fish, have an anti-inflammatory function (39, 40). Nonetheless, we cannot exclude that disease-related symptoms such as mobility deficits and cognitive impairment in more disabled PwMS might have affected the access to healthier food, that usually require a more dedicated preparation rather than easy snacks or fast-foods, partly accounting for the observed relationship between food hubs intake and disability (41). However, in our sample of PwMS we have not recruited those with an EDSS higher than 5.5, also in order to restrain this potential bias.

Our post-hoc analysis disclosed an independent predictive role for “negative foods,” exercise and smoking on motor disability. This finding suggests that not only the intake of “positive foods” is related to disability, but that the global composition of the dietary regimen as well as the lifestyle adopted potentially affect MS outcomes. Specifically, both smoking habit and low level of exercise were related to disability and fatigue independently from diet, as previously suggested (42, 43). Although reverse causality and simultaneity cannot be excluded when interpreting these associations, the relevance of a balanced dietary regimen, rich in “positive foods” and poor in “negative foods,” associated to a healthy lifestyle, might explain why studies exploring the beneficial effects of isolated dietary supplementations are often unsuccessful (44).

Our work is not without limitation. First, the cross-sectional design of our work limited the ability to explore the temporal relationship between lifestyle habits and changes in disability and fatigue over time, preventing us from drawing conclusions about direct causality. Second, we used a qualitative questionnaire to assess the adequacy to MeDi and food intake, intrinsically limiting the exact estimation of nutrients intake. We did not measure serum markers (i.e., antiflammatory or antioxidant) that could better reveal the association between nutrition intake and disability or fatigue if inserted in the network analysis, but on the basis of our results, future more comprehensive study on the matter might lead to further elucidation. Moreover, the use of family members as part of the control group might have limited the possibility to reveal real differences in dietary and lifestyle habits with PwMS. We have tried to minimize this occurrence introducing another set of controls in the HC group (healthcare staff). Indeed, the choice of family members as control group lie on the basis that family peers share characteristics such as socioeconomic and educational level with their relatives, that are known to influence diet behaviors (45).