In this observational study, 166 consecutive patients with remissive cancer were recruited at the Local Health Unit of Naples 3 South, Naples, Italy. Inclusion criteria were:

Exclusion criteria were:

All enrolled volunteers underwent detailed and bioimpedance phenotyping, including measurement of height, weight, waist and hip circumferences. Height was measured with a wall-mounted stadiometer GIMA™ 27335, while waist and hip circumferences with a SECA™ 201 non-stretchable measuring tape. Bioimpedance parameters were assessed by TANITA ™ MC-780U multifrequency segmental body composition analyzer, and included Weight, Fat Mass (FM), Fat-Free Mass (FFM), Basal Metabolic Rate (BMR), Total Body (TBW) and Extra-Cellular Water (ECW), Visceral Adipose Tissue (VAT), Bone Weight, and Phase Angle (°). Body Mass Index (BMI) was calculated as a ratio of body weight/height² (kg/m²).

Clinical and Pathological data were collected at time of enrolment (T0) and after a 6-month follow-up (T1). Data included cancer type, presence of metastasis, cancer recurrence, and eventual ongoing chemotherapy/treatment status, as well as eventual comorbidities, like Metabolic Syndrome (METS), Type-2 Diabetes (T2D), cardiovascular diseases (CVD), Non-Alcoholic Fatty Liver Disease (NAFLD), Hyperlipaemia (HL), and Anemia (A). Moreover, a detailed questionnaire about anamnestic and behavioral data, such as smoking habit, physical activity, and weekly food frequency, was administered to each volunteer (as described below). Cytometric blood counts, biochemical analyses, and tumor-marker screenings were collected as well.

A serum sample from all donors was obtained at T0 and T1, and stored at -20°C.

Investigations were carried out following the rules of the Declaration of Helsinki of 1975, revised in 2013. Informed consent was obtained from every volunteer before the procedure. The protocol was approved by the ethical committee of the University of Naples.

The assessment of volunteers’ nutrition habits was based on the procedure described in D’Esposito et al. (2022) (8). Briefly, a food frequency questionnaire (FFQ) was de visu administered by two expert nutritionists. The FFQ involved information about the last 7 days of food frequency and amount and was based on the Italian EPIC study (16, 17). Consequently, a list of 12 food groups (namely Cereals, Milk, Seafood, Crustaceans, Meat, Poultry, Processed Meats, Vegetables, Fruits, Cakes, Pizza, and Beverage) was developed in line with the local food availability, cultural specificity and dietary habits. Food frequency type with portion size was estimated by means of pictures. Pictures showing different portion sizes – arranged by increased amount – followed the question on frequency and corresponded to a specific portion in grams. Questionnaires were finally digitalized, and data related to the weekly total consumption of each food item were gathered and analyzed.

The dietary intervention consisted of a qualitative dietary change characterized by low glycemic and insulinemic indexes, with the Mediterranean model as a blueprint. The intervention objectives were the promotion of a healthy diet, the mitigation of antineoplastic therapy side effects, the improvement of patients’ metabolic derangement, inflammatory status, and body composition.

To achieve these goals, all participants were provided with the Italian Dietary Guidelines for Healthy Eating, released by the Research Center for Food and Nutrition of CREA (Council for Agricultural Research and Economics) (18, 19). These guidelines, grounded in the Mediterranean dietary model, take into account the recommendations of World Cancer Research Fund (WCRF) (20), and encourage to:

In detail, the Mediterranean dietary protocol administered to patients consisted of three main meals (breakfast, lunch, and dinner) and two snacks (mid-morning and mid-afternoon). Breakfast included fermented milk products (e.g., yogurt) and baked goods or cereals (flakes or puffed), preferably whole-grain or based on ancient grains. Lunch and dinner were structured around a generous portion of fresh or cooked seasonal, locally sourced vegetables; a source of complex carbohydrates such as whole grains (e.g., pasta, bread, or grain kernels such as rice, spelt, and barley) or potatoes (twice per week); and a moderate portion of proteins, with a preference for legumes and fish (at least three times per week), white meat (2–3 times per week), ricotta or aged cheeses such as Parmigiano Reggiano™, and eggs (twice per week), while limiting red meat and processed meats to once per week. Extra-virgin olive oil was recommended as the main condiment for both lunch and dinner. For snacks, options included fresh fruit, fermented milk products (yogurt), nuts, and oilseeds. In addition, participants were allowed one glass of red wine if desired, and a dessert once per week.

These recommendations allowed to standardize the dietary approach even for patients with metabolic comorbidities (e.g. T2D, Hyperlipaemia, NAFLD).

To ensure greater adherence to the prescribed dietary plan, periodic nutritional follow-up visits were conducted approximately every 30 days, allowing for the evaluation of anthropometric and bioimpedance parameters. During these sessions, adherence to the Mediterranean diet model was monitored through an intervention-modulated FFQ, and nutritional counseling was provided to guide patients through their dietary transition and behavioral change, as well as tracking the overall quality-of-life and clinical follow-up.

Serum samples at T0 and T1 were screened for the concentration of interleukin (IL)-1β, IL-1ra, IL-2, IL-4, IL-5, IL-6, IL-7, IL-8, IL-9, IL-10, IL-12 (p70), IL-13, IL-15, IL-17A, Eotaxin, basic Fibroblast Growth Factor (bFGF), Granulocyte-Colony Stimulating Factor (G-CSF), Granulocyte and Macrophage-Colony Stimulating Factor (GM-CSF), Interferon-γ (IFN-γ), Interferon-γ inducible Protein-10 (IP-10), Monocyte chemoattractant Protein-1 (MCP-1), Macrophage Inflammatory Protein-1 alpha (MIP-1α), MIP-1β, C-C motif chemokine ligand 5/Regulated on Activation, Normal T-cell Expressed and Secreted (CCL5/RANTES), Tumor Necrosis Factor-α (TNF-α), Platelet-Derived Growth Factor-bb (PDGF), and Vascular Endothelial Growth Factor (VEGF), using the Bio-Plex^® Multiplex Human Cytokine, Chemokine, and Growth Factor Kit (cat. number M500KCAF0Y, Bio-Rad, Hercules, Ca, USA), as previously described (8, 21, 22).

Additionally, a second array of metabolic markers, including C-Peptide, Ghrelin, Gastric Inhibitory Peptide (GIP), Glucagon-Like Peptide-1 (GLP-1), Glucagon, Insulin, Leptin, total Plasminogen Activator Inhibitor-1 (PAI-1), Resistin, and Visfatin, was screened using the Bio-Plex Pro^® Human Diabetes 10-plex Assay (cat. number 171A7001M, Bio-Rad, Hercules, Ca, USA), as previously described (23–25). Both magnetic bead-based assays were performed on a Bio-Plex 200 analyzer with the Bio-Rad Bio-Plex Manager software (Bio-Rad, Hercules, CA, United States).

High sensitivity C-Reactive Protein (CRP, cat. number L2KCR2, Immulite, USA) assay was performed using the IMMULITE^® 2000 Analyzer (DPC, Los Angeles, Ca, USA), according to the manufacturer’s protocol. High Molecular Weight Adiponectin (cat. number 296752, Fujirebio, Tokyo, Japan) CLIA assay was performed using the Lumipulse^® G600II analyzer (Fujirebio, Tokyo, Japan).

Statistical Analysis was performed using both GraphPad^® Prism 8.4.2 software (GraphPad Software Inc, La Jolla, Ca, USA) and the R version 4.1.2 programming language (R Foundation for Statistical Computing, Vienna, Austria) through the RStudio^® Integrated Development Environment (Posit PBC, Boston, Ma, USA).

Categorical Variables were described as the number of occurrences and relative frequency as percentage and compared using Fisher’s exact test. Continuous variables were described as either mean ± standard deviation or median and interquartile range (IQR, 25th to 75th percentile). Shapiro-Wilk normality test was used to evaluate whether the continuous data were normally distributed.

Correlations among two continuous variables were reported and analyzed with either Pearson’s r correlation coefficient or Spearman’s rho rank correlation coefficient, and their respective 95% confidence intervals (95% CI). for two-group comparisons, either Welch’s two tailed t-test for independent samples (for parametric data) or Mann Whitney U-test (for non-parametric data) were used. Multiple comparisons among more than two groups were assessed using either ANOVA test with Tukey’s post hoc correction (for parametric data) or Kruskal-Wallis test with Dunn’s post hoc correction (for non-parametric data). Intra-patient comparisons of variables between T0 and T1 were reported as mean variation (Δ), with 95% CI, and assessed through either paired t-test or paired Wilcoxon test, along with Cohen’s d evaluation and relative 95% CI for effect size assessment. Outliers were detected and accordingly removed with the ROUT (Robust regression and Outlier removal) method of regression. Briefly, a robust nonlinear regression was used to fit a curve not influenced by outliers; then, the residuals of the robust fit analyzed to identify any outliers. This step used an outlier test adapted from the False Discovery Rate approach of testing for multiple comparisons; finally, outliers were removed and a least-squares regression on the remaining data was performed.

High dimensional data visualization algorithms, such as t-distributed stochastic neighbor embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP), were used for visualizing the high dimensionality data on two-dimensional projections. After empirical fine-tuning, t-SNE was used with a perplexity factor equal to 30, while UMAP parameter for neighbors was set to 15. Moreover, Multiple Factor Analysis (MFA) was used to both analyze a dimensionality reduction with a proper variable grouping, the single-variable impact on the cumulative overall variance, and the tumor-relative clustering. All analyses were performed with a fixed random seed (04031960) to ensure reproducibility.

To investigate the collinearity effects among risk factors and food groups, data were analyzed with a multivariable linear regression analysis. Generated models were analyzed with an ANOVA test to evaluate the goodness of fit. Regression coefficients were reported as an estimate and their 95% CI.

All p-values <0.05 were considered statistically significant.

166 consecutive patients with cancer in remission were recruited. Of these, 64 had Breast Cancer (BC), 52 had Colorectal Cancer (CC), and 16 had Prostate Cancer (PC). 34 of them had poorly represented tumors (e.g. Endometrial Cancer, n: 6; Bladder Cancer, n: 5; Kidney Cancer, n: 4) and were not included in further analyses.

At time of enrolment (T0), BC, CC, and PC patients displayed a well distinct characterization, with the BC groups being the youngest (mean age: 55.88 ± 11.06 years) but with the highest BMI (median: 29.56 Kg/m², IQR 26.81 to 33.98 Kg/m²). These differences were maintained after stratifying age in 10-year range subgroups and BMI in Normal Weight, Overweight, Obese. Apart from slightly lower alcohol exposure for BC patients, no differences in behavioral, blood count and biochemical data were noted (Table 1, Supplementary Table S1).

Bioimpedance data reflected the marked anagraphical and physical differences among the three groups. Presence of Metastasis was reported mainly for CC patients (23.08%), while chemotherapic treatment for BC patients (51.56%) (Table 1).

Dietary information was collected with a weekly food-frequency questionnaire, as described in the “Materials and Methods”, and further analyzed. No gross differences in dietary habits based on food frequency and distribution were noted (Supplementary Table S2).

Patients’ serum was screened for the concentration of two panels of molecules: a 27-molecules array of cytokines, chemokines, and growth factors, and a 10-molecules array of metabolic markers. At the same time, circulating levels of C-Reactive Protein (CRP) and Adiponectin were screened. Overall, all factors were detectable in serum specimens (Table 2). Screened molecules displayed an overall balanced distribution among the three cancer subgroups, with the exception of a slight increase of IL-12 in PC patients, and a remarkably higher concentration of Leptin in BC patients (Table 2).

We have investigated the complex association between all patients’ data by using multiple Machine Learning approaches on the overall enrolled population. This strategy was also chosen to avoid potential biases related to small sample subgroup sizes.

Firstly, all parameters evaluated at T0 were stratified in six variable groups (namely “Anagraphic”, “Bioimpedential”, “Clinical”, “Dietary Habits”, “Cytokines”, and “Metabolic Markers”) and subjected to a Multiple Factor Analysis (MFA).

The overall difference in variables distribution is shown in the biplot, in which very homogeneous clusters of variables can be noted (Figure 1A). The variables-group representation also depicts a very strong association among cytokines and metabolic marker distributions, which resulted very close to dietary habits variables; conversely, these molecular markers resulted the furthest from bioimpedential parameters (Figure 1B).

The two main dimensions (Dim1 and Dim2) obtained from the MFA allowed to show the contribution of single variables (Figures 1A, C, D, Supplementary Table S3). In Dim1, the primary contributors were variables from the Bioimpedential group, with BMI, FM, and VAT showing the highest contributions to the overall variance. Additional positive contributors included age and Sex stratification, alongside clinical evidence of METS and CVD. Conversely, variables such as therapy and Adiponectin levels negatively contributed to the association among variables (Figure 1C, Supplementary Table S3).

In contrast, although cytokines and other molecular markers contributed notably to Dim1, their impact was more pronounced in Dim2, in which bFGF, Visfatin, IL-1ra, IL-2, TNF-α and others had the most substantial positive contribution to variance. Conversely, anagraphic and bioimpedential variables – e.g. Sex, BMI, and VAT – exhibited negative contributions for this dimension’s variance (Figure 1D, Supplementary Table S3).

This approach revealed latent structures within the multileveled dataset. Notably, Adiponectin was positioned in opposition to a well-grouped cluster of inflammatory cytokines, including Visfatin, IL-6, IL-8, TNF-α, and CRP. At the same time, this inflammatory cluster resulted associated, along Dim1, with bioimpedential markers, such as BMI, FM, and VAT. Consistently, metabolic markers like Insulin and C-Peptide were positively linked to clinical and metabolic variables, such as Glucagon and T2D. Another notable positive association was identified between RANTES, IL-8 and the presence of Metastasis. These associations were further confirmed by the assessment of the single variables to the contribution for MFA dimensions (Figures 1A, C, D, Supplementary Table S3).

Finally, when the MFA output was clustered for tumor type, no separation among BC, CC, and PC groups was revealed, depicting a heterogeneous global population with no clear neoplastic-specific patterns (Figure 1E). The overlapping distribution among the three tumor subgroups was also emphasized after using non-linear dimensionality reduction approaches, such as t-Stochastic Neighbor Embedding (t-SNE) or Uniform Manifold Approximation and Projection (UMAP) (Supplementary Figure S1). All techniques, despite the limitations given to the small sample sizes, showed an overall homogeneous dispersion and a lack of proper separation for tumor types.

The population adhered to a Mediterranean-based dietary regimen. After 6 months (T1), bioimpedential data and a sample of serum were collected. At follow-up, a drop-out of 46.97% occurred, with 70 out of 132 starting patients remaining in the study. Baseline comparisons between participants who completed the 6-month intervention and those who discontinued showed no significant differences across demographic, clinical, nutritional, bioimpedance, cytokine, or metabolic parameters, with the only exception for glucagon levels, which resulted higher in patients who completed the intervention compared to dropouts (Supplementary Table S4).

Overall, upon dietary intervention, patients displayed significant reduction in body weight, BMI, FM, and VAT (all p-values <0.001) (Figure 2A, Supplementary Table S5). The screening of metabolic markers revealed a significant increase of Adiponectin (p-value: 0.012), accompanied by significant reduction of circulating Insulin (p-value: 0.013), Leptin (p-value: 0.018), and Visfatin (p-value: 0.013) (Figure 2B, Supplementary Table S5). These outcomes were accompanied by an overall reduction of inflammatory- (such as IL-2, Eotaxin, IFN-γ, IP-10, TNF-α, CRP) and immunomodulatory- (such as IL-4, IL-10, IL-12) molecules, and growth factors (e.g. bFGF, G-CSF, GM-CSF, PDGF) (Figure 2C, Supplementary Table S5).

Next, we investigated the effects of dietary intervention in the three subgroups. In BC patients, at follow-up, a drop-out of 43.76% was obtained, with 36 out of the 64 patients remaining in the study. Overall, these patients displayed a significant reduction for bioimpedential parameters, such as body weight, BMI, FM, and VAT (Figure 2A); these reductions were accompanied with significant reductions for cytokines, including IL-2 (p-value: 0.012), IL-7 (p-value: 0.025), IL-13 (p-value: 0.03), IP-10 (p-value <0.001), TNF-α (p-value: 0.007), and CRP (p-value: 0.038), but also for metabolic markers, such as C-Peptide (p-value: 0.004), GIP (p-value <0.001), Insulin (p-value: 0.005), Leptin (p-value: 0.013) (Figures 2B, C, Supplementary Table S6).

In regard to the CC subgroup, 25 of the 52 starting patients remained in the study, with a drop-out of 51.93%. Overall, these patients did not display any significant reduction for bioimpedential parameters or metabolic markers; however, they showed significant reduction of inflammatory chemokines and growth factors, among which IL-1β (p-value: 0.018), IL-2 (p-value: 0.025), IL-6 (p-value: 0.02), IL-8 (p-value: 0.027), bFGF (p-value: 0.04), GM-CSF (p-value: 0.023), IP-10 (p-value: 0.005), MIP-1α (p-value: 0.041), MIP-1β (p-value: 0.033), PDGF (p-value: 0.011), TNF-α (p-value: 0.027), and CRP (p-value: 0.02) (Figures 2B, C, Supplementary Table S7). Notably, all cytokine reductions showed relevant effect size, with Cohen’s d value ranging from -0.60 to -0.97 (Supplementary Table S7).

Finally, on the 16 original PC patients, 9 (with a drop-out of 44.75%) remained in the study. Overall, these patients showed significant reduction for various bioimpedential parameters (Figure 2); this reduction was accompanied by a significant re-duction of Eotaxin (p-value: 0.027), GM-CSF (p-value: 0.02), MCP-1 (p-value: 0.031), and CRP (p-value: 0.016). Interestingly, these patients also showed reduced levels of Leptin (p-value: 0.08), and increased levels of Adiponectin (p-value: 0.025) and Resistin (p-value: 0.039) (Figure 2, Supplementary Table S8).

We retrospectively investigated the possible associations among these diet-modified molecules with anagraphical, clinical, and nutritional variables at baseline. Univariable comparisons were analyzed between any single molecule with all the aforementioned anagraphic, clinical, and dietary factors, and results were shown in Figure 3 and Supplementary File S1.

In BC patients, univariable analysis displayed a total of 42 significant differences that were noted, with 28 showing an increasing concentration of molecules in association with a specific factor, and 14 showing reductions (Figure 3, Supplementary File S1). For instance, higher levels of IL-2 in serum of BC patients were associated with presence of T2D (p-value: 0.04); Il-7 and IL-13 were associated with presence of METS (p-value: 0.003 and p-value: 0.005, respectively), CRP was associated with higher levels of BMI and METS (p-value <0.001 and p-value: 0.004, respectively), C-Peptide was associated with higher weekly intake of Milk and dairy products (Mid vs Low intake p-value: 0.087; High vs Low intake p-value: 0.025); GIP and Insulin were likewise associated with higher intake of Milk (p-value: 0.001 and p-value: 0.018, respectively), as well as Cereals (p-value: 0.006 and p-value: 0.016, respectively); finally, Leptin was significantly associated with higher levels of patients’ BMI (p-value: 0.031). These associations were then validated via consecutive univariable and multivariable linear regression models, which confirmed their significance (Supplementary Table S9).

In CC patients, 43 significant associations were noted, with 24 reductions and 19 increases (Figure 3, Supplementary File S1). Notably, a higher intake of fruits resulted associated with reduced levels of inflammatory markers, such as IL-1β (Mid vs Low intake p-value <0.001; High vs Low intake p-value: 0.005), IL-6 (Mid vs Low intake p-value: 0.025; High vs low Intake p-value: 0.011), IL-8 (Mid vs Low intake p-value: 0.027; High vs Low intake p-value: 0.021), MIP-1α (Mid vs Low intake p-value: 0.005; High vs Low intake p-value: 0.001), MIP-1β (Mid vs Low intake p-value: 0.026; High vs Low intake p-value: 0.003), and TNF-α (Mid vs Low intake p-value: 0.018; High vs Low intake p-value <0.001). Other significant associations were noted for BMI with IL-6 (p-value: 0.019) and CRP (p-value: 0.011), for age with IL-1β (p-value: 0.026) and bFGF (p-value: 0.03), and between smoking exposure with IL-8 (p-value: 0.046) (Figure 3, Supplementary File S1). These associations retained their significance also upon univariable and multivariable linear regression Analysis (Supplementary Table S10).

Finally, in PC patients, 12 significant associations were noted, with only 3 reductions and 9 increases (Figure 3, Supplementary File S1). Eotaxin, CRP, and Leptin levels correlated with BMI (p-values: 0.034, 0.033, and <0.001 respectively); IL-12 (p-value: 0.043) and GM-CSF (p-value: 0.042) were associated with aging; CRP levels were also associated with higher intake of red meats (Low vs Mid intake p-value: 0.031), and finally, physical activity was associated with significantly lower levels of Resistin (p-value: 0.031) (Figure 3, Supplementary Table S11). These associations were subsequently confirmed through univariable and multivariable linear regression models, maintaining their significance (Supplementary Table S11).

This pilot observational study has different limitations. First, the lack of a control group and randomization, together with the absence of information on potential confounder changes between T0 and T1 (e.g., physical activity, socioeconomic status, medication changes, etc.), restricts the ability to draw causal inferences. Second, the study did not include the assessment of validated Mediterranean diet adherence scores (e.g., PREDIMED, MEDI-LITE), as well as standardized quality-of-life assessment scores (e.g., WHOQOL-BREF, QOLS, EORTC QLQ-C30), which limits comparability with other investigations. Third, methodological constraints related to the intervention should be acknowledged, including a relatively high dropout rate (47% of participants did not complete the 6-month follow-up) and the reliance on nutritionist-administered FFQ at each time point. Finally, although the study investigated circulating inflammatory and metabolic markers, the underlying molecular mechanisms remain unexplored.