Cancer is one of the deadliest diseases in the world. Biologically, it includes a collection of diseases in which the cells of the body start dividing uncontrollably without stopping. Normally, human cells grow and duplicate according to physiological needs and their number is controlled by the apoptotic process. In pathological conditions such as cancer, however, cells grow abnormally, old and damaged cells are not eliminated, and new cells duplicate uncontrollably. Malignant tumors invade surrounding tissues and compete with normal cells for nutrients. Additionally, these types of tumors can metastasize to new sites, often distant from the primary tumor site, through the blood or lymphatic system (1, 2). According to the American Cancer Society annual report, 1,918,030 new cases of cancer were recorded in 2022 resulting in 609,360 deaths (3). The most common site of cancer development is the digestive system, and 343,040 new cases were discovered in that same year. The 5-year survival rate for all cancers has grown dramatically since the early 1960s, reaching nearly double, from 39 to 70% among white individuals and triple, from 27 to 63% among black patients. In particular, for gastrointestinal tumors, interest is now focused not only on the tumor growth rate but also on the cellular microenvironment, which is responsible for tumor initiation, progression and metastasis (4, 5). One of the many causes of this dysregulation is certainly an incorrect diet (Figure 1), lifestyle, and genetic background (6) which leads to an ever increasing rate of obesity and continues to remain the “elephant in the room” (7, 8).

Dietary intakes of macro- and micronutrients have been implicated in the etiology of chronic disease as gastrointestinal cancer (9, 10). Several studies have shown that high dietary intakes of macronutrients (as fat, protein and carbohydrate) may have a positive association with the risk of developing of cancer (11). Additionally, fruits and vegetables as sources of dietary fiber, folate, vitamin C, and a lot of phenols, and flavonoids could protect, because are involved on trapping free radicals and reactive oxygen molecules at the cellular level [induced lipid peroxidation has been implicated in malignant transformation (12)], thus acting as a protective mechanism against the oxidative damage during digestion process (13, 14). However, micronutrients with a significantly reduced intake also are involved in other kind of cancer, as the breast cancer risk in obese women (15).

The obesity, which is a slow and progressive condition, is mainly based on an ever increasing fatty acid intake vs. low amounts of fiber, vitamins and minerals, widely present in fruit and vegetables (16, 17). For example nuts contain magnesium, unsaturated fatty acids, potassium, fiber, and vitamin E (18), while dairy products and eggs (19) contain a well-balanced composition of micronutrients that could help for the prevention of hypertension (20). However, also consumption of particular type of vegetable food as N. sativa could reduce oxidative stress and inflammation with various mechanisms including a reduction of lipid peroxidation via its antioxidant properties; agonist of PPAR-γ in adipose tissue; activation of AMPK, increased antioxidants inhibition of NF-κB pathway (21, 22).

The obesity rate associated with the development of gastrointestinal cancer (GC) is estimated to rise by about 18% overall by 2025, affecting 18% of men and 21% of women. Studies have shown that weight disorders such as overweight and obesity are usually associated also with sleep problems, and some studies have shown that adherence to DASH diet has helped to correct this trend (23). Furthermore, these disorders could be associated with the Non-alcoholic Fatty Liver Disease (NAFLD) dyslipidemia, and insulin resistance (24).

One of the many opportunities available to study this increase, as well as the classical clinical approaches, is to apply machine learning (ML) as a basis on which to predict cancer survival. A huge amount of research has been conducted across a wide range of statistical approaches. Classical survival analysis is a well-established methodology for estimating a patient’s probability of survival. Although this is now validated, research is also focused on identifying new methods for prediction and calculating their accuracy. In this work we aim to evaluate survival of GC patients using the classical statistical method, Cox Proportional Hazards (CPH) and compare it with new generation algorithms such as Random Survival Forest (RSF), and Survival Support Vector Machine (SSVM), together with a new approach to quantify the power of parameters’ prediction on GC, denominated Shapley Additive Explanation (SHAP) (Table 1).

The use of ML approaches proved to enhance the classical statistical methods, improving cancer diagnosis, prognosis, detection and prediction (25), thus providing researchers with powerful tools. The use of classical approaches, ML methods and SHAP is helpful to solve the problem of the ML algorithm, because it is based on the “black-box” concept (26).

We randomly subdivided patients and allocated 75% to the training set, and the remaining 25% to the test set.

The demographic and food habits are reported in Table 2 for the total cohort. Mean age was 65.01 ± 8.76 years and 38.26% of patients were male.

Table 3 shows the difference between alive and dead GC patients. Patients who were older, male, separated, divorced or widower, and with lower education levels were more represented than the others, likewise patients with diabetes and myocardial infarction (p < 0.001 respectively). Lower intakes of both macro- and micronutrients were confirmed in dead patients, where all comparisons resulted statistically significant (p < 0.05).

Figures 2, 3 describe the traditional Cox model adjusted for age and gender in the training set. For macronutrients (Figure 2) only alcohol intake was associated with risk of mortality, with a low protective role (HR = 0.99, p = 0.008, 0.98–1.0, 95% C.I.), while for micronutrients intake (Figure 3) iron intake had a protective role against mortality (HR = 0.88, p = 0.002, 0.81–0.95, 95% C.I.).

Figures 4, 5 describe the importance plot for the RSF for the training set; in both models (Figures 4, 5), alcohol and iron intake, as also the Cox model, had the most important values, contributing to the good prediction of mortality. To identify the features that influenced the prediction, the SHAP summary plot was used (Figures 6, 7). Each point corresponds to a single patient in the testing set. The position on the x-axis, i.e., SHAP value, was the impact of the variable on the model output. Statistically, this represents the logarithm of the mortality risk. Patients with higher SHAP values had a higher risk of death, while lower SHAP values had lower risk. For macro- and micronutrients (Figures 6, 7, respectively) except the time and age variables, all patients had SHAP value near to 0, except for fatty acids (for macronutrients) and Sodium intake (for micronutrients), when more patients had higher SHAP values.

Figures 8, 9 depict the quantification of the C-Index in the graph bar. Among macro- and micronutrients CPH and RSF had similar values (0.7649284 and 0.7653711, 0.764206 and 0.7725246, respectively), but the values for SSVM were statistically different for each statistical technique (0.5667753 and 0.545222, respectively, p < 0.0001).

In this study we enhanced predictive performances using different survival statistical techniques in an elderly cohort.

Alcohol consumption had a protective role on mortality in our cohort, unlike in the literature (43), probably due to wine consumption rather than spirits. Red wine is a source of polyphenols, such as resveratrol and anthocyanin, with demonstrated antioxidant properties and protective role against cardiovascular diseases (44–46).

In the same way, iron, like alcohol, had a protective role and a high impact on mortality both in the CPH model and RSF (47). Among these methods, RSF demonstrated the best performance followed by CPH, as reported in the literature (48) while the SVM showed the worst performance because generally used for a wide array of applications for its high potential in data transformation and geometric structure (49, 50). The results reveal a promising ML potential for predicting mortality risk in clinical practice. SHAP values on the summary plot were an interesting way to illustrate key features and a good method for describing the power of prediction of outcome. RSF followed by CPH are known to be the best for predicting mortality, while the use of SHAP is still rarely used. The power of SHAP was to allow a description of the single patients variability and also of how individual patients were distributed in terms of SHAP values. The advantage of using SHAP was to overcome the problem of interpreting the classical ML techniques, widely described in the literature regarding the prediction of macro- and micronutrients, and to couple it with other techniques to improve the prediction (51).

The main limitation of this study is the use of an elderly cohort. Mathematically, the use of larger cohorts will surely lead to different results and therefore external validation could be necessary.

Our report can be seen as a first step toward evaluating personalized nutrition in an elderly cohort, and studying the relationship between macro- and micronutrients intake and mortality with different approaches.

Understanding eating habits and therefore the correct intake of macro and micronutrients will be useful for improving the lifestyle of people with obesity who also suffer from other disabling pathologies (18, 24).

This suggests that combining different techniques could be useful in order to personalize the ideal dietary intake. Furthermore, a future follow-up project will be to use these tools to predict changes in food habits in the Southern Italian population. The promise of using ML tools to achieve nutritional phenotyping needs to be explored further in order to set up standard paradigms based on different epidemiological parameters.

Precision nutrition directly addresses metabolic heterogeneity and may serve as a treatment for obesity and other metabolic diseases. Future interventions should examine ways to increase dietary self-monitoring adherence and intervention exposure and consider the development and testing of a specific predictive algorithm (52).

The results of this study demonstrate that both the CPH and RSV methods performed well, but the combination with SHAP was best. It is difficult to draw conclusions about preferring one model over the other based on this study and the others in the literature, because each model had advantages and disadvantages. In clinical practice none of the mathematical models described could replace another, but all should be used together to make future decisions. In the last years, Artificial intelligence offers major opportunities to improve public health management.

The datasets presented in this article are not readily available because IRCCS “S. de Bellis” Hospital property. Requests to access the datasets should be directed to corresponding author.

The studies involving human participants were reviewed and approved by the National Institute of Gastroenterology “S. de Bellis,” Research Hospital, Castellana Grotte (BA), Italy. The patients/participants provided their written informed consent to participate in this study.

RD and ML: conceptualization and writing—original draft preparation. RD: methodology, software, formal analysis, investigation, and data curation. RD, ML, and PP: writing—review and editing. LP: supervision. VG: methodology. All authors have read and agreed to the published version of the manuscript.