Endometrial cancer (EC) is the most frequent gynecological malignancy in developed countries, with an increasing incidence rate (1, 2). The diversity of EC dictates the need for precise early diagnosis and pre-operative stratification to appropriately select the extent of surgery and lower both the recurrence rate and risk for overtreatment. A dualistic model, built mainly on the histological findings and prognosis, was introduced by Bokhman back in 1983 (3–5). Recent advances based on molecular classification stratify EC into four risk categories: POLE ultramutated, microsatellite instability hypermutated, copy-number low, and copy-number high (6, 7). This classification of endometrial cancer has been validated and incorporated in the ESMO/ESTRO risk stratification and is currently used in clinical practice to guide management decisions. However, refinements of the current classification with additional biomarkers are likely to further improve and de-escalate treatment in certain subtypes of EC (8).

Biomarkers represent a noninvasive approach for more precise stratification of various malignant diseases (9). No single serum/plasma biomarker alone has yet been classified for clinical use in the diagnostics of EC (10). Emerging findings in genomics, transcriptomics, and proteomics may present a pivotal role for noninvasive early diagnostic options (11).

Angiogenesis is the process of the formation of new vessels from the preexisting vasculature. The process may be activated due to ischemia and tissue trauma as part of normal tissue healing, or it can be one of the key processes in cancerogenesis, allowing fast growth of cancerous tissue and spreading to distant organs. Tumor tissue receives oxygen and nutrition at its very early stage via diffusion independently of the vascular network. When the tumor size exceeds 1–2 mm³, the existing capillary network becomes insufficient, causing hypoxia in solid tumors. The resulting shortage of cellular oxygen and nutrients causes the production of angiogenic factors (AFs), mostly cytokine molecules, which are secreted into the surrounding tissue and provoke the growth of new vessels (12–16). This angiogenic switch makes AFs potential biomarker candidates for early EC detection and assessment of prognosis (12, 17, 18).

In our previous research, we investigated 37 AFs as potential biomarkers for EC. AFs’ concentrations in preoperative plasma samples of patients with endometrioid EC (n = 38) and control patients with benign gynecological conditions (e.g., prolapsed uterus or myoma; n = 38) were measured using Luminex xMAP™ multiplexing technology. Our discovery study demonstrated significant differences in the plasma levels of six AFs: sTie-2, G-CSF, and leptin were present in different concentrations in EC versus control patients, and IL-8, neuropilin-1 and follistatin differed among different EC subgroups (19). The roles of these AFs have been described in detail in our recent review paper (12). In the present study, we aimed to validate our previous findings in a larger group of patients.

Recently, machine learning algorithms, a part of artificial intelligence, have evolved to help develop cancer risk stratification systems with great precision. With the aim to differentiate between patient groups, the machine learning approach simultaneously consider multiple disease-specific risk factors, which would otherwise present an impossible statistical obstacle, especially in the heterogeneous type of disease like EC (20, 21). In our study, besides basic statistical methods, we used machine learning algorithms to create a robust diagnostic model based on the plasma concentration of tested AFs.

Non-invasive diagnostic approaches invaluably assist clinicians in the detection, management and treatment of patients in daily clinical practice. Recently a proof-of-principle was provided in the area of cytological analysis for EC screening: O’Flynn et al. demontrated that endometrial cancer can be detected with high accuracy in urine and vaginal fluid (35). However, there are currently no validated diagnostic or prognostic biomarkers for EC that could accurately predict the presence and extent of disease, and thus pathohistological examination remains the gold standard for EC diagnosis.

In our previous discovery study, we investigated 37 AFs as potential biomarkers for EC. Preoperative plasma samples of 38 EC patients and 38 control patients were analyzed using Luminex xMAP™ multiplexing technology. Six out of 37 AFs were present in significantly different concentrations between groups of patients. sTie-2 and G-CSF were lower in EC compared to control patients, and plasma level of leptin was higher in EC patients. Neuropilin-1 plasma level was higher in patients with type 2 EC (G3) compared to patients with lower grade cancer or controls. Follistatin level was higher in patients with LVI, and IL-8 plasma level was higher in patients with metastases (19). In this validation study, we further evaluated those six AFs on a larger group of patients. While our previous study was limited to post-menopausal women with endometrioid EC, in the present validation study, women with other EC types were also included, regardless of their menopausal status. In the previous study, where 60.5% of included EC patients were in IA stage EC, and no patients in stage II or IV were included, we detected lower values of sTie-2 in EC patients compared to the control group. In the current validation study, we observed a similar trend in the early stages, whereas the concentration of sTie-2 significantly increased as the disease progressed, particularly in FIGO stage IV. This is in accordance with results from Saito et al., who reported higher expression of Tie-2 in endometrial adenocarcinoma than in normal epithelial cells (36).

EC often involves patients burdened with other comorbidities, such as hypertension, diabetes or obesity (37). In our study, diabetes status was balanced between EC and control patients, and only leptin differed among patients when stratified according to hypertension status or obesity. The effect of BMI difference between groups was intensely further investigated through machine learning modelling.

Obesity is an established risk factor for EC and presents a larger risk for this malignancy than any other cancer type (38–41). Adipose tissue is an endocrine organ which synthesizes adipokines—biologically active substances participating in cell growth and differentiation, apoptosis, angiogenesis, and carcinogenesis. Leptin is among the most important adipokines during EC development, and numerous recent studies have already proposed it as a new candidate marker in determining the potential risk of EC (42–45). In accordance with the literature (46–50), our study showed higher leptin levels in EC patients, particularly in Type I EC, at higher stages and in patients with present metastasis or MI and LVI.

During carcinogenesis, leptin promotes tumor angiogenesis. Early studies showed that endothelial leptin receptor Ob-R generates a growth signal involving a tyrosine kinase-dependent intracellular pathway, promoting angiogenic processes (51, 52). Furthermore, leptin induces proliferation, migration, and invasion and suppresses apoptosis of cancer cells (42, 53). However, cancerogenic potential of leptin has been shown to differ between different tissues (54, 55). In EC, it has been shown in in vitro studies that leptin inhibits the apoptosis of endometrial carcinoma cells through activation of the nuclear factor κB-inducing kinase/IκB kinase pathway (49) and stimulates endometrial carcinoma cell proliferation via enhancing P450arom expression and estradiol synthesis (56).

Leptin has an important role in the regulation of energy balance and glucose metabolism and is considered to play an important part in the link between obesity and EC (57). There is still an ongoing debate in the literature on whether the effect of leptin on EC risk is related to higher BMI or whether it is an independent risk factor for EC. Wang et al. (58) performed a meta-analysis of 6 preceding studies and found that after adjusting for BMI, leptin was still associated with an increased risk of EC. On the other hand, another meta-analysis performed by Ellis et al. observed that BMI and diabetes appeared to affect the association between leptin levels and EC risk (43). Some other studies found a strong positive correlation between patients’ BMI and serum leptin levels (46, 50), while a recent Mendelian randomization analysis showed a causal effect of BMI on EC but failed to find evidence for leptin to be causally implicated in EC risk (59).

By utilizing an automated machine learning approach and comparing univariate models applying BMI and leptin as predictor variables, our research indicates that leptin might be able to predict EC better than BMI. This would seem to be further indicated by the results of the automated feature prediction selection described in the methodology, which identified a combination of age, IL-8, leptin and G-CSF, but not BMI, as the most important features for multivariate model building and the performance of the corresponding model utilizing those features. This supports findings that leptin might be involved in EC development via pathways beyond obesity-related pathophysiology, including through angiogenesis (51, 52).

However, as always with machine learning models, this should be interpreted with care. Due to the relatively small data set, we observed overfitting on univariate models, and classification close to random on the test dataset could simply mean the model picked features that were not well generalizable. On the other hand, leptin-based multivariate models performed significantly better than random chance on the test dataset with an AUC very similar to the one reached on the training dataset. This indicates that the models were able to generalize and could be useful for predicting the risk of EC even on unseen data. Therefore, these models show good diagnostic characteristics, with AUC in the range from 0.69 to 0.89.

Another link between adipose tissue and EC is through chronic low-grade inflammation resulting from increased secretion of pro-inflammatory molecules such as cytokines and chemokines (60). IL-8 is a pro-inflammatory cytokine secreted by adipocytes, with well-defined functions in tumor-associated inflammation. It is chemotactic for lymphocytes and neutrophils and also has an important role in angiogenesis (61, 62). Its levels increase with BMI and waist circumference, which is also associated with EC incidence and outcomes (34, 63). Kotowicz et al. (64) demonstrated the clinical usefulness of IL-8 measurements as potential prognostic factors in type 1 EC, where elevated pretreatment IL-8 serum levels were independently associated with shorter disease-free and overall survival.

While in our study, leptin was significantly increased in type I EC, IL-8 was higher in type II EC. Its level also increased through higher grades, and it was particularly elevated in poorly differentiated EEC G3, however the difference between G1/G2 and G3 cases was not statistically significant. We assessed EEC G3 cancers together with type II tumors since several important reports have strongly demonstrated that high-grade endometrioid cancers have molecular characteristics, risk factors, clinical behavior, and prognosis overlapping with those of non-endometrioid cancers (65, 66). We demonstrated that higher levels of IL-8 are suspect for higher-grade tumors. However, there were only 10 G3 EC patients in our cohort. If this result could be validated on a larger set of patients, a simple blood test for IL-8 prior to the surgery might lead to a better stratification for the extent of surgery.

The angiogenic switch occurs early in the process of cancerogenesis. Correspondingly, we observed indicated trend for leptin, sTie-2 and IL-8 progression through early EC stages, which illuminates how growing tumor mass dictates a higher need for additional oxygen and nutrient supply and accelerates angiogenic activity. Nevertheless, due to low number of cases in each EC stage, this should be taken with care and studies on bigger cohort are needed to confirm this.

We also showed a statistically significant correlation between EC myometrial invasion and IL-8 levels. This was previously reported by Fujimoto et al., who suggested IL-8 might act as an angiogenic switch in myometrial invasion in stage I EC (67). According to the currently valid guidelines (10), LVI, among other histopathological characteristics, is the cornerstone of risk stratification. In our study, IL-8 levels significantly increased during LVI, which further demonstrates the role of IL-8 in the angiogenesis of EC. Since IL-8 is also associated with obesity, the prognostic value of IL-8 was compared to BMI based on the presence of MI and LVI. Plasma IL-8 level was a better prognostic biomarker than BMI in terms of EC patient stratification according to both MI and LVI. Levels of IL-8 were also higher in patients with present metastasis, which is in accordance with our previous results (19) and in vitro study on cell lines (68).

Due to the relatively low number of patients with LVI and MI in our dataset, the IL-8 link to MI and LVI could not be analysed using the machine learning approach; however, the data presented indicate that it would be a good candidate for a larger study with larger and more balanced datasets.

In our study, we also evaluated neuropilin-1, follistatin and G-CSF as potential biomarkers for EC. While they were individually able to differentiate between EC patients and control patients (with the exception of neuropilin-1), the univariate models utilizing those AFs did not generalize well to unseen (test) data (so-called overfitting). However, when used in conjunction with other data, the AFs significantly improved the classification capabilities of models, and the model utilizing a combination of all AFs, BMI and age reached a ROC AUC of 0.89 on both the training and test dataset, strongly indicating the usefulness of the combination of AFs.

The best multivariate model on the training set (“Selected Features”) has also proven to be robust on both the training and test datasets, where it revealed good diagnostic characteristics with the ROC AUC of 0.94 and 0.81, respectively. It also greatly outperformed individual results of univariate models for included AFs, i.e. leptin, IL-8 and G-CSF.

Our study has also confirmed the importance of existing, well-known risk factors, namely age and BMI, and the values for both were significantly different between the EC and control patients (69, 70). To confirm AFs can valuably add to diagnostics of EC, we created and tested models using only the variables that are – at most – weakly correlated with known risk factors. Using those models, we have confirmed that even without knowing age, BMI or hypertension status, we can reach relatively robust results using only AFs (71% accuracy on the test dataset using AFs without leptin). Nevertheless, combining AFs with the existing risk factors yielded better results (80% accuracy on the test dataset), which confirms the value of already established risk factors in extended models.

Finally, the fact that models utilizing less data sometimes outperformed models utilizing more data might seem counterintuitive but would seem to hint at some combinations of factors containing more noise and thus causing overfitting. In our study, due to noise generation, models using a combination of both BMI and leptin performed weaker than models with either BMI or leptin alone, indicating some collinearity between them. Consequently, a larger study providing more data would be useful to fine-tune the developed diagnostic models and determine the minimal subset required to achieve good classification results.

Our study has the following limitations. First, relatively low number of patients in different EC subgroups – in relation to presence of LVI and MI, EC stage and grade, and EC subtypes. Second, the control group was chosen randomly among women with a prolapsed uterus or myoma who were admitted for surgical procedure, whereas ideal control group for endometrial cancer biomarker discovery would be symptomatic post-menopausal women. In our study patients in control group were relatively younger and with less additional comorbidities than patients with EC. Since these limitations may affect the generalisability of the study conclusions, the results of this study should be validated on bigger cohort in a multicentric clinical study.

To conclude, AFs – especially leptin and IL-8, represent valuable biomarkers candidates with a potential for early diagnostics and risk stratification of EC. Leptin represents a candidate for a diagnostic biomarker of EC, while IL-8 might be valuable in EC patient stratification according to prognostic characteristics, e.g. LVI, MI and EC grade. Other AFs further increased the performance of multivariate models. As revealed through the machine learning approach, the plasma concentrations of AFs, in conjunction with other clinical data, show good diagnostic characteristics. They could, provided they are confirmed in a large-scale multicentre validation study, represent a valuable supplementary diagnostic tool for EC’s early detection and prognostic characterization. This could guide the decision-making regarding the extent of surgery and the choice of adjuvant therapy for EC.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the National Medical Ethics Committee of the Republic of Slovenia (No. 0120-515/2017/4 and 0120-541/2019/7. The patients/participants provided their written informed consent to participate in this study.

LR: Conceptualisation, Data curation, Writing, Result Analysis, Discussion. MP: Methodology; Data analysis, Data curation. IR: Data curation, Writing, review & editing, Statistics, Modelling. MK: Data curation, Statistics, ML Modelling, Results, Writing – review & editing. BP: Methodology; Data curation. ŠS: Conceptualisation; Writing –review & editing; Resources; Supervision. TR: Conceptualisation; Writing – review & editing; Resources; Supervision. All authors contributed to the article and approved the submitted version.

This research was funded by two grants: the Slovenian Research Agency grants J3-2535 to TLR; and University Medical Centre Ljubljana tertiary project TP 202110160 to ŠS.

The authors would like to thank study participants who kindly donated their samples and time. We also thank the personnel from the Division of Gynecology and Obstetrics, University Medical Centre, Ljubljana, especially Milena Osredkar, Dr. Leon Meglič, MD, and Vesna Sekelj Rangus, for their help and support in enrolling the study participants and sample collection. We thank Vera Troha Poljančič and Prof. Dr. Joško Osredkar, from the Clinical Institute of Clinical Chemistry and Biochemistry of the University Clinical Centre Ljubljana, for processing and storing the samples according to our standard operating procedures.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.