Endometrial carcinosarcoma is a biphasic tumor with both carcinomatous (epithelial) and sarcomatous (mesenchymal) elements (Raffone et al., 2022). It is a rare, aggressive high-grade endometrial cancer, accounting for about 5% of all uterine cancers and nearly 20% of non-endometrioid endometrial carcinomas (Siegel et al., 2022). Although non-endometrioid tumors make up 10–20% of endometrial malignancies, they are responsible for over 40% of endometrial cancer deaths (Lu and Broaddus, 2020).

The treatment can be complex, including the need to perform surgery, platinum-based chemotherapy, and radiotherapy. Despite this, the prognosis remains poor (Travaglino et al., 2022). Median overall survival is less than 2 years, and the 5-year overall survival is under 30% (approximately 50% for early stage and 20% for advanced stage disease). Even patients with early-stage disease have a 45% 5-year recurrence rate and 50% 5-year disease-specific mortality (Toboni et al., 2021). The rising incidence and poor outcomes of endometrial carcinosarcoma highlight an unmet need of personalising the management of these challenging patients, in order to allow more informed and targeted decision-making even in the presence of a complex prognosis (Grasso et al., 2017).

Over the past few years, due to the increased availability of data and the greatest computing power, Artificial Intelligence (AI) has emerged as a potential tool to deal with this big data, with the aim of optimizing cancer research, improving clinical practice, and promoting precision in healthcare (Farina et al., 2022). Specifically, Machine Learning (ML) is a subfield of AI which exploits mathematical and statistical approaches to develop learning models able to detect hidden patterns in the data and to improve model performance (Cuocolo et al., 2020). Anyway, ML models are often related to the concept of “black box” whereby even the operators who engineered the model cannot explain the reasoning behind their predictions. Hence, eXplainable Artificial Intelligence (XAI) algorithms have been designed to enable users to understand and appropriately trust ML model decisions (Gunning et al., 2019; Doshi-Velez and Kim, 2017).

Recently, ML methods have been also employed in the identification of prognostic factors capable of predicting cancer survival and the risk of disease recurrence, with the purpose of supporting clinicians in optimizing the clinical follow-up plan (Wang et al., 2019). The ability of ML methods to handle survival data by modelling the relationship between the event of interest and the predictor variables, could allow the design of personalized therapeutic options, showing better accuracy than conventional statistical approaches (Mazzaki et al., 2021). As a matter of fact, ML algorithms are able to capture and shape non-linearities between variables without having to specify the form of the relationship between them a priori.

So far, in the state-of-the-art, several ML models able to predict early diagnosis, as well as response to therapy and disease recurrence in the gynaecological cancers field have been proposed (Fiste et al., 2022; Sheehy et al., 2023; Akazawa and Hashimoto, 2021). However, there is a lack of ML models to predict recurrence-free survival in patients affected by endometrial carcinosarcoma. To the best of our knowledge, in this study, we proposed the first explainable ML approach which addresses this task exploiting the predictive power of clinical and histopathological data, as well as chemotherapy and surgical information of 80 endometrial carcinosarcoma patients.

At the end of each cross-validation round, features were ranked in descending order according to their estimated significance, and only features with a rank ≤3 were selected. The frequency of selection of each feature within all rounds of the cross-validation procedure is shown in Figure 1. Four variables, namely, myometrial invasion, omentectomy, surgery and histotype were always selected, presenting a frequency equals to 100%. Conversely, the age at diagnosis has never been selected as an important feature over the training process.

The designed ML survival model, trained on features resulted important by turns, was able to well describe the observed sequence of events, and its discrimination power was evaluated in terms of C-index along with its 95% confidence interval: 70.00% [59.38–84.74].

The model predictions also allowed to describe both a survival and a cumulative hazard function for each patient. The respective functions estimated by means of the best performing model are depicted in Figure 2. Due to the specularity of these functions, in both cases the curves resulted well separated into two groups. The first group of patients was characterized by a lower survival probability and, consequently, a higher risk of recurrence since the first months after diagnosis. Conversely, patients belonging to the second group were identified by a survival probability always greater than 80%, and which trend remained constant even after several months after diagnosis. A comparison highlighted that patients with a higher risk of recurrence all share the following feature values: a heterologous MMMT type, a CBDCA CT scheme and a laparotomic-LPT surgery performed.

Despite the good performances of our ML survival model in predicting recurrence-free survival, the reasoning behind its predictions is unknown. To this aim, we provided local explanations of the contribution of important features to the model prediction, both at patient and all dataset level. Figures 3, 4 shows some examples of explanations at patient level. Each explanation consists of a feature importance diagram in which the feature contributions to the outcome are displayed in descending order, using a red colour palette for the features that increase the Cumulative Hazard Function and a blue palette for those that decrease it.

Considering Figure 3, for both patients the ML survival model correctly returned a high predicted hazard. This can be related to having performed the lymphadenectomy, a CBDCA CT scheme and a laparotomic-LPT surgery. On the other hand, patients illustrated in Figure 4 are related to a correctly low predicted hazard. In these cases, explanation diagrams highlighted how differences among their feature values are related to different feature contributions in terms of weight to the outcome prediction.

Lastly, Figure 5 depicts the local explanation at all dataset level. In this feature importance diagram, feature contributions are displayed in descending order by means of box plots representing the feature contribution distributions computed over the entire dataset. The feature positive or negative contribution in increasing the Cumulative Hazard Function is pictured by a red colour or blue colour palette, respectively.

Uterine carcinosarcoma is a high-grade tumor including both epithelial and mesenchymal malignant cell components. Typically, the former shows low differentiation and a mix of characteristics, possibly displaying traits like endometrioid, clear-cell, or serous features. Tumor cells may organize in gland-like structures. The latter can resemble either endometrial stromal sarcoma or leiomyosarcoma, known as “homologous,” or it may exhibit features akin to specialized connective tissues outside the uterus, such as muscle, cartilage, and bone, termed “heterologous.” In both scenarios, angiolymphatic invasion is frequently observed (Bogani et al., 2023).

Despite surgical treatment and timely adjuvant multimodal therapy, more than half of the cases of endometrial carcinosarcoma will recur within the first 2 years (Concin et al., 2021). The management of the recurrent disease is highly personalized and should consider several factors, such as the performance status of the patient, the size and sites of recurrences, and prior therapies (Pezzicoli et al., 2021). Importantly, it depends on whether the relapse is locoregional, oligometastatic, or disseminated and, second, on whether the patient has already received radiotherapy, as radiotherapy rechallenge is generally avoided for safety reasons. Again, the best treatment approach is multimodal. Patients with recurrent disease (including peritoneal and lymph node relapse) should be considered for surgery only if it is anticipated that complete removal of macroscopic disease can be achieved with acceptable morbidity and be treated in specialized centres (Beckmann et al., 2021). External beam radiotherapy can be used in radiotherapy-naïve patients or those who had received only prior vaginal brachytherapy. Immunotherapy (with or without tyrosine kinase inhibitor) is the emerging preferred second-line systemic treatment. After the failure of immunotherapy, chemotherapy alone (generally mono-chemotherapy) is the preferred treatment in cases of disseminated metastases (Abu-Rustum et al., 2021).

Owing to the rare and aggressive nature of endometrial carcinosarcoma, the complexity of its management both at diagnosis and recurrence, as well as its high recurrence rate (Amant et al., 2005), identifying an ensemble of prognostic factors able to accurately predict recurrence-free survival in patients affected by this malignancy could allow more informed and targeted decision-making, enabling proactive clinical management even in the presence of a complex prognosis. Actually, the ability to identify patients at a major risk of recurrence at an early stage could allow clinicians to tailor treatment plans, both adopting more aggressive strategies such as intensive or combination chemotherapy protocols, adjuvant radiotherapy or experimental approaches in patients with a high probability of recurrence and sparing low-risk patients from unnecessary interventions. Moreover, an accurate predictive model can guide the frequency and intensity of clinical follow-up, allowing an early detection of recurrence and improving the likelihood of disease control. Finally, thanks to predictive modelling, it is possible to more selectively enrol patients which may be candidates for clinical trials of new drugs or experimental therapies, especially when standard options have obvious limitations.

Over the past years, several efforts have been made to develop a greater awareness and deeper understanding of endometrial carcinosarcoma pathogenesis, with the purpose of identifying new targeted therapies and providing specific guidelines for the management of this tumor (Bogani et al., 2023). Besides, endometrial cancer treatment has provided new changes by incorporating biological, clinical, genomic, and clinico-pathologic characteristics of the women affected by this tumor, and recent studies showed that molecular targets such as L1CAM (L1 cell adhesion molecule) plays an important role as prognostic factor and could provide a potential useful tool for tailoring the need of adjuvant therapy (Giannini et al., 2024; Vizza et al., 2021). As well, a prognostic nomogram to predict the overall survival rate in endometrial carcinosarcoma patients by exploiting lymph-node metastasis information has been proposed (Gao et al., 2021).

However, there is a lack of research studies focused on the prediction of disease recurrence risk.

In this study, we proposed the first explainable ML method designed to predict recurrence-free survival in patients affected by endometrial carcinosarcoma. The nested feature importance approach allowed us to identify the most relevant variables for this outcome prediction. Accordingly, promising results were achieved in providing a reliable ranking of the survival times based on the individual risk scores (C-index: 70%). Finally, with the aim to enable clinicians to understand the reasoning behind the ML model predictions, the implemented XAI algorithm computed the contribution of each of the most significant features to the model decision, both at patient and at all dataset levels.

To conclude, the proposed explainable ML model represents the first effort in devising an artificial intelligence-based tool to be enclosed in clinical practice to support clinicians in discriminating between endometrial carcinosarcoma patients at low-risk or high-risk of recurrence in a non-invasive and inexpensive way, also providing an intelligible explanation on how the clinical characteristics considered for those patients contributed to the estimated risk. Accordingly, the ability of this model in detecting the risk for a patient of experiencing recurrence could aid clinicians to personalise therapeutic options, by candidating high-risk patients to adjuvant chemotherapy and saving low-risk patients from unnecessary aggressive treatments.

Besides, a limitation of our study deals with its retrospective design and the limited dimension of the dataset. As far as the limited dataset size, this could affect the robustness and generalizability of the ML model which generally require larger dataset, in contrast to what classic survival approaches, such as Cox regression, need. However, the advantages in exploiting ML techniques rather than classical methods are the increased flexibility, the ability to adapt to non-linear relationships, together with improved predictive performances. Actually, relationships between variables are often non-linear or complex, and some effects may depend on interactions between them. ML algorithms are able to capture and shape these non-linearities and interactions without having to specify the form of the relationship between variables a priori. Definitely, employing ML models with larger datasets, it could be possible to achieve higher performances and improve the model. For this purpose, in our future work we will collect an external dataset for prospective validation, in order to establish documented evidence that the model is able to consistently produce the desired results within predetermined specifications and quality attributes.