Researchers have introduced a new tool called DeepMerkel, a hybrid machine-learning model designed to predict survival outcomes for patients with Merkel cell carcinoma (MCC).
Study: A hybrid machine learning approach for the personalized prognostication of aggressive skin cancers. Image Credit: David A Litman/Shutterstock.com
Published in Nature, the study outlines how this web-based tool combines deep learning with explainability analysis to deliver personalized, time-dependent survival predictions. DeepMerkel has been shown to outperform traditional staging systems, offering a significant improvement in MCC prognostication.
The Challenge of MCC Prognosis
Merkel cell carcinoma is the most aggressive form of skin cancer, often presenting in advanced stages and disproportionately affecting elderly patients. Current prognosis relies heavily on the American Joint Cancer Committee (AJCC) staging system, which focuses on tumor size and spread. While widely used, this system has limitations—it provides population-level estimates, doesn’t differentiate between disease-specific survival (DSS) and overall survival, and often underestimates outcomes due to its inability to account for competing risks.
Machine learning (ML) has previously been explored to improve cancer prognostication, but existing models have struggled to address the unique challenges posed by MCC. DeepMerkel fills this gap, leveraging a hybrid ML framework to analyze complex data and provide accurate, personalized survival predictions.
How DeepMerkel Works
The DeepMerkel model was developed using patient data collected from two major sources: the NIH Surveillance, Epidemiology, and End Results (SEER) database in the United States and 30 hospitals within the UK’s National Health Service (NHS). The dataset included MCC cases diagnosed between 2000 and 2020.
To create the DeepMerkel model, researchers employed advanced data processing techniques and machine learning frameworks. They used deep learning and a modified version of extreme gradient boosting (XGBoost) to handle the complex data requirements of the study. Missing values, a common challenge in large datasets, were addressed using neural network-based imputation methods, ensuring that the analysis remained accurate and reliable.
During the model training phase, DeepMerkel was designed to predict time-to-event outcomes, specifically focusing on DSS. This was achieved by integrating TabNet and XGBoost frameworks. TabNet allowed the model to process tabular data efficiently while preserving important feature relationships, while XGBoost contributed to extracting meaningful patterns from the data. Together, these methods enabled the model to deliver precise and insightful predictions.
To make these advanced predictions accessible to clinicians, the researchers also developed a user-friendly, web-based tool called the DeepMerkel survival calculator. This tool allows healthcare providers to input patient-specific data and receive real-time DSS predictions, offering a practical and personalized approach to supporting clinical decision-making.
Key Findings
The DeepMerkel model consistently demonstrated superior performance compared to traditional systems like the AJCC staging system and other machine learning models.
In terms of predictive performance, the model achieved an area under the receiver operating characteristics (AUROC) score of 0.89 in the US cohort and 0.81 in the UK cohort. These scores were significantly higher than the 0.55 achieved by the AJCC staging system and the 0.68 recorded by an ML staging model.
The model also excelled in its concordance index (C-Index), a metric that reflects the reliability of survival predictions. DeepMerkel achieved an impressive C-index of 0.93, far surpassing the AJCC system’s score of 0.51 and the DL staging model’s 0.67.
Additionally, DeepMerkel demonstrated excellent calibration, with a Brier score of 0.053, indicating minimal deviation between predicted and observed outcomes. This level of accuracy enhances confidence in the model’s ability to make reliable predictions.
The study also identified key predictors of survival, including lymph node involvement, distant metastasis, patient age, and tumor location. These critical factors were analyzed using Shapley additive explanations (SHAP), which provided clear visualizations of how each feature influenced individual predictions. This transparency in the model's decision-making further underscores its value as a tool for personalized prognostication.
Why DeepMerkel Stands Out
DeepMerkel distinguishes itself from traditional staging systems by incorporating both staging and non-staging features into its prognostication process. While conventional models like the AJCC staging system primarily focus on tumor size and spread, DeepMerkel goes further by considering additional factors such as patient age, tumor location, and the extent of metastasis. These non-staging features, often overlooked in traditional approaches, were shown to have a significant impact on survival outcomes, enabling a more nuanced and accurate prediction.
One of DeepMerkel’s key innovations is its web-based survival calculator, which allows clinicians to input specific patient details and receive real-time, personalized survival predictions. This tool doesn’t just estimate survival likelihood; it also predicts time-to-death, providing critical insights that can guide therapeutic decisions. By offering tailored, actionable data, DeepMerkel empowers healthcare providers to make more informed and precise treatment plans, elevating the standard of care for patients with Merkel cell carcinoma.
Challenges and Future Outlook
While DeepMerkel is a major step forward in predicting outcomes for Merkel cell carcinoma (MCC), it’s not without its limitations. For instance, the model doesn’t yet include key biomarkers like polyomavirus status or immune response—factors that are known to play an important role in how MCC progresses. Adding these into the mix could make the predictions even more accurate. The team behind DeepMerkel plans to address this in future research, along with testing the tool in real-world clinical trials to validate its performance further.
That said, DeepMerkel is already a game-changer. By combining advanced machine learning frameworks like TabNet and XGBoost, it offers accurate, personalized predictions that go far beyond what traditional staging systems can provide. And with its web-based survival calculator, clinicians have a practical, user-friendly tool for making more informed decisions about treatment plans in real time.
As researchers work on refining the model and incorporating additional data, DeepMerkel has the potential to truly transform how we manage aggressive cancers like MCC. By delivering personalized insights and helping tailor treatments to individual patients, tools like this can improve outcomes and raise the bar for cancer care.
Journal Reference
Andrew et al., 2025. A hybrid machine learning approach for the personalized prognostication of aggressive skin cancers. Npj Digital Medicine, 8(1). DOI:10.1038/s41746-024-01329-9 https://www.nature.com/articles/s41746-024-01329-9
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.