A smarter approach to Prognosis: AI's groundbreaking role in skin cancer prediction

While widely used, the AJCC system focuses on overall survival rather than disease-specific survival (DSS), overlooking competing risks. It ignores non-staging factors like patient demographics, socioeconomic conditions, and tumor biological features and provides generalized predictions that fail to account for the unique complexities of individual cases. These limitations necessitate a paradigm shift toward precision tools capable of delivering granular, patient-specific prognoses.

Skin cancer, while common, has subtypes like Merkel cell carcinoma (MCC) that are aggressive and fatal. MCC’s case fatality rate is over twice that of melanoma, making its prognosis and management challenging. Traditional tools for predicting outcomes, such as the American Joint Cancer Committee (AJCC) staging system, rely heavily on basic anatomical features like tumor size and spread. These systems often fall short in providing personalized insights critical for guiding treatment decisions.

A groundbreaking study titled “A Hybrid Machine Learning Approach for the Personalized Prognostication of Aggressive Skin Cancers”, authored by Tom W. Andrew, Mogdad Alrawi, Ruth Plummer, and colleagues, and published in npj Digital Medicine, introduces DeepMerkel, a hybrid machine learning (ML) model. This innovation addresses the prognostic limitations in MCC by leveraging data-driven insights and personalized predictions.

The challenges of Merkel Cell Carcinoma

MCC is the most aggressive form of skin cancer, often presenting at advanced stages, particularly in elderly patients. This disease is associated with dismal survival rates due to its rapid progression and limited therapeutic options. Prognostication is primarily performed using AJCC’s staging system, which categorizes patients based on tumor size (T), lymph node involvement (N), and metastasis (M).

While widely used, the AJCC system focuses on overall survival rather than disease-specific survival (DSS), overlooking competing risks. It ignores non-staging factors like patient demographics, socioeconomic conditions, and tumor biological features and provides generalized predictions that fail to account for the unique complexities of individual cases. These limitations necessitate a paradigm shift toward precision tools capable of delivering granular, patient-specific prognoses.

DeepMerkel: A New Era of Prognostication

The study employed a hybrid ML approach to develop DeepMerkel, a model designed to enhance prognostic accuracy and generalizability in MCC. Combining deep learning for feature selection with XGBoost for predictive modeling, DeepMerkel was trained and validated using data from the NIH SEER database (11,342 patients) and the UK NHS dataset (121 patients).

The key components of DeepMerkel include:

Advanced feature selection

• DeepMerkel integrates both staging and non-staging features, including patient demographics, socioeconomic factors, and tumor characteristics.
• Non-staging features such as marital status, income level, ethnicity, and tumor invasion depth were identified as significant predictors of DSS.

Data imputation and robustness

• Missing data were addressed using neural network-based iterative imputation, ensuring accuracy and reducing bias in predictions.

Model performance

• DeepMerkel outperformed existing systems, achieving AUROCs of 0.89 (US cohort) and 0.81 (UK cohort) compared to AJCC’s 0.55.
• Its ability to combine staging and non-staging features provided a significant edge over other machine learning models, including logistic regression, random forests, and decision trees.

Explainability with SHAP

• Using SHAP (Shapley Additive Explanations) values, the study identified lymph node involvement, distant metastasis, tumor size, and patient demographics as key drivers of survival outcomes. This level of explainability enhances clinical trust in AI-driven tools.

Real-world implications of DeepMerkel

Precision medicine at scale

DeepMerkel’s predictions are not constrained by population-based averages, as seen in traditional staging. Instead, it provides individualized survival probabilities, enabling clinicians to tailor interventions to each patient’s unique risk profile. For instance, two patients with stage IIB MCC might receive drastically different survival predictions based on DeepMerkel’s inclusion of non-staging factors like income, marital status, and tumor invasion depth.

Web-based clinical tool

To bring its capabilities to bedside practice, the researchers developed a user-friendly web application, the DeepMerkel Survival Calculator. Clinicians can input patient-specific details and receive a Kaplan-Meier survival curve with personalized predictions. This tool democratizes access to advanced AI prognostication, making it feasible for routine clinical use.

Enhanced calibration and performance

DeepMerkel demonstrated superior calibration metrics, with a C-index of 0.93 and a low Brier score of 0.053, indicating reliable and unbiased survival predictions.

Key insights and findings

The study revealed several critical insights into Merkel cell carcinoma (MCC) prognostication. One notable finding was the importance of non-staging features such as socioeconomic status, marital status, and tumor location (e.g., trunk lesions). These factors significantly influenced disease-specific survival (DSS) in ways not captured by conventional staging systems. For instance, married patients and those from higher-income households exhibited better survival outcomes, likely due to enhanced access to healthcare and robust social support networks.

Regarding tumor characteristics, smaller tumor size and superficial invasion were linked to better DSS, while larger tumors and deeper invasion beyond the dermis indicated poorer outcomes, underscoring the critical need for early detection. Additionally, older age and male gender were associated with worse DSS, emphasizing the need for focused surveillance and targeted interventions for these high-risk groups.

Another key factor was nodal involvement; both macroscopic and microscopic nodal disease substantially affected survival outcomes, with patients having distant metastasis showing a 1-year DSS of only 38.2%. Although socioeconomic factors played a pivotal role in survival predictions, they were deliberately excluded from the clinical tool to ensure the focus remained on actionable medical factors.

Limitations and future directions

The study also acknowledged several limitations that warrant attention. Key biomarkers, such as polyomavirus status and immune markers, were missing from the datasets, which may have restricted the model’s predictive capability. Furthermore, the study's retrospective nature, though validated with independent cohorts, necessitates prospective clinical trials to confirm its utility in real-world applications.

The exclusion of socioeconomic factors from the clinical tool, while simplifying decision-making, removes a critical dimension of risk stratification that could aid in personalized care.

Future iterations of DeepMerkel could bridge these gaps by integrating comprehensive biomarker data, incorporating socioeconomic factors where appropriate, and validating the model across diverse patient populations to enhance its precision and generalizability.