A hybrid machine learning approach for the personalized prognostication of aggressive skin cancers

Abstract

Accurate prognostication guides optimal clinical management in skin cancer. Merkel cell carcinoma (MCC) is the most aggressive form of skin cancer that often presents in advanced stages and is associated with poor survival rates. There are no personalized prognostic tools in use in MCC. We employed explainability analysis to reveal new insights into mortality risk factors for this highly aggressive cancer. We then combined deep learning feature selection with a modified XGBoost framework, to develop a web-based prognostic tool for MCC termed 'DeepMerkel'. DeepMerkel can make accurate personalised, time-dependent survival predictions for MCC from readily available clinical information. It demonstrated generalizability through high predictive performance in an international clinical cohort, out-performing current population-based prognostic staging systems. MCC and DeepMerkel provide the exemplar model of personalised machine learning prognostic tools in aggressive skin cancers.

Fig. 1. Study design and workflow of patient selection.

*DeepMerkel model consists of TabNet/XGBoost framework for 5-year DSS prediction. **DeepMerkel Survival Analysis consists of TabNet/XGBoost framework for time-series DSS prediction. Machine Learning (ML) staging model consists of TabNet/XGBoost framework of staging features only.

Fig. 2. DeepMerkel outperforms other models in predicting DSS.

Graphical representation of AUROC comparison of DeepMerkel (US test cohort, blue, AUROC = 0.89), DeepMerkel (UK test cohort, green, AUROC = 0.81), and ML modeling of staging features (purple, AUROC = 0.68) (p < 0.001). Tiered AJCC staging has not been represented in the graph (AUROC = 0.55). Table represents performance metrics of comparators models and DeepMerkel.

Fig. 3. DeepMerkel survival analysis can accurately perform time-series prediction of survival.

Kaplan-Meier curve demonstrates the predicted mean survival of DeepMerkel Survival Analysis (blue) compared to the actual mean survival observed (black) (c-index = 0.93). This is compared to predicted mean survival in deep learning modeling of AJCC staging features only (purple) (c-index = 0.67). Mean survival of tiered AJCC staging is not presented in this graph (c-index= 0.51). 95% confidence intervals are demonstrated.

Fig. 4. Staging and non-staging features are critical for DeepMerkel performance.

Top 20 Shapley values representing feature importance in DSS have been displayed. This has been validated against clinical domain knowledge. Features in bold are staging features, all others are non-staging features.

Fig. 5. User-friendly DeepMerkel MCC Survival Calculator can predict patient-level DSS.

Representative screenshot of two patients with Stage IIb MCC. These Kaplan–meier plots demonstrate a significant difference in DSS based on non-staging features. Blue line = average predicted survival in stage IIb patients (CI included), Orange line- predicted survival of individual based on patient features, tumor.