Using Real-World Data for Machine-Learning Algorithms to Predict the Treatment Response in Advanced Melanoma: A Pilot Study for Personalizing Cancer Care

Abstract

Purpose: The use of real-world data (RWD) in oncology is becoming increasingly important for clinical decision making and tailoring treatment. Despite the significant success of targeted therapy and immunotherapy in advanced melanoma, substantial variability in clinical responses to these treatments emphasizes the need for personalized approaches to therapy.

Materials and methods: In this pilot study, 239 patients with melanoma were included to predict the response to both targeted therapies and immunotherapies. We used machine learning (ML) to incorporate RWD and applied explainable artificial intelligence (XAI) to explain the individual predictions.

Results: We developed, validated, and compared four ML models to evaluate 2-year survival using RWD. Our research showed encouraging outcomes, achieving an AUC of more than 80% and an estimated accuracy of over 74% across the four ML models. The random forest model exhibited the highest performance in predicting 2-year survival with an AUC of 0.85. Local interpretable model-agnostic explanations was used to explain individual predictions and provide trust and insights into the clinical implications of the ML model.

Conclusion: With this proof-of-concept, we integrated RWD into predictive modeling using ML techniques to predict clinical outcomes and explore their potential implications for clinical decision making. The potential of XAI was demonstrated to enhance trust and improve the usability of the model in clinical settings. Further research, including foundation modeling and generative AI, will likely increase the predictive power of prognostic and predictive ML models in advanced melanoma.