Dados do Resumo

Título

Clinicopathological Predictors of Immune Checkpoint Inhibitors Efficacy in Melanoma: A Random Forest-Based predictive model

Introdução

Melanoma has the highest mortality rate among skin cancers due to its metastasis and recurrence. Immune checkpoint inhibitors have shown promise in advanced stages, improving survival compared to traditional therapies. However, response rates vary, emphasizing the need for predictive biomarkers to guide patient selection. Despite the high cost of immunotherapy, no approved biomarkers currently exist for its use in melanoma, prompting further research efforts.

Objetivo

This study aimed to apply a machine-learning model to identify clinicopathological biomarkers with potential predictive value for responses to immunotherapy in melanoma patients.

Métodos

This observational, translational study analyzed data from 82 melanoma patients in unresectable stage III and IV, treated with immune checkpoint inhibitors at a Tertiary Oncology Hospital. Clinical and histopathological data were extracted from standardized medical records. Response to immunotherapy was assessed according to the Response Evaluation Criteria in Solid Tumours for Immunotherapy. BRAF mutation status was determined by real-time PCR (Cobas Z480) or next-generation sequencing (Illumina Univariate logistic regression identified predictors with p-values <0.15 for inclusion in a Random Forest model, implemented with the ranger package. Model tuning involved 10-fold cross-validation with optimization of parameters like mtry, minimum node size, and split criterion (Gini impurity). Predictive performance was evaluated using ROC curve analysis. Statistical analyses were conducted with R software. The study was approved by research ethics committee (CEP-HCB 1772/2019).

Resultados

Univariate analysis indicated 7 potential predictors to be used in the random forest model. After exclusion of missing data, 61 patients were included in the multivariate analysis. The Random Forest model achieved an accuracy of 80.33% and a balanced accuracy of 77.93%, with a Kappa value of 0.5754, indicating moderate agreement between predictions and actual values. Variables with the highest impact in the model included pre-treatment lung metastasis, microscopic satellitosis, and the use of chemotherapy before or after immunotherapy. The confusion matrix revealed that the model had a sensitivity of 66.67% and a specificity of 89.19%, showing a good balance between the ability to detect positive and negative cases. These results suggest that the model performs well in predicting responses, highlighting the relevance of the selected variables.

Conclusões

This study underscores the importance of integrating clinical and pathological factors to predict immunotherapy responses in melanoma patients. Machine learning and artificial intelligence models are crucial for analyzing complex data and enhancing prediction accuracy. These advanced tools enable personalized treatment strategies, improving the effectiveness of immunotherapy and patient outcomes.

Financiador do resumo

This research was funded by the São Paulo Research Foundation (FAPESP, #2019/07111-9), National Council for Scientific and Technological Development (CNPQ, # 444017/2023-2) and Brazilian Melanoma Group (GBM).

Palavras Chave

Histopathology; Machine Learning; immunotherapy