Immunotherapy toxicity prediction in patients with melanoma using machine learning algorithms.

Abstract

e21567 Background: Immune checkpoint inhibitor (ICI) related toxicity is common in melanoma patients, but identifying who will experience toxicity can be challenging. Precision medicine tools that accurately predict ICI toxicity could improve patient outcomes. This study aimed to use Machine Learning (ML) to predict ICI toxicity in patients with melanoma. Methods: 384 datasets were available for patients started on immunotherapy between 2014-2023 in a single, large regional cancer center in Kent, U.K. Raw data was preprocessed using a standard scaling function and one hot encoding, followed by SMOTE to balance the imbalanced classes. Six ML algorithms were developed using 80% of the training data, tested on 20% of validation data and used the Grid Search Cross-Validation technique for hyperparameter optimization. Shapley Additive Explanations (SHAP) explainable artificial intelligence was used to interpret the toxicity prediction model to improve the interpretation and transparency of the decision boundary. Results: Of the 384 patients, 112 patients (29%) had completely resected disease and then had adjuvant treatment while 272 (71%) were unresectable or had metastatic disease and had palliative treatment intent. In the metastatic setting, the commonest immunotherapy regimen was Pembrolizumab (155, 57%), followed by Ipilimumab/Nivolumab (85, 31%). The commonest regimen in the adjuvant setting was Pembrolizumab (86, 77%). 95 patients (25%) experienced ICI-related toxicity. The number of patients with Grades 1-4 toxicity was 2 ,4, 18 and 4 respectively, the remaining had unspecified toxicity grades. The commonest class of toxicity (CTCAE v5) was gastrointestinal (36, 38%), followed by endocrine (20, 21%) and skin (19, 20%). Linear Regression and Gaussian Naïve Bayes predicted toxicity with the most accuracy, 92% (Table 1). The model showed that increasing age, female gender and not receiving Nivolumab predicted toxicity. Regardless of the regimen or treatment intent, our ML model could also predict immunotherapy response with 90% accuracy. Conclusions: These ML algorithms used clinically available data to predict ICI toxicity accurately. Key predictors of toxicity included increasing age, female gender and not receiving Nivolumab in the setting of both adjuvant and palliative intent. Future work will aim to externally validate these models in other centers. We will also explore if models could be trained for use in patients treated with nivolumab in combination with new anti-LAG-3(relatlimab). [Table: see text]