Anti-PD-1 Based Immunotherapy in Melanoma: Application of Machine Learning to Predict Survival and Elucidate Complex Relationships

Abstract

Project 1 Background: We assessed if machine learners outperform Cox-PH models in a high dimensional setting in the context of anti-PD-1 treatment for melanoma brain metastases as accurate prediction of overall survival/OS is critical to prioritize limited treatment options. Methods: In this retrospective cohort study, metastatic melanoma patients had radiological evidence of brain metastases (treated/untreated) at the time of anti-PD-1 (monotherapy/combination) immunotherapy (June-2014–September-2016). The primary outcome was OS. Five tasks were evaluated (Task-1: 22-features; Task-2: diagnosis specific-graded prognostic assessment/ds-GPA; Task-3: Recursive Partitioning Analysis/RPA; Task-4: ds-GPA plus 20-features; Task-5: RPA plus 21-features) using machine learners (Random Survival Forests/RSF, Boosting, Cox-Boost, Survival-Tree) and Cox-PH models based on test-set performance (5-fold cross-validated C-index) in a benchmark experiment followed by feature filtering and sequential forward selection (best performing learner). Results: The median OS was 59 days (IQR:13-97) for ds-GPA 0-1, 473 days (IQR:150-Not Reached/NR) for ds-GPA 2 and was NR for ds-GPA 3-4 (N=56 patients/33 deaths). RSF had the lowest mean rank and performed significantly better than Cox-PH (mean-rank: 1.3vs.4.3; Critical-Difference=2.73; p-value=0.023). Average C-index for Task-2/ds-GPA (0.867;95%CI:0.777-0.947) and Task-4 (0.878; 95%CI:0.760-0.961) feature-subset was comparable. Task-3/RPA performed poorly (C-index:0.664; 95%CI:0.560-0.779). Task-4 subset included ds-GPA, ≥4 extra-cranial metastases, prior surgical resection and stereotactic radiotherapy/SRT, neurological symptoms and spine metastases. Prior surgical resection and SRT predicted improved survival for ds-GPA>1. Conclusions and Impact: RSF had the best relative performance compared to Cox-PH models in a high dimensional setting. OS was prolonged for ds-GPA>1 with anti-PD-1 immunotherapy compared to historical estimates (surgery and/or radiation). Baseline dsGPA was the strongest predictor. Project 2 Background: Obesity paradox (improved survival in overweight/obese patients compared to normal weight) is well documented in cancer but not in metastatic melanoma. We characterized the relationship of Body Mass Index (BMI) with survival outcomes and explored complex interactions in the context of PD-1 blockade. Methods: Unresectable or metastatic melanoma patients who received at least one dose of pembrolizumab, nivolumab, or nivolumab plus ipilimumab (combination) from June- 2014-September-2016 were included in this retrospective cohort study. Overall Survival and Progression Free Survival were the main outcomes. Analysis was performed using Random Survival Forests (RSF) and multi-variable Cox Proportional Hazards models. Results: A “U” shaped relationship was noted for pre-treatment BMI where overweight/Class-I (25-<35kg/m2) obese patients had a lower risk of mortality (HR:0.26; 95%CI:0.1-0.71; p-value=0.008) and progressive disease (HR:0.43; 95%CI:0.19-0.95; pvalue:0.038) compared to normal-weight (18.5-<25kg/m2) (Total N=139). Exploration of interactions (RSF) showed that the association was predominantly driven by males; it was attenuated in patients with serum creatinine<0.9mg/dL, who were predominantly females. Findings were similar for anti-PD-1 monotherapy(n=79)/combination(n=60). Conclusions: The findings support the presence of “obesity paradox” where overweight/Class-I obesity was associated with substantially lower risk of mortality and progressive disease compared to normal weight; this association was driven predominantly by males who largely had higher serum creatinine levels (surrogate for skeletal muscle mass in the setting of metastatic disease). These findings suggest that instead of BMI alone, sarcopenia (low skeletal muscle mass) or direct measures of body mass composition may be more suitable predictors of survival in melanoma patients treated with PD-1 blockade.