Evaluating novel imaging-based mathematical modeling prediction of immunotherapy response by individual melanoma brain metastasis and patient prognosis.

Abstract

e14013 Background: Ipilimumab/Nivolumab (Ipi/Nivo) significantly transformed outcomes for patients with melanoma brain metastases (MBM), however inter- and intra-patient responses can vary widely. The purpose of this study is to develop and calibrate an imaging-based mathematical model to predict lesion-specific immune response to Ipi/Nivo in patients with MBM to guide early local treatments such as radiotherapy and evaluate the prognostic value of identifying treatment resistant lesions. Methods: Baseline and follow-up 3D T1-W contrast MRIs of patients who solely received Ipi/Nivo after diagnosis with MBM were used to segment individual brain metastases and track serial volumetric data for each lesion to develop a mechanistic model (eq 1). After applying appropriate assumptions to our model, tumor burden ( ρ), intrinsic lesion growth rate ( α 0), Ipi/Nivo-induced kill rate ( μ) and immune response strength ( Λ) were quantified for each lesion. Growth rate shortly after the start of treatment ( α 1 ) was calculated using the short-term model solution (eq 2). dρ/dt= (α0- μ+μ.Λ).ρ - μ.Λ.ρ2 (eq 1) ρ(t)≈eα1.t (eq 2) Change in BM volume between baseline and last follow-up was used to classify each lesion as either ‘responder’ or ‘non-responder’; these were compared with the Wilcoxon-rank-sum test. The highest α1 value for any one BM within a patient tumor was evaluated for its potential to sort patient overall survival (OS). Because the median survival was not reached for this cohort, we determined the α1 threshold of 0.005 to sort patients using an optimization routine that generates the maximum AUC to separate the cohort into 2 groups on an ROC curve. Kaplan-Meier (KM) curves were compared between these 2 groups. Results: In 23 patients, we evaluated 61 lesions. At the lesion level, model parameters representing μ and Λ were significantly different between responder and non-responder lesions (p-values <0.0001 and 0.0004, respectively). The highest α1value for any one BM within a patient tumor was used to sort patient overall survival (OS) into prognostically ‘favorable’ or ‘unfavorable’ groups based on patient survival to ±1,200 days. In this limited cohort, OS KM curves by most aggressive lesion for favorable (α1 < 0.005) vs. unfavorable (α1 > 0.005) were not significant by log-rank test (p = 0.3263). Conclusions: In this limited cohort, we demonstrate feasibility to leverage serial conventional imaging and tumor volumetric data to calibrate a mechanistic model to predict per-lesion treatment response of MBM to Ipi/Nivo. Although the p-value between groups was found to be insignificant when predicting survival using α1, this is likely because of the small number of patients included in this study (n=23). Furthermore, when taken together with the notable AUC (=0.7917) observed in the ROC curve, this promising result supports further investigation in a larger cohort.