A Radiomic Machine Learning Model to Predict Treatment Response to Methotrexate and Survival Outcomes in Primary Central Nervous System Lymphoma (PCNSL)

Introduction: Primary CNS lymphomas (PCNSL) are heterogeneous, aggressive, extra-nodal non-Hodgkin lymphomas limited to the neuraxis. Published response rates to high-dose methotrexate (MTX) based induction regimens for PCNSL range from 35-78%. However, >50% of patients relapse and have a median survival of 2 months without additional treatment. Our ability to prognosticate outcomes is limited to clinical models like the International Extranodal Lymphoma Study Group (IELSG) score and Memorial Sloan-Kettering Cancer Center (MSKCC) classifier. There is an urgent need to develop improved biologic and radiologic predictive models for PCNSL to facilitate therapeutic advances. We hypothesize that a machine learning model using advanced magnetic resonance imaging (MRI) tumor characteristics will improve the accuracy of clinical models to predict response to MTX and survival outcomes. Methods: Data from patients with PCNSL treated at UT Southwestern and Parkland Health and Hospital System hospitals from 2008-2020 (n=95) were collected. An analytical dataset of 61 patients was selected based on the availability of T1 postcontrast (T1c) and T2w FLAIR MR images. A subset of 47 patients was used to evaluate MTX treatment response. Expert neuroradiologists drew regions of interest (ROIs) on the multiparametric MR images including whole tumor (consisting of edema + enhancing tumor + necrosis), enhancing tumor and necrosis (Figure 1). Response to methotrexate-based induction was defined per the International Primary CNS Lymphoma Collaborative Group (IPCG) criteria. For overall- and progression-free survival (OS and PFS) analysis, short (≤1 year) and long-term (>1 year) survivor groups were defined. A support vector machine (SVM) network was used for predicting treatment response to MTX and for predicting the OS groups. A Multinomial Naive Bayes (MNB) network was used for predicting the PFS groups. PyRadiomics package was used to extract 106 texture-based features from the combination of each MR image and tumor ROI. A total of 642 features were extracted from the imaging parameters. Clinical features including age, race, performance status, MSKCC class, IELSG score, histology, delay from 1st MRI to start of treatment, induction and consolidation treatments used were included in the analysis. Feature reduction methodology based on the feature importance derived from the gradient boost model was applied to reduce the number of features. 17 features (imaging = 14, clinical = 3) were used for predicting OS/PFS and 7 features (imaging = 5, clinical = 2) were used for predicting treatment response to MTX. Networks utilizing only clinical features were analyzed for comparison. The sklearn package in python was used for the machine learning analysis. 5-Fold cross validation was performed to generalize the network performance. Results: Baseline wclinical characteristics of the study population is shown in Table 1. Table 2 lists the accuracy, F1 score, sensitivity, specificity, positive predictive value, negative predictive value and area under the curve (AUC) values averaged for the 5-fold cross validation. The SVM network achieved a mean testing accuracy of 81.1 ± 12.3% for predicting the treatment response to MTX-based induction. Sensitivity, specificity and AUC values were 90.5 ± 13.1%, 63.3 ± 22.1% and 0.81 ± 0.14 respectively. The SVM and the MNB network achieved mean testing accuracies of 80.3 ± 11.4% and 83.3 ± 11.8% for predicting the long and short survival groups in OS and PFS respectively. Sensitivity, specificity and AUC values for the SVM and MNB networks were 79.3 ± 6.5%, 80.5 ± 16.5% and 0.86 ± 0.12 and 85.3 ± 12.9%, 81.9 ± 11.8% and 0.86 ± 0.13 respectively. The accuracy values for predicting treatment response to MTX, OS and PFS using only the clinical features were 61.6 ± 9.2%, 59.1 ± 16.4% and 62.1 ± 17.5% respectively. Conclusion: This machine learning model boosted the accuracy (≥20%) over currently validated clinical models alone in predicting response to methotrexate-based therapies and survival outcomes in PCNSL. The current analysis is limited by the small sample size, and we plan to statistically test this model across a larger dataset and report results at the meeting. Our preliminary results suggest that machine learning based radiomic analysis may predict biologic aggressiveness in PCNSL and has the potential to be integrated in clinical predictive tools and design of clinical trials. Disclosures Awan: Blueprint medicines: Consultancy; Celgene: Consultancy; Sunesis: Consultancy; Karyopharm: Consultancy; MEI Pharma: Consultancy; Astrazeneca: Consultancy; Genentech: Consultancy; Dava Oncology: Consultancy; Kite Pharma: Consultancy; Gilead Sciences: Consultancy; Pharmacyclics: Consultancy; Janssen: Consultancy; Abbvie: Consultancy. Desai:Boston Scientific: Consultancy, Other: Trial Finding.