Development of a machine learning model to predict overall survival in patients with peripheral T-cell lymphoma in a minority enriched population Free

Introduction: Peripheral T-cell lymphomas (PTCL) are a group of heterogeneous lymphomas that arise from mature T-cells and comprise about 10% of all non-Hodgkin lymphomas. While diverse clinically and pathologically, these lymphomas often present at advanced stages and have a poor prognosis. Accurate prognostication for these patients is critical to inform clinical decision making and treatment strategies. Traditional tools to estimate overall survival (OS) can fall short when it comes to handling complex clinical data. Established indices such as the International Prognostic Index (IPI), Prognostic Index of PTCL-NOS (PIT), modified PIT, and others typically include factors such as age, disease stage, performance status, and lactate dehydrogenase (LDH), but may not capture the full spectrum of clinical and laboratory data in individual patient cases. In this study we aimed to develop and validate a machine learning (ML) model to predict OS using comprehensive demographic, clinical, and laboratory data from diagnosis.

Methods: We identified 97 patients with PTCL diagnosed at Montefiore Medical Center between January 1^st, 2010 and December 31^st, 2022. Data was collected via manual chart review. Inclusion criteria included a diagnosis of PTCL, age >18, and complete data at diagnosis. Patients with incomplete data were excluded. Data processing and modeling was performed in Python using libraries such as scikit-learn, scikit-survival, and pandas. A random forest survival model was used for model development. Model performance was evaluated using concordance index (C-index) to assess discriminatory ability and Brier scores were calculated at different time points (30, 365, 730, and 1095 days) to assess calibration. Permutation feature importance (PFI) was used to evaluate individual variable contribution, and partial dependence plots were used to interpret model behavior.

Results: Among the 97 patients with PTCL, 42% were diagnosed with Adult T-cell Leukemia/Lymphoma (ATLL), 28% with Peripheral T-cell Lymphoma not otherwise specified (PTCL-NOS), 8.2% with Extranodal NK/T-cell Lymphoma, Nasal Type (ENKTCL), 7.2% with Angioimmunoblastic T-cell Lymphoma (AITL), and 13.4% with other PTCL subtypes. The cohort was 54.6% male and 45.4% female. 49.6% were Black, 36.1% were categorized as Other, 9.2% were White, 4.1% were Unknown, and 1% was Asian. 38% were of Hispanic ethnicity. The average age at diagnosis was 60 years. During the study period, 57.7% of patients died and 25.8% were lost to follow-up. The average time from diagnosis to death or censorship was 760 days.

Our ML model demonstrated predictive performance in the full dataset with a C-index of 0.86, suggesting strong overall discriminative performance, though on the validation set the C-index decreased to 0.68. Brier scores at 30, 365, 730, and 1095 days were 0.07, 0.21, 0.21, and 0.19, respectively, indicating good calibration over time. The most predictive features for the full data set included baseline LDH, ECOG at diagnosis, Charlson Comorbidity Index score, ejection fraction (EF), absolute neutrophil count (ANC), hemoglobin, and if the patient received a transplant after first- or second-line therapy (PFI of 0.03, 0.02, 0.02, 0.02, 0.02, 0.01, and 0.01 respectively). Creatinine (PFI= 0.009), number of extranodal sites (VI=0.009), Medicaid coverage (PFI= 0.007), and GFR (PFI= 0.007) had the next highest feature importance scores. Baseline LFTs and bilirubin did not significantly contribute, and neither did age >60. Partial dependence plots were suggestive of shorter predicted survival among patients of Hispanic ethnicity and those with Medicaid.

Conclusions: Our machine learning model demonstrated predictive capacity for OS in our full cohort. The discrepancy in the C-index in the full cohort and validation sets is likely reflective of the impact of limited sample size and indicates potential overfitting in our model. This emphasizes the need for further model refinement and validation on independent data. Overall, these findings do support the feasibility of machine learning in generating survival prediction models in PTCL but again highlight the importance of external validation and a larger sample size to improve model robustness and possibility for clinical use. Future work will focus on refining the model, incorporating further molecular and genetic data, and validation in multi-institutional cohorts.