Leveraging a Novel Machine Learning Method to Predict Outcomes in Latin American (LATAM) Patients with Newly Diagnosed T and NK Cell Lymphomas (TNKL)

Introduction:

Our PETAL consortium group recently characterized the largest global cohort of relapsed/refractory TNKL patients demonstrating statistically significant survival differences between patients in high-income and developing countries like LATAM (Koh MJ, et al. 2023). Estimating outcomes for newly diagnosed (ND) TNKL patients in developing countries can be challenging due to the lack of comprehensive lymphoma registries, limited clinical and molecular data, and substantial missing information. This missing data often does not occur randomly and may contain crucial insights regarding patient characteristics and prognoses. However, traditional analyses often exclude these patients, which ultimately affects the validity of the conclusions. Therefore, developing cutting edge machine learning (ML) tools to estimate survival with high precision is warranted.

Methods:

We conducted a retrospective cohort study of 487 patients with ND TNKL from 9 LATAM countries (Argentina, Chile, Colombia, Guatemala, Mexico, Paraguay, Uruguay, and Venezuela). Eligible patients were aged ≥18 years and diagnosed between 2000-2022. Overall survival (OS) was calculated from diagnosis to death or loss to follow-up using the Cox proportional-hazards method, based on two covariates: histologic subtype and IPI or PIT score at diagnosis. The analysis was performed using both complete and imputed data. To adjust for unobserved confounders arising from missing data in retrospective datasets, OS was estimated using these covariates with a novel causal inference method called survival synthetic intervention (SSI).

SSI utilizes available data to estimate survival for a specific group of patients by first converting the data to panel data with the Kaplan-Meier estimator. In the panel data setting, SSI leverages principal component regressions to estimate outcomes using observations from (a) all patient groups for a reference intervention and (b) observations under an intervention of interest for a “representative” subset of patient groups. Specifically, SSI learns from the high-dimensional survival information of the group of patients with missing data to better characterize the OS of the entire dataset. We evaluated the empirical performance of SSI and compared it to the Cox method. The data were randomly split into a training set (80%) to build the model and a testing set (20%) to test performance.

Results:

We included 475 patients with a median age of 52.8 years (range, 15-95), of whom 283 (60%) were male. ATLL was the predominant histology, observed in 183 (39%) patients, followed by PTCL-NOS in 157 (33%) and ENKTCL in 75 (16%). Using histologic TNKL subtype and IPI or PIT scores at diagnosis as covariates, we found that SSI outperformed multivariate Cox models in both the training and testing sets. For Model I (incorporating histologic subtype and IPI score), the testing concordance index (CI) for predicting OS was 0.595 for the Cox model trained with patients who had complete data only, 0.612 for the Cox model trained with imputed data, and 0.651 for SSI. For Model II (incorporating histologic subtype and PIT score), the testing CI for predicting OS was 0.556 for the Cox trained with patients who had complete data only, 0.612 for the Cox model trained with imputed data, and 0.635 for SSI.

Conclusions:

Low survival rates for TNKL in LATAM highlight the urgent need for improved prognostication tools and access to novel therapies. We applied SSI, a novel machine learning method, to correct biases in a retrospective real-world dataset from LATAM, achieving accurate outcome estimates. Excluding patients from analyses due to missing data leads to biased outcomes. SSI learns from patients with missing information, enhancing overall outcome predictive accuracy.

Sorial:MJH Life Sciences: Honoraria; The Dedham Group: Consultancy; SecuraBio: Research Funding; Daiichi Sankyo: Research Funding. Jain:Myeloid Therapeutics: Consultancy, Membership on an entity's Board of Directors or advisory committees; Acrotech: Membership on an entity's Board of Directors or advisory committees, Research Funding; Mersana Therapeutics: Consultancy, Membership on an entity's Board of Directors or advisory committees; SecuraBio: Membership on an entity's Board of Directors or advisory committees, Research Funding; Kyowakirin: Research Funding; SIRPant Immunotherapeutics: Consultancy, Membership on an entity's Board of Directors or advisory committees, Research Funding; Abcuro Inc: Consultancy, Membership on an entity's Board of Directors or advisory committees, Research Funding; Daiichi Sankyo: Membership on an entity's Board of Directors or advisory committees, Research Funding; Crispr therapeutics: Membership on an entity's Board of Directors or advisory committees.