A machine learning-based triage system for systemic EBV-positive T/NK cell lymphoproliferative diseases of childhood.

Abstract

Systemic Epstein-Barr virus-positive (EBV-positive) T/NK cell lymphoproliferative diseases of childhood (sEBV+T/NK-LPD) are a spectrum of rare diseases that have highly variable biological behavior, from indolent conditions to highly aggressive malignancies. Clinicians currently face substantial challenges in promptly assessing disease severity and predicting patient outcomes, leading to limitations in treatment planning. To address this challenge, we constructed a comprehensive triage system to aid in rapid clinical interventions. The study included 156 patients with newly diagnosed sEBV+T/NK-LPD from 42 institutions. An independent prospective cohort of 35 newly enrolled patients was further included to evaluate the model's performance. An additional 45 patients from the literature and 18 patients who underwent hematopoietic stem cell transplantation were included to test the score's generalizability. An integrative machine learning strategy was applied to identify robust and optimal factors and to integrate multiple algorithms to enhance the system's performance and stability. This system, termed COLLAPSED, identifies critical factors and provides a stable, high-performing ensemble. This model was validated externally and simplified into a risk score to improve interpretability and accessibility. The COLLAPSED system substantially enhances clinicians' ability to rapidly and precisely identify high-risk patients, thus enabling timely clinical decision-making and expedited initiation of potentially lifesaving treatments.