Time-Series Clustering Captures Patterns of Early Immune Effector Cell-Associated Hematotoxicity That Are Predictable...

Area of Science:

• Hematology
• Immunotherapy
• Data Science

Background:

• Immune effector cell-associated hematotoxicity (ICAHT) is a significant cause of mortality post-chimeric antigen receptor (CAR) T-cell therapy.
• Current grading systems for early ICAHT (eICAHT) may not fully capture the complexity of hematotoxicity patterns.

Purpose of the Study:

• To investigate if unsupervised time-series clustering can identify distinct patterns of early hematotoxicity after CAR T-cell therapy.
• To compare the effectiveness of clustering patterns versus the eICAHT grading system in identifying patient outcomes.

Main Methods:

• Applied k-means time-series clustering to longitudinal absolute neutrophil count (ANC) data from 691 patients.
• Trained a random forest (RF) model using early ANC values (days +3, +4, +5, +26, +27) to predict cluster assignments.
• Validated the RF model's predictive accuracy and compared cluster separation with eICAHT criteria.

Main Results:

• Identified four distinct ANC recovery clusters: very good, good, poor, and very poor.
• RF-predicted clusters showed better separation and compactness than eICAHT criteria.
• The RF model identified patients in the 'good' recovery cluster with intermediate overall survival, a finding missed by grade 2 eICAHT.

Conclusions:

• Unsupervised time-series clustering effectively identifies clinically relevant hematotoxicity patterns post-CAR T-cell therapy.
• A trained RF model accurately predicts these patterns using only five ANC measurements.
• An online web application is available for generating predictions.