Meta-analysis informed machine learning: Supporting cytokine storm detection during CAR-T cell Therapy

Area of Science:

• Oncology
• Immunology
• Computational Biology

Background:

• Cytokine release syndrome (CRS), or cytokine storm, is a significant adverse effect of chimeric antigen receptor (CAR) T-cell therapy.
• CRS onset is characterized by specific, yet variable, cytokine and chemokine profiles across patients.
• Accurate and timely diagnosis of CRS is crucial for managing cancer patients undergoing CAR T-cell therapy.

Purpose of the Study:

• To develop a novel, meta-review informed machine learning method for identifying CRS.
• To address the challenge of limited patient data in machine learning models for CRS detection.
• To enhance the accuracy and interpretability of CRS diagnosis using clinical data and literature.

Main Methods:

• Utilized machine learning algorithms to analyze similarities in patient cytokine profiles.
• Augmented limited patient cytokine concentration data with statistical knowledge extracted from domain literature.
• Pioneered a meta-review informed approach combining clinical study evidence with patient data.

Main Results:

• Achieved over 90% accuracy in identifying CRS.
• Demonstrated that the proposed method outperforms purely data-driven approaches.
• Showcased the interpretability of the model's results when applied to real-world CRS clinical data.

Conclusions:

• The developed machine learning method effectively identifies CRS onset using cytokine peak concentrations and literature data.
• Augmenting data with domain knowledge improves model performance and addresses data scarcity.
• This approach offers a valuable tool for clinicians to support swift and accurate CRS diagnosis in cancer patients.