CAR-T CRS Prediction

Early CRS Detection Using Wearables + ML (84.62% Accuracy, 7-Hour Lead Time)

Overview

Machine learning system for early prediction of Cytokine Release Syndrome (CRS) in CAR-T therapy patients, developed as part of an investigator-initiated trial at Mount Sinai Hospital. Achieved 84.62% accuracy within a 6-hour prediction window with 7-hour median lead time before standard nursing detection.

My Role: Built the machine learning models for CRS prediction (data analysis).

The Problem

CAR-T therapies (ide-cel, cilta-cel) have revolutionized treatment for relapsed/refractory multiple myeloma, but CRS remains a common and potentially severe complication:

  • CRS: Life-threatening inflammatory response requiring inpatient monitoring
  • Current approach: Reactive monitoring every 4 hours by nursing staff
  • Clinical need: Early warning system to enable outpatient CAR-T delivery

Study Design

Prospective, single-center observational pilot study (2023-2024):

Parameter Value
Patients Enrolled 30
Final Analysis 25 (sufficient monitoring data)
CRS Events 20/25 patients (80%)
Products Idecabtagene vicleucel (ide-cel), Ciltacabtagene autoleucel (cilta-cel)
Median Length of Stay 13 days (IQR 12-14)

Data Sources

Wearable Device

Current Health Gen 2 (FDA-approved) + Feverscout axillary patch (VivaLnk):

  • Pulse rate, oxygen saturation, respiratory rate
  • Skin temperature (upper arm) + axillary temperature
  • Motion data
  • Continuous transmission to cloud

Cytokine Profiling

Olink proximity extension assay (PEA) platform:

  • 92 immune- and oncology-related proteins
  • Normalized Protein Expression (NPX) values on log2 scale
  • Samples collected pre-infusion and at predetermined intervals post-infusion

Machine Learning Approach

Feature Engineering

  • Time-lagged features from rolling windows (6-14 hours)
  • Biomarker fold-changes from baseline
  • Linear, spline, and polynomial interpolation to align cytokine with wearable data

Model Evaluation

5 classifiers evaluated via StratifiedKFold CV (5 splits):

  • Random Forest
  • Gradient Boosting
  • Support Vector Machine
  • Logistic Regression
  • k-Nearest Neighbors

SHAP analysis for feature importance and interpretability.

Results

Wearable-Based CRS Detection

| Metric | Value | |——–|——-| | CRS Episodes Detected | 18/20 | | Median Lead Time | 7:01 hours before nursing detection | | Device Adherence | 71% (IQR 55-84%) during high-risk periods |

Combined Model (Temperature + Cytokines)

Product Best Classifier Accuracy Prediction Window
Ide-cel Random Forest 84.62% 6 hours
Cilta-cel Gradient Boosting 80.62% 6 hours

IFN-γ as Predictive Biomarker

Metric Cilta-cel Ide-cel
Precision 90% 86%
Sensitivity 75% 75%
Accuracy 73% 67%
Mean Lead Time 40 hours (median 22, range 7-120) -

Key Biomarkers by Product

  • Ide-cel: IL5, IFN-γ, NOS3, CD4, TNFRSF9
  • Cilta-cel: IL10, CCL4, MCP-2, MCP-3, IFN-γ
  • IFN-γ: Cross-product predictor present in both models

Mentorship Component

Mentored 6 Carnegie Mellon University graduate students (Sep-Dec 2024) on this project as part of their capstone:

  • Guided experiment design, baselining, and reporting
  • Coordinated with clinical team on project scope and deliverables
  • Students contributed to model development and validation

Clinical Impact

Wearable devices demonstrated feasibility for early CRS detection, offering:

  • Actionable lead times for intervention
  • Support for outpatient CAR-T models reducing hospital stays
  • Integration with cytokine data enhances predictive accuracy

Technical Stack

ML:             Scikit-learn (Random Forest, Gradient Boosting, SVM, LR, k-NN)
Interpretability: SHAP
Cytokines:      Olink PEA (92 proteins)
Wearables:      Current Health Gen 2, Feverscout
Validation:     StratifiedKFold CV (5 splits)
Statistics:     Linear mixed-effects (DREAM framework)
Patients:       N=30 enrolled, N=25 analyzed

Publication

  • SSRN Preprint - “Early Detection of Cytokine Release Syndrome Using Wearable Devices and Cytokine Profiling Following CAR-T Therapy for Myeloma” (Contributing Author)