Forecasting immune effector cell-associated neurotoxicity syndrome after CAR T-cell therapy

Abstract

Background. Immune effector cell-associated neurotoxicity syndrome (ICANS) can occur days to weeks after chimeric antigen receptor (CAR) T-cell therapy, with manifestations ranging from encephalopathy and aphasia to cerebral edema and death. Because the onset and time course of ICANS are currently unpredictable, prolonged hospitalization for close monitoring is a frequent standard of care.

Methods. Using a cohort of 199 hospitalized patients treated with CAR T-cell therapy at Brigham and Women's Hospital (April 2015–February 2020), we developed a combined hidden Markov model and lasso-penalized logistic regression model to forecast the course of ICANS. Model development used leave-one-patient-out cross-validation.

Results. Among 199 patients, 97 (48.7%) developed ICANS and 59 (29.6%) experienced severe (grade 3–4) ICANS; median onset was day 9. Selected predictors included maximum daily temperature, C-reactive protein, IL-6, and procalcitonin. The model predicted development of ICANS and of severe ICANS with area under the curve of 96.7% and 93.2% when forecasting 5 days ahead, and 93.2% and 80.6% when forecasting the entire future risk trajectory from day 5. Forecasting performance was also evaluated over horizons of 1–7 days using bias, mean absolute deviation, and weighted average percentage error.

Conclusion. The model accurately predicts the risk and time course of ICANS following CAR T-cell infusion.

This project provides the MATLAB code and de-identified data to reproduce the analyses. See the GitHub repository for a one-command reproduction pipeline.

---

Background

Immune effector cell-associated neurotoxicity syndrome (ICANS) is a common, potentially life-threatening complication of chimeric antigen receptor (CAR) T-cell therapy whose onset and time course are currently unpredictable, leading to prolonged precautionary hospitalization. This project releases the model and de-identified data for a forecasting approach that predicts both whether a patient will develop ICANS and how the syndrome will evolve day by day.

---

Software Description

MATLAB source code and de-identified data. The patient identifier is a surrogate SID (no MRNs, names, or dates). Committed data: AllBWH_5days_deID.xlsx (per patient-day clinical features and ICANS meanscore), data_new.mat (grade trajectories + augmented HMM states), LR_prob_allp_Lasso_LOO.mat (leave-one-out onset probabilities), time_LR.mat (time-of-onset risk curve).

---

Technical Implementation

A cohort of 199 patients treated with CAR T-cell therapy at Brigham and Women's Hospital (April 2015-February 2020) was used to develop a combined hidden Markov model (for the latent ICANS-grade trajectory) and lasso-penalized logistic regression (for onset risk from daily clinical features: maximum temperature, C-reactive protein, IL-6, procalcitonin, ferritin, WBC, age). Model development and evaluation used leave-one-patient-out cross-validation. Forecasting performance was assessed over 1-7 day horizons using forecast bias, mean absolute deviation, and weighted average percentage error, and via AUC for predicting ICANS and severe ICANS.

---

Installation and Requirements

MATLAB R2019b or later with the Statistics and Machine Learning Toolbox. No installation beyond cloning the repository. From the repo root run run_all in MATLAB to regenerate the leave-one-out lasso probabilities and the HMM forecast metrics. See REPRODUCE.md and DATA_SOURCE.md in the repository.

---

Usage Notes

Reproduce with the run_all driver (MATLAB). REPRODUCE.md maps each figure/table/metric to its script and committed input; DATA_SOURCE.md documents provenance and de-identification (the de-identified spreadsheet reproduces the raw ICANS-grade matrix exactly). Code and updates: https://github.com/bdsp-core/ICANS-forecasting-after-CAR-T-cell-therapy

---

Release Notes

First public release: code + de-identified data + one-command reproduction pipeline.

---

Ethics

Data were collected under IRB approval at Mass General Brigham; all released data are de-identified.

---

Acknowledgements

We thank the patients and care teams at Brigham and Women's Hospital.

---

Conflicts of Interest

MBW and SSC are co-founders of Beacon Biosignals, which played no role in this work. Other disclosures as stated in the associated publication.

---