AI-002 - APPLICATION OF MACHINE LEARNING COMBINED WITH POPPK AND QSP MODELS FOR ADVERSE EVENTS RISK PREDICTION OF A CD19-TARGETING BISPECIFIC THERAPY

B. Sun^1,2, X. Zhu³, K. Smart³, M. Lai³, C. Pichardo³; ¹University of Minnesota College of Pharmacy, ²Summer Intern, AstraZeneca, Waltham, MA, USA, ³AstraZeneca.

Background: Commonly observed adverse events associated with T cell engagers (TCEs) are cytokine release syndrome (CR S) and neurotoxicity (NT). In this study, we aim to develop a framework to combine PopPK, QSP with machine learning models to predict CRS/NT risks using patient-level risk factors and dosing information.

Methods: Data collected from an ongoing ph1 study for a CD3×CD19 TCE was used using the dataflow described in Figure 1. Patient-level factors and PK and PD variables (from a mechanistic QSP model) at each dosing event were considered as potential predictors. The most predictive variables (from univariate and stepwise regression) were used as features for machine learning (ML) algorithms that were tested, evaluated, and calibrated. The CRS prediction model was developed based on data linked to first dosing events and validated with later dosing events. The best performing and calibrated models were applied to predict probabilities for CRS and NT.

Results: The study cohort includes 60 patients with 187 dosing events, of which 50% (30/60) developed CRS events and 28.3% (17/60) NT events. The final selected features are ECOG status, Max IL-6 and AUC after first dosing event for CRS and age, creatinine at screening and AUC changes between dosing events for NT prediction. Logistic regression (LR) models outperformed other ML algorithms with a mean precision, recall and AUROC score of 0.71, 0.67, 0.87 for CRS and 0.34, 0.75, 0.78 for NT prediction. The LR model for CRS was also validated with dataset from later dosing events as an AUROC score of 0.75. Both LR models achieved improved calibration with Brier scores decreasing from 0.33 to 0.28 for CRS and 0.21 to 0.08 for NT.

Conclusion: Combining multifactorial data such as baseline patient factors and outputs from PopPK and QSP models can achieve reasonable prediction in CRS and NT risk using machine learning techniques. Ongoing data from this clinical trial can provide prospective validation of the model. Future study is needed to explore the model generalizability to TCEs.