PT-027 - UTILIZING ADAPTIVE DOSING SIMULATIONS AND MACHINE LEARNING TO IMPROVE PREDICTION AND MANAGEMENT OF NEUTROPENIA RATES.

X. Huang¹, C. Grant², K. Kowalski³, B. Jermain³; ¹UC San Diego Skaggs School of Pharmacy and Pharmaceutical Sciences, San Diego, CA, USA, ²Mayo Clinic Graduate School of Biomedical Sciences, Rochester, MN, USA, ³Pfizer, La Jolla, CA, USA.

Background: Palbociclib (palbo) is used for adult HR+/HER2- advanced breast cancer, with a daily dose for 21 days followed by 7 days off per 28-day cycle. Neutropenia, marked by low absolute neutrophil count (ANC), affects ~80% of patients in clinical trials. This study uses machine learning (ML) to identify predictors of Grade 4 Neutropenia (GR4) with data from adaptive dosing simulations, then leverages top predictors to develop a new dosing plan to reduce GR4 events.

Methods: The ML model, random forest (RF), uses data from adaptive dosing simulations with 1000 virtual subjects over 4 cycles. Reference Jermain ACOP 2023 poster, palbo US Package Insert (USPI), and Wan PKPD 2017 publication, the outcome variable is presence of GR4 (ANC < 500/mm3), and predictors include individual clearance, baseline ANC, changes in ANC, palbo concentration and area under the curve over the first 4 weeks. The data is split 80/20 into training and testing sets, respectively. In training, hyperparameters were tuned, and the model with the highest area under the receiver operating characteristic curve was selected for testing on the test data. Model performance is assessed with accuracy, sensitivity, specificity, and precision. Validated models identify key GR4 predictors and cutoff values, which are implemented as new dosing rules in the PKPD model.

Results: The USPI simulation data showed that 18.1% of patients developed GR4 over 16 weeks, with 79.8% of these cases occurring by day 21. The RF model using days 7 and 14 data achieved 93% accuracy, 89.2% sensitivity, 93.9% specificity, and 76.7% precision, identifying the ANC %change from baseline to day 14 as a key predictor. Implementing a new dosing rule—reduce and hold the dose until ANC ≥ 1000/mm³ if ANC drops ≥ 77% from baseline at day 14—resulted in a 6.1%, 9.2%, and 11.6% reduction (Chi-squared test p< 0.001) in GR4 events over 16 weeks, at day 21 or later, and at day 21, respectively (Figure 1). The new rule led to a 5.5% reduction in average palbo concentration (standard deviation: 6.8%) in GR4 patients from USPI dosing.

Conclusion: ML was able to identify early predictors of GR4 for a virtual population receiving palbo with USPI dosing. Adjusting dosing rules based on RF findings reduced GR4 events while maintaining similar palbo exposure through 4 cycles. Ongoing work involves daily sampling predictions and exploring other ML algorithms.