Carlos Serra Traynor, Marija Kekic, David Boulton, Diansong Zhou
1 Clinical Pharmacology & Quantitative Pharmacology, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca (Barcelona, Spain), 2 Predictive AI & Data, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca (Barcelona, Spain), 3 Clinical Pharmacology & Quantitative Pharmacology, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca (Gaithersburg, United States), 4 Clinical Pharmacology & Quantitative Pharmacology, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca (Waltham, United States)
Introduction: Neutropenia remains a common and clinically important toxicity of cytotoxic chemotherapy, and early risk prediction is central to safer treatment delivery and more targeted use of supportive care such as G-CSF. Docetaxel provides a well characterised and event rich setting in which to develop and evaluate predictive approaches. In this study, we combined a Friberg style kinetic pharmacodynamic (KPD) model of absolute neutrophil count (ANC) with machine learning methods, using stochastic gates (STG) for covariate screening and XGBoost for event prediction, to assess whether model informed ANC features improve early prediction of neutropenia (and agranulocytosis) relative to conventional clinical heuristics.
Methods: From the all-comers Flatiron Health database [1], an initial cohort of 3,847 patients treated with docetaxel was identified. Of these, 1,771 had sufficient baseline and follow up data on demographics (race, ECOG performance status, sex), renal function, baseline laboratory values, and post-dose ANC to support development of a Friberg-style model with KPD as docetaxel or G-CSF input. Docetaxel dosing was body surface area based. Most patients received 60 to 100 mg/m^2 every three weeks (81.4%), with the remainder receiving lower dose weekly or fortnightly schedules (typically 20 to 40 mg/m^2), consistent with tolerability driven dose modification. The analysed endpoints were sustained neutropenia, defined as two consecutive ANC values < 1.0 × 10^9/L (242 events, 13.6%), and a composite endpoint of sustained neutropenia or clinically recorded neutropenia or agranulocytosis (312 events, 17.6%). The dataset was randomly split 50:50 into training and validation sets. Hamiltonian Monte Carlo (NUTS) in Torsten Stan [2] was used, with literature informed population priors [3], and 4 chains with 2,000 post warm-up draws to develop KPD models (with or without G-CSF) which describe individual ANC trajectories in the training set over 56 days post dose. Both models were then prior-posterior updated using baseline characteristics and the first 7-days post-dose to predict ANC profiles in the validation set. Individual model informed ANC parameters were benchmarked against conventional clinical heuristics, including observed nadir and time to nadir, for predicting sustained neutropenia and the composite endpoint. Candidate covariates were screened using a stochastic gates (STG) neural network [4], and events were predicted using XGBoost (boosting and DART), with performance assessed in the validation set using AUC and false positive rate (FPR) controlled classification.
Results: Sensitivity and structural identifiability assessments showed that a 3-transit compartment KPD model with a docetaxel driven proliferation inhibition slope, with or without G-CSF stimulation, adequately describe docetaxel/G-CSF input. Bayesian NUTS diagnostics indicated good convergence and stable posterior estimates for all identifiable parameters (R-hat = 1.00 throughout, with strong effective sample sizes). Model comparison favoured the KPD model with G-CSF stimulation, with substantially better out of sample fit than the model without G-CSF (ELPD LOO −4808 vs −5629; WAIC 9351 vs 11160). Visual predictive checks indicated good agreement between observed and model predicted ANC trajectories at the individual level, both with and without G-CSF. Diagnostics and visual predictive checks following the validation set update indicated that the model was robust and well captured the ANC trajectories for the validation set. The STG screening identified model predicted ANC nadir and early G-CSF use as key predictors of both severe neutropenia and the composite endpoint. At an FPR of ≤10%, the KPD model informed features delivered a modest but consistent improvement over the clinical early observation heuristic, with better discrimination for both endpoints (AUC 0.773 vs 0.647 for sustained neutropenia; AUC 0.799 vs 0.651 for the composite endpoint) and improved operating characteristics, most notably higher recall (0.448 vs 0.176; 0.509 vs 0.242) at comparable false positive count (76 and 73 false positives, respectively). Finally, feature importance was consistent across both endpoints, with predicted nadir, percentage decrease, and time below 1.0 × 10^9/L emerging as the key predictors.
Conclusions: The G-CSF augmented KPD model was parameter identifiable and provided clinically useful improvements in early prediction of docetaxel induced neutropenia, with model informed nadir allowing for higher sensitivity at the cost of a manageable increase in alert count, a trade off that is often appropriate in neutropenia screening.
References:
References:
[1] Flatiron Health. (2025). Database characterization guide. Retrieved Dec 15, 2025, from https://flatiron.com/database-characterization
[2] Margossian, Charles C., Yi Zhang, and William R. Gillespie. "Flexible and efficient Bayesian pharmacometrics modeling using Stan and Torsten, part I." CPT: Pharmacometrics & Systems Pharmacology 11.9 (2022): 1151-1169.
[3] Friberg, Lena E., et al. "Model of chemotherapy-induced myelosuppression with parameter consistency across drugs." Journal of clinical oncology 20.24 (2002): 4713-4721.
[4] Kekic, Marija, et al. Stochastic Gates for Covariate Selection in Population Pharmacokinetics Modeling. CPT: PSP 15.2 (2026): e70147.
Reference: PAGE 34 (2026) Abstr 12119 [www.page-meeting.org/?abstract=12119]
Poster: Clinical Applications