Proof of Concept / Machine Learning in Pharmacokinetics
A gradient boosting model trained on synthetic pharmacokinetic data to predict vancomycin AUC₂₄ from patient covariates. This proof of concept demonstrates how machine learning can augment traditional population PK approaches — learning non-linear feature interactions, providing prediction transparency through feature importance, and offering a foundation for clinical decision support integration.
Disclaimer: This is a proof-of-concept demonstration using synthetic data generated from published population PK models. The gradient boosting model is trained entirely in-browser on simulated patients. It is NOT validated for clinical use and should NOT inform patient care decisions. Clinical dosing should be performed using validated, institution-approved software and verified by a licensed pharmacist.
Generating synthetic patients & training model...
Decision stump ensemble trained via gradient descent on residuals. 80 iterations with learning rate 0.1. Each weak learner finds the optimal single-feature split minimizing MSE across 20 quantile thresholds per feature.
500 virtual patients generated using the Crass (2018) population PK model with realistic demographic distributions. AUC₂₄ computed analytically (Dose₂₄/CL) with ±10% proportional noise to simulate inter-individual variability.
SHAP-like additive decomposition groups each tree's contribution by split feature, producing a waterfall of per-feature effects. This mirrors how clinical pharmacists reason about dosing — which patient factors most influence the expected exposure.