2021 FDA Science Forum
Decision Tree Based Approaches with Application to a Rabbit Left Ventricular Wedge Preparation Study
- Authors:
- Center:
-
Contributing OfficeCenter for Drug Evaluation and Research
Abstract
Background
Drug-induced Torsde de Points (TdP) is a rare but potentially fatal side effect. The risk of TdP is recommended to be assessed at early drug development period. The Comprehensive in Vitro Proarrhythmia Assay (CiPA) steering team published a selection of 28 drugs categorized as low, intermediate and high TdP risk, and set up general principles to quantify models and metrics to be used to predict proarrhythmia risk.
Purpose
A type of ex vivo study, rabbit ventricular wedge assay (RVWA), was proposed to be used to guide the early development decision and assist the evaluation of potential risk. Multiple ECG intervals and an ordinal variable representing the early afterdepolarization related incidence are recorded from each sample in a blinded validation study. This research project uses a decision tree based model to predict proarrhythmia based from RVWA results.
Methodology
A Bayesian additive regression tree (BART) model was used to analyze data from the RVWA validation study. Each drug is assigned to a risk category based on the predicted probabilities. The model uses a regularization prior to summarize information from multiple binary prediction trees. The posterior probability vectors yielded from MCMC are further summarized and visualized in a distance metric for easier understanding and clearer explanation.
Results
The results based on the training and validation framework suggested by CiPA program have 75% (4 out of 16) correct predictions of risk categories. The misclassified drugs are likely to have posterior densities close to boundary of two categories or showing greater uncertainty from the data. The results based on leave one drug out validation has 82% (23 out of 28)correct predictions.
Conclusion
The BART model demonstrated robust prediction results. More importantly, the posterior densities provide descriptive information regarding the uncertainty of the categorization. The corresponding distance visualization provides a better understanding of the relationship between a testing compound and reference drugs in different categories.
