Predicting Individual Treatment Effect from Randomized Clinical Trial Data: A Comprehensive Evaluation Framework based on nested Cross-Validation
Although clinical doctors have patient’s information such as demographics, medical history, and lab test results, the decision they made is based on average effects across many subjects. There is an increasing need on estimation of individual treatment effect (ITE).
In this study, we develop and evaluate preliminary supervised machine learning pipelines to predict two related outcomes for an individual patient: the probability of diabetes onset (a binary classification task) and the net-benefit of possible treatments. We utilize logistic regression and random forests models to make predictions on individual outcome and benefit (ITE). We also applied state-of-the-art models causal forest and causal boosting, which optimize the treatment effects when building the models, to estimate the treatment effects. The results are evaluated with area under roc, c statistics and quantile calibration plots.
We developed prediction and evaluation pipeline for ITE estimation in the study and hopefully will benefit future clinical trial research.
Please join in Sococo VH 475, or join via Zoom:
Zoom Link: https://tufts.zoom.us/j/93918876139
Password: please see colloquia email for password
Dial-in option not available.