Comparison of Different Machine Learning Methodologies for Predicting the Non-Specific Treatment Response in Placebo...

Area of Science:

• Psychopharmacology
• Computational Neuroscience
• Clinical Trial Methodology

Background:

• The placebo effect significantly confounds the assessment of treatment efficacy in clinical trials, especially for antidepressant medications.
• Current randomized, placebo-controlled trial designs cannot estimate individual non-specific treatment effects.
• High placebo response rates hinder the development of novel antidepressant drugs.

Purpose of the Study:

• To compare machine learning (ML) methodologies for estimating individual probabilities of non-specific treatment effects.
• To identify ML models that can predict individual placebo response.
• To support novel methods for controlling the impact of high placebo response in clinical trials.

Main Methods:

• Comparative analysis of six ML methodologies: gradient boosting machine, lasso regression, logistic regression, support vector machines, k-nearest neighbors, and random forests.
• Utilized multilayer perceptrons artificial neural network (ANN) for predicting individual non-specific treatment response probability.
• Employed fivefold cross-validation to evaluate model performance and overfitting risks.

Main Results:

• Artificial neural network (ANN) demonstrated the highest overall accuracy in predicting individual non-specific treatment response compared to other ML methods.
• ANN model performance was validated using fivefold cross-validation.
• Excluding subjects with non-specific effects significantly increased signal detection and effect size.

Conclusions:

• Machine learning, especially ANN, can accurately estimate individual non-specific treatment effects, offering a potential solution to the placebo confounder.
• Accurate estimation of individual placebo response can lead to more sensitive clinical trials and improved antidepressant drug development.
• This approach may enhance the ability to detect true treatment effects by better controlling for placebo response.