Can machine learning improve randomized clinical trial analysis?

Area of Science:

• Clinical trial methodology
• Machine learning in medicine
• Epilepsy research

Background:

• Randomized clinical trials (RCTs) are costly and statistically inefficient.
• A realistic patient seizure diary simulator has been developed.
• There is a need for more statistically efficient outcome metrics in clinical trials.

Purpose of the Study:

• To identify a more statistically efficient outcome metric using realistic simulation and machine learning.
• To compare deep learning architectures against traditional metrics.
• To evaluate the type 1 error rates of proposed models.

Main Methods:

• Five deep learning architectures with 54 hyperparameter permutations were evaluated.
• Models were compared to the median percent change (MPC) standard.
• Type 1 error rates were assessed for all models.

Main Results:

• The simplest model, logistic regression of a histogram of individual percentage changes (LPC), demonstrated comparable outcomes to other models.
• LPC requires 21-22% fewer patients to achieve 90% power in discriminating drug from placebo.
• All tested models exhibited appropriate low type 1 error rates.

Conclusions:

• The novel LPC metric offers improved statistical efficiency for clinical trials.
• Validation of LPC could lead to faster, more cost-effective clinical trials.
• This approach has potential applications in epilepsy research and beyond.