A clinical trial termination prediction model based on denoising autoencoder and deep survival regression

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Machine Learning in Healthcare

Background:

• Effective clinical trials are vital for medical progress, but early termination leads to resource waste.
• Survival models predict trial outcomes, yet sparse data challenges existing models like DeepSurv, limiting accuracy and generalization.
• Clinical trials often exclude pregnant women, necessitating specialized prediction models for this population.

Purpose of the Study:

• To develop an improved survival prediction model for clinical trial completion using sparse data.
• To enhance the feature representation capabilities for survival analysis in clinical trials.
• To specifically address the prediction of trial completion in studies involving pregnant women.

Main Methods:

• Proposed a hybrid model combining a denoising autoencoder (DAE) with the DeepSurv model (DAE-DSR).
• Utilized DAE to extract robust feature representations from raw clinical trial data.
• Trained the DAE-DSR model on a dataset from ClinicalTrials.gov, focusing on trials in pregnant women.

Main Results:

• The DAE-DSR model effectively extracted meaningful and robust features for survival analysis.
• Achieved a C-index of 0.74 on the training dataset and 0.75 on the test dataset.
• Demonstrated superior performance and robustness compared to traditional Cox proportional hazards and standalone DeepSurv models.

Conclusions:

• The proposed DAE-DSR model significantly enhances survival prediction accuracy for clinical trials with sparse data.
• The model's ability to capture robust features improves generalization and predictive power.
• This approach offers a more reliable tool for predicting clinical trial completion, particularly in underrepresented groups like pregnant women.