Generative adversarial networks for imputing missing data for big data clinical research.

Area of Science:

• Machine Learning in Clinical Research
• Data Imputation Techniques
• Big Data Analytics

Background:

• Missing data is a significant challenge in clinical research.
• Generative Adversarial Imputation Nets (GAIN) is a novel machine learning approach for data imputation.
• GAIN's efficacy in large, real-world clinical datasets remains unevaluated.

Purpose of the Study:

• Evaluate the accuracy of GAIN for imputing missing values in large clinical datasets with mixed variable types.
• Assess the computational efficiency of GAIN.
• Compare GAIN's performance against MICE and missForest.

Main Methods:

• Utilized two large real-world clinical datasets (diabetes and hypertension cohorts).
• Simulated missing at random data at 20% and 50% missingness rates.
• Measured imputation accuracy using Normalized Root Mean Square Error (NRMSE) and Proportion of Falsely Classified (PFC).
• Recorded computation time for each imputation method.

Main Results:

• GAIN and missForest were more accurate than MICE.
• GAIN demonstrated superior accuracy over missForest at 50% missingness.
• GAIN excelled in imputing skewed continuous and imbalanced categorical variables.
• GAIN exhibited significantly faster computation times (32 min vs. 1300 min for missForest on 50,000 cases).

Conclusions:

• GAIN is a more accurate and efficient imputation method for missing data in large clinical datasets compared to MICE and missForest.
• GAIN is robust to high missingness rates (up to 50%).
• GAIN's speed and accuracy make it a promising tool for big clinical data research.