RINet: synthetic data training for indirect estimation of clinical reference distributions

Area of Science:

• Clinical chemistry and laboratory medicine
• Biostatistics and data science
• Machine learning in healthcare

Background:

• Indirect methods estimate clinical reference intervals (RIs) using statistical analysis of routine testing data.
• Supervised learning shows promise but is limited by real-world data constraints.
• Synthetic data offers advantages for developing and benchmarking indirect RI estimation methods.

Purpose of the Study:

• To develop and evaluate deep learning models for indirect estimation of reference distributions (RDs) and RIs.
• To leverage synthetic data for training models capable of handling both univariate and bivariate clinical data.
• To compare the performance of these models against existing indirect RI estimation algorithms.

Main Methods:

• Trained two convolutional neural networks (CNNs) using synthetic data: one for univariate and one for bivariate data.
• The bivariate CNN was designed to predict covariance between clinical analytes.
• Evaluated model performance on both synthetic and real-world clinical datasets, comparing against four alternative algorithms.

Main Results:

• CNN model predictions closely matched directly estimated RIs and RDs in real-world and synthetic data.
• Models outperformed GMM, refineR, reflimR, and RINet_v1 in indirect RI estimation.
• Predicted multivariate reference regions (MRRs) demonstrated improved coverage of healthy patients and reduced region size compared to univariate RIs.

Conclusions:

• Training deep learning models with synthetic data is a viable strategy for accurate indirect RI estimation.
• This approach effectively addresses limitations associated with real-world data and traditional univariate RIs.
• The developed models offer a data-driven solution for precise RI estimation in both univariate and bivariate contexts.