Effect of Missing Data Imputation on Deep Learning Prediction Performance for Vesicoureteral Reflux and Recurrent...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Pediatric Nephrology

Background:

• Missing data is a significant challenge in clinical datasets, particularly for conditions requiring long-term patient follow-up like vesicoureteral reflux (VUR) and recurrent urinary tract infections (rUTI).
• Hospital records for VUR and rUTI often exhibit substantial missingness across demographic, clinical, laboratory, and imaging variables, hindering accurate diagnosis and analysis.
• Deep learning (DL) methods offer potential solutions for handling datasets with high proportions of missing observations.

Purpose of the Study:

• To compare the performance of multiple imputation techniques, specifically MICE (Multiple Imputation by Chained Equations) and FAMD (Factor Analysis of Mixed Data), in conjunction with deep learning for the differential diagnosis of VUR and rUTI.
• To evaluate the effectiveness of DL-based approaches in managing datasets with significant missing values for pediatric urological conditions.

Main Methods:

• A retrospective cross-sectional study involving 611 pediatric patients (425 with VUR, 186 with rUTI) was conducted, with an observed missing data ratio of 26.65%.
• Deep learning models were implemented using CNTK and R 3.6.3 to analyze 34 features, including physical, laboratory, and imaging findings.
• The study compared the performance of DL models after applying MICE and FAMD imputation techniques versus DL models capable of handling missing data directly.

Main Results:

• Deep learning utilizing the MICE algorithm achieved the highest performance in the differential diagnosis of VUR and rUTI, yielding 64.05% accuracy, 64.59% sensitivity, and 62.62% specificity.
• The FAMD algorithm, using 3 principal components, resulted in a DL model performance of 61.52% accuracy, 60.20% sensitivity, and 61.00% specificity.
• DL-based approaches demonstrated the capability to process datasets without prior imputation or omission of missing values, though performance was enhanced when combined with imputation techniques.

Conclusions:

• Combining deep learning with appropriate missing data imputation techniques, such as MICE, significantly enhances predictive performance for the differential diagnosis of VUR and rUTI in pediatric patients.
• DL models show promise in managing complex clinical datasets with high missingness, offering a valuable tool for improving diagnostic accuracy in pediatric urology.
• The findings suggest that MICE imputation followed by DL analysis is a superior strategy compared to FAMD imputation or DL models handling missing data alone for this specific clinical problem.