K-nearest neighbor algorithm for imputing missing longitudinal prenatal alcohol data

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning

Background:

• Longitudinal studies on pregnant women often face challenges with missing alcohol consumption data.
• Inaccurate or incomplete data can bias findings in studies of prenatal alcohol exposure.
• Robust imputation methods are crucial for reliable analysis of such data.

Purpose of the Study:

• To evaluate the effectiveness of the K Nearest Neighbor (k-NN) machine learning algorithm for imputing missing daily alcohol consumption data.
• To assess the accuracy of k-NN imputation in a large prospective cohort of pregnant women (Safe Passage study).
• To determine optimal parameters for k-NN imputation to minimize error.

Main Methods:

• Utilized data from the Safe Passage study (n=11,083) with missing alcohol consumption data (11.4%).
• Applied the k-NN algorithm, weighting distances and matching by day of week for imputation.
• Validated imputation accuracy by randomly deleting data segments and comparing imputed to actual values.

Main Results:

• The k-NN algorithm with 5 nearest neighbors (K=5) and 55-day segments yielded the lowest imputation error.
• Imputed values closely matched actual values for 64% of deleted segments and were within +/-1 drink/day for 31%.
• Imputation accuracy demonstrated variability across study sites due to differences in drinking patterns and missing data proportions.

Conclusions:

• The K Nearest Neighbor (k-NN) algorithm offers a highly accurate method for imputing missing alcohol data in longitudinal pregnancy studies.
• k-NN imputation can significantly enhance the quality and reliability of data in studies examining alcohol use during pregnancy.
• This machine learning approach provides a valuable tool for addressing data gaps in sensitive public health research.