Time to 01:00 pm Add to Calendar 2024-10-23 12:00:00 2024-10-23 13:00:00 Predicting Long-Term Gestational Weight Gain Using Early-Stage Data: Leveraging Imputation Models for Missing Data HHD101 Population Research Institute hxo5077@psu.edu America/New_York public
Location HHD101
Presenter(s) Our speaker for this week is Young Won Cho, a doctoral student in HDFS (Human Development and Family Studies).
Description

Abstract

Excessive gestational weight gain poses significant health risks to both pregnant women and their children. Early prediction of long-term weight gain trajectories enables timely interventions to prevent overweight. However, traditional time series models require extensive datasets, making them less effective with minimal early-stage data, and their accuracy declines over longer forecasting windows. In contrast, growth curve models can leverage group-level effects to predict individual trajectories but struggle with too many covariates or missing data. Imputation-based approaches offer a promising alternative by filling in missing values and learning relationships among multiple variables simultaneously. This study investigates the feasibility of using imputation models to predict long-term gestational weight gain (specifically in the third trimester, from 28 weeks onward) using early-stage data (12-17 weeks of pregnancy). Data were collected from 12 pregnant women over a 10–40-week period, including gestational weight and covariates measured daily (e.g., energy intake, sleep) and weekly (e.g., physical activity intention). Three methods were compared: a univariate time series model (Kalman filter for local linear trend), a growth curve model, and an imputation-based approach using MICE (multiple imputation chained equations) with predictive mean matching. Each model was evaluated on its ability to predict weight gain using leave-one-out testing with RMSE as the evaluation metric. The univariate time series model showed low RMSE for short-term predictions but often produced unreasonable long-term forecasts. The growth curve model provided more reasonable predictions; however, including weekly variables was problematic because missing data in those variables resulted in substantial data reduction, affecting overall prediction quality. And it underperformed when less relevant variables were included. In contrast, the MICE-based imputation model yielded the most accurate predictions, particularly for the third trimester, with improved performance when incorporating weekly measured psychological variables, such as physical activity intention.

Contact Person Hyungeun Oh
Contact Email hxo5077@psu.edu