Machine learning methods to predict attrition in a population-based cohort of very preterm infants

Area of Science:

• Medical research methodology
• Biostatistics
• Machine learning in healthcare

Background:

• Timely identification of participants at high risk for attrition is crucial for efficient research resource utilization and early intervention.
• Conventional methods for predicting attrition may not fully capture complex interactions among predictors.
• Machine learning offers potential advantages in improving attrition prediction accuracy.

Purpose of the Study:

• To develop and compare predictive models for participant attrition in a very preterm infant cohort.
• To evaluate the performance of various machine learning methods against a conventional regression model.
• To identify key predictors of attrition in longitudinal research.

Main Methods:

• Applied conventional regression and eight machine learning methods (AdaBoost, Artificial Neural Networks, Functional Trees, J48, J48Consolidated, K-Nearest Neighbours, Random Forest, Logistic Regression) to predict attrition.
• Utilized data from 542 very preterm infants in the European Effective Perinatal Intensive Care in Europe (EPICE) cohort.
• Compared model performance using Area Under the Curve-Precision Recall (AUC-PR), Accuracy, Sensitivity, and F-measure for both fixed (Baseline) and dynamic (Incremental) predictor sets.

Main Results:

• Both Baseline and Incremental models showed good predictive performance, with AUC-PR ranging from 69 to 97.1.
• Random Forest consistently outperformed other methods across all follow-ups, achieving the highest AUC-PR values.
• Key predictors for attrition included birthweight, gestational age, maternal age, and length of hospital stay, consistent across models.

Conclusions:

• Machine learning, particularly Random Forest, provides robust and interpretable models for predicting participant attrition in research cohorts.
• These models can assist researchers in proactively identifying high-risk individuals, enabling targeted interventions to improve participant retention.
• The findings support the use of advanced analytical techniques to optimize the efficiency and success of longitudinal studies involving vulnerable populations.