Mortality risk score prediction in an elderly population using machine learning.

Area of Science:

• Epidemiology
• Machine Learning
• Biostatistics

Background:

• Parametric regression is standard for prediction in epidemiology.
• Machine learning algorithms like random forest and neural networks are emerging.
• Selecting the optimal prediction algorithm a priori is challenging.

Purpose of the Study:

• To apply the super learner ensemble method for predicting mortality risk.
• To evaluate the performance of super learner against individual algorithms.
• To assess the utility of super learner in epidemiological research.

Main Methods:

• Super learner, an ensemble machine learning approach, was utilized.
• The study involved 2,066 participants aged 54+ in the Study of Physical Performance and Age-Related Changes in Sonomans (SPPARCS).
• Prediction of death (risk score) was performed using cross-validated mean squared error.

Main Results:

• Super learner improved prediction over all single algorithms tested.
• Super learner outperformed neural networks by 44% in cross-validated mean squared error.
• Super learner achieved an R2 value of 0.201, approximately doubling the R2 of random forest.

Conclusions:

• Super learner offers a robust and improved approach for risk score prediction in epidemiological studies.
• Ensemble methods like super learner can enhance predictive performance beyond individual algorithms.
• This method provides a valuable alternative for mortality risk prediction.