Reorienting Latent Variable Modeling for Supervised Learning.

Area of Science:

• Psychometrics
• Machine Learning
• Statistical Modeling

Background:

• Supervised learning typically assumes clear, pre-defined outcomes, making outcome validation before prediction uncommon.
• Latent variable (LV) modeling is primarily used for inference, not prediction, necessitating a conceptual shift for its application in supervised learning.

Purpose of the Study:

• To outline methodological and conceptual adjustments for integrating LV modeling into supervised learning.
• To demonstrate the feasibility of using LV modeling for generating and validating prediction targets within a supervised learning framework.

Main Methods:

• Combining principles from LV modeling, psychometrics, and supervised learning to create an interdisciplinary framework.
• Utilizing flexible LV modeling to generate a pool of candidate outcomes.
• Systematically validating generated outcomes using clinical validators.

Main Results:

• Demonstrated that integrating LV modeling into supervised learning is achievable.
• Showcased the generation of numerous candidate outcomes via flexible LV modeling using data from the Longitudinal Assessment of Manic Symptoms (LAMS) Study.
• Illustrated how this exploratory approach allows for tailoring prediction targets using scientific and clinical insights.

Conclusions:

• The integration of LV modeling with supervised learning offers a novel approach to prediction target generation and validation.
• This interdisciplinary framework facilitates the development of more refined and clinically relevant prediction models.
• The study provides a methodological foundation for leveraging LV modeling's inferential power for predictive tasks.