Structural modeling of clinical factors for simultaneous validation and prediction of future conversion to mild cognitive impairment

Abstract

AbstractBackgroundIdentifying risk factors of future conversion from cognitively normal to mild cognitive impairment is crucial for dementia prevention and understanding the clinical processes that result in the development of cognitive deterioration. One challenge which is faced with this task is that, when identifying these factors, not only the individual contribution but also its coherency with other indicators must be measured.MethodTo map the relationship between clinical risk factors and conversion to mild cognitive impairment, we created and validated a set of structural equation and growth mixture models based on input from clinical experts in the field of dementia research. These models are powerful as they allow the researcher to test hypotheses between observed variables and their underlying latent constructs. In addition to verification of the factor structure, we integrated a method for predicting individual factors, such as mild cognitive impairment, on out‐of‐sample converters and controls directly on the structural model, a feature which is normally uncommon in structural equation modeling partly due to its sensitivity for model misspecification.ResultIn a study of 1213 elderly participants we found that these structural models perform similar to current state‐of‐the‐art logistic regression models in terms of Area‐Under‐the‐Curve (AUC) score on a 1‐year prediction (AUC > 0.9), yet offer a significant increase in flexibility and insight into the direct and indirect effects of endogenous variables on conversion to mild cognitive impairment. We also observe a substantial increase in predictive power for models fitted with baseline data compared to baseline with follow‐up visit for both 5 and 9‐year prediction on neuropsychological and demographic data alone (AUC increase +‐ 0.05), allowing us to monitor the development of cognitive health in subjects over time.ConclusionThe results from this study suggest a powerful technique for creating structural models with the ability to perform robust path analysis, longitudinal monitoring of cognitive health and prediction of future conversion to mild cognitive impairment. We expect that clinical experts can use this method to continue testing and validating new hypotheses in dementia research.