Application of machine learning and deep learning approaches for prediction modeling with time-to-event outcomes in clinical epidemiology. Methods comparison and practical considerations for generalizability and interpretability

Abstract

Purpose

Clinical prediction models (CPM) are essential tools for diagnosis and prognosis in clinical epidemiology. Machine learning (ML) and deep learning (DL) approaches provide flexible methods that can complement regression-based methods for CPM when complex predictors such as clinical biomarkers are of interest. However, concerns have been raised on the ability of ML and DL to address desired properties of CPMs such as parsimony, generalizability, and interpretability.

Methods

In this study, we evaluated and applied selected regression-based, ML and DL approaches for time-to-event outcomes in a clinical study integrating protein biomarkers and lipids in an existing CPM for cardiovascular risk.

Results

We observed considerable advantages from the application of gradient boosting machines (GBM: C-statistic=0.72; Brier Score=0.052), which provided the best balance between model flexibility, discrimination, calibration, and parsimony, the latter being directly related to one of the model parameters (shrinking rate). Further, GBM results can be used for individual risk prediction, providing an interpretable tool for CPM implementation.

Conclusions

We compared ML and DL methods for CPM with time-to-event outcomes and discussed practical aspects of their implementation in clinical epidemiology including generalizability and interpretability. Adequately trained ML approaches can provide advantages in prediction modeling, especially when integrating complex predictors.