Comparing machine learning algorithms for multimorbidity prediction: An example from the Elsa-Brasil study.

Area of Science:

• Public Health
• Computational Medicine
• Biostatistics

Background:

• Multimorbidity, the simultaneous occurrence of multiple chronic diseases, presents a significant global health challenge, impacting quality of life and increasing mortality.
• Early prediction of multimorbidity is essential for developing effective preventive strategies and integrating care.
• Current knowledge on predicting multimorbidity is limited due to the inherent complexity of modeling multiple conditions.

Purpose of the Study:

• To develop and evaluate cost-effective multimorbidity prediction models using machine learning.
• To compare the performance of various multilabel classification algorithms for multimorbidity prediction.
• To identify key predictors for early multimorbidity detection.

Main Methods:

• Utilized a machine learning approach with seven multilabel classifiers, including random forest and support vector machine-based models.
• Employed easily obtainable clinical predictors: sociodemographic, clinical, family history, and lifestyle data.
• Analyzed data from 15,105 participants in the Brazilian Longitudinal Study of Adult Health (ELSA-Brasil).

Main Results:

• The classifier chain with a random forest base classifier demonstrated superior performance (accuracy = 0.34, subset accuracy = 0.15, Hamming Loss = 0.16).
• Random forest-based classifiers generally outperformed support vector machine-based classifiers across different feature sets.
• Key predictors for multimorbidity included Body Mass Index (BMI), blood pressure, sex, and age.

Conclusions:

• Random forest-based classifiers are recommended for effective multimorbidity prediction.
• The study provides a foundation for developing practical, data-driven tools for early disease detection.
• Accessible clinical data can be leveraged for robust multimorbidity prediction models.