Comparison of machine learning techniques with classical statistical models in predicting health outcomes

Area of Science:

• Biomedical informatics
• Machine learning applications in healthcare
• Predictive modeling for health outcomes

Background:

• Accurate mortality risk prediction is crucial for patient care and public health.
• Machine learning offers potential for improving health outcome predictions.
• Diverse data sources, including clinical records and population surveys, contain valuable prognostic information.

Purpose of the Study:

• To evaluate the performance of various machine learning techniques in predicting mortality risk.
• To compare prediction accuracy across different datasets (patient records vs. population surveys).
• To identify the most influential factors (data nature vs. algorithm) in mortality prediction.

Main Methods:

• Application of multiple machine learning algorithms: multilayer perceptron (artificial neural network), single layer perceptron, logistic regression, least square linear separation, and support vector machines.
• Utilized two distinct biomedical datasets: patient care records and a population survey.
• Incorporated diverse data features such as symptoms, medical history, lab tests, medications, health attitudes, and functional disabilities.

Main Results:

• High mortality prediction accuracy (Area Under the Curve [AUC] up to 0.89) was achieved with acute patient data.
• Fair prediction accuracy (AUC 0.70-0.76) was observed for six-year mortality in population survey data.
• The nature of the data was found to be more critical than the specific machine learning technique employed.

Conclusions:

• Machine learning models demonstrate significant potential for mortality risk prediction.
• Artificial neural networks (multilayer perceptron) exhibited superior performance, highlighting the value of capturing nonlinear relationships.
• Future research should focus on data quality and the integration of advanced algorithms for enhanced health outcome prediction.