Improving hospital decision making with interpretable associations over datacubes.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Mining

Background:

• Electronic Health Records (EHRs) contain complex patient data.
• Current classification methods may oversimplify patient conditions, leading to misdiagnosis.
• Interpretability and robustness are crucial for clinical decision support.

Purpose of the Study:

• To introduce a new Classification based on Association Rules (CAR) algorithm.
• To improve the interpretability and reduce the complexity of classification results from EHR data.
• To enable the study of patients holistically and provide reasoned alternative diagnoses.

Main Methods:

• Developed a novel CAR algorithm utilizing hierarchies defined over datacube dimensions.
• Modified the approach for obtaining and evaluating association rules in the classification process.
• Applied the algorithm to a cancer discrimination dataset for validation.

Main Results:

• The CAR algorithm demonstrated improved interpretability of results.
• The method effectively works with real-world EHR data, analyzing patients comprehensively.
• Successfully identified hidden relationships between diseases and provided reasoned alternative diagnoses, reducing misdiagnosis risk.

Conclusions:

• The proposed CAR algorithm offers a more reliable and interpretable approach to EHR data analysis.
• It enhances diagnostic accuracy by considering patient complexity and avoiding overfitting.
• The method shows significant utility in clinical applications like cancer discrimination.