An Accurate Data Preparation Approach for the Prediction of Mortality in ACLF Patients using the CANONIC Dataset

Area of Science:

• Hepatology and Medical Informatics
• Clinical Data Science

Background:

• Chronic liver disease incidence is rising in Europe, increasing the risk of Acute on Chronic Liver Failure (ACLF).
• ACLF presents high mortality due to multisystem organ failure, necessitating continuous patient monitoring and rapid treatment adjustments.
• Predicting ACLF patient mortality is crucial for effective clinical management.

Purpose of the Study:

• To validate a data preparation approach for predicting ACLF-associated mortality.
• To assess the utility of combining data imputation, class balancing, feature selection, and hyperparameter optimization.
• To evaluate the performance of machine learning models using optimized datasets for mortality prediction.

Main Methods:

• A multivariate Mutual Information (MI) based feature selection approach was employed to identify informative patient characteristics.
• XGBoost and Logistic Regression models were trained using prepared data from the CANONIC dataset.
• Model performance was evaluated on a test set for predicting mortality rates at various time points post-diagnosis.

Main Results:

• The data preparation approach, including MI-based feature selection, resulted in smaller, informative subsets of patient characteristics.
• Machine learning models trained on the prepared data demonstrated high accuracy in predicting ACLF patient mortality.
• In some cases, reduced feature sets led to improved prediction accuracy compared to broader characteristic sets.

Conclusions:

• The proposed data preparation strategy effectively enhances the prediction of mortality in ACLF patients.
• Optimized feature selection and machine learning model tuning are valuable for improving prognostic accuracy in critical care settings.
• This approach supports clinicians in making timely treatment decisions for ACLF patients by providing reliable mortality predictions.