Evaluating Performance and Interpretability of Machine Learning Methods for Predicting Delirium in Gerontopsychiatric...

Area of Science:

• Geriatric Psychiatry
• Computational Medicine
• Artificial Intelligence in Healthcare

Background:

• Delirium is an acute mental disturbance common in hospitalized elderly patients.
• Clinical assessment tools like the Delirium Observation Screening Scale (DOSS) and Confusion Assessment Method (CAM) are currently used.
• Detecting delirium accurately is crucial for timely intervention and improved patient outcomes.

Purpose of the Study:

• To evaluate the performance of various machine learning (ML) approaches for delirium detection.
• To analyze the effectiveness of ML models using DOSS and CAM data in geropsychiatric settings.
• To compare different ML classification methods and assess their interpretability.

Main Methods:

• Utilized DOSS and CAM questionnaire data as features for ML models.
• Trained and compared seven popular classification algorithms.
• Investigated the impact of sampling techniques (down-sampling, up-sampling, hybrid) on imbalanced datasets.
• Used ICD-10 diagnoses of delirium as the target class variable.

Main Results:

• Random Forest demonstrated high predictive accuracy, effectively handling imbalanced data.
• Advanced ML methods show strong potential for delirium detection.
• A trade-off exists between model performance and interpretability.
• High predictive performance is prioritized for clinical applications in electronic health record systems.

Conclusions:

• Machine learning, particularly Random Forest, offers a promising approach for automated delirium detection.
• Model performance is a critical factor for clinical adoption, potentially outweighing interpretability in electronic settings.
• Further research may be needed to balance predictive power with the need for explainable AI in healthcare.