A modular and interpretable framework for tabular data analysis using LLaMA 7B: Enhancing preprocessing, modeling,...

Area of Science:

• Healthcare Informatics
• Machine Learning Applications
• Artificial Intelligence in Medicine

Background:

• Healthcare systems face inefficiencies due to missed medical appointments, impacting resource allocation.
• Predictive modeling for patient no-shows is crucial for optimizing healthcare operations.
• Existing methods often require extensive manual data preprocessing.

Purpose of the Study:

• To introduce a novel, local LLM-assisted pipeline for automating semantic preprocessing in tabular prediction tasks.
• To evaluate the effectiveness of using LLaMA 7B for preprocessing tasks like column renaming and data cleaning.
• To assess the performance of classical machine learning models following LLM-guided preprocessing.

Main Methods:

• Development of a local pipeline integrating LLaMA 7B for semantic preprocessing and XGBoost for classification.
• Application of the pipeline to the Medical Appointment No-Shows dataset, including analysis, transformation, and explainability.
• Utilizing SHAP (SHapley Additive exPlanations) for model interpretability.

Main Results:

• The LLM-assisted pipeline achieved 80% overall accuracy with an XGBoost classifier.
• High F1-score (0.89) for the majority 'Show' class, but a low F1-score (0.03) for the minority 'No-show' class due to class imbalance.
• AUC-ROC of 0.65 and precision-recall AUC of 0.87, with SHAP identifying waiting days, age, and SMS notifications as key predictors.

Conclusions:

• Local large language models can effectively enhance preprocessing and interpretability in tabular prediction workflows.
• Classical supervised models remain vital for the final prediction tasks in this hybrid approach.
• The demonstrated pipeline offers an efficient and deployable solution for improving healthcare operational efficiency through predictive analytics.