Year: 2025
Type: dissertation
Abstract: The rapid integration of Artificial Intelligence (AI) into healthcare has generated both significant opportunities and critical challenges, particularly concerning transparency and trust in algorithmic decision-making. As many AI models, especially, those based on deep learning operate as “black boxes,” there is a growing demand for explainable artificial intelligence (XAI) techniques that make these systems interpretable and actionable in clinical settings. This thesis explores the comparative performance of various machine learning models and the utility of explainability methods in predicting 30-day Intensive Care Unit (ICU) readmissions using the MIMIC-IV dataset. A set of machine learning models including logistic regression, decision trees, random forests, gradient boosting machines, and neural networks were trained and evaluated using standard metrics such as Area Under the Curve (AUC), F1-score, and recall. SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) were applied to selected models to assess both global and local interpretability. The analysis reveals that ensemble models, particularly gradient boosting, achieve higher predictive accuracy. SHAP proved effective in identifying clinically relevant risk factors such as abnormal vital signs and laboratory values, whereas LIME provided individualized explanations suitable for point-of-care interpretation. The study contributes to the growing body of literature on XAI in healthcare by empirically comparing model performance and interpretability. However, it also identifies limitations, including the exclusive use of a single dataset, absence of external validation, and lack of clinician-involved usability testing. These limitations highlight the need for further research into the clinical integration, generalizability, and user-centered evaluation of XAI-based CDSS. Ultimately, this thesis underscores the importance of balancing accuracy with transparency to foster trust and adoption of AI systems in high-stakes clinical environments.
Source: Doria (University of Helsinki)
Author: Ahmed, Syed Sabbir
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7115382010
