Explainable deep learning framework incorporating medical knowledge for insulin titration in diabetes.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Diabetes Management

Background:

• Deep learning (DL) models show potential in diabetes management but lack transparency due to their "black-box" nature.
• Existing explainable artificial intelligence (XAI) methods often overlook feature interactions and clinical domain knowledge.
• This limits the trustworthiness and real-world application of AI in complex medical environments.

Purpose of the Study:

• To develop an expert-guided XAI framework for transparent and trustworthy deep learning models in insulin titration.
• To address the limitations of current XAI methods by incorporating feature interactions and clinical expertise.
• To improve the accuracy and reliability of AI-assisted diabetes management.

Main Methods:

• Utilized two Electronic Health Record (EHR) cohorts of hospitalized patients with type 2 diabetes (T2DM).
• Introduced an expert-guided XAI framework using the Shapley Taylor Interaction Index (STII) to capture feature interactions.
• Refined the model iteratively through a doctor-in-the-loop (DIL) process, encoding clinical constraints.

Main Results:

• The STII-DIL model effectively explored interaction factors and reduced unreasonable explanations compared to other models.
• XAI system explanations showed strong alignment with expert clinicians' reasoning and improved correctness.
• AI-assisted insulin titration accuracy significantly improved for junior clinicians, with increased confidence reported by both junior and senior clinicians.

Conclusions:

• Presented an explainable deep learning framework combining post-hoc XAI and expert knowledge for insulin titration in T2DM.
• The framework provides transparent, expert-aligned explanations, enhancing decision-making accuracy and confidence.
• This approach may facilitate wider clinical adoption of AI tools in diabetes care.