DILIGENT Clinical Assistant: Practical Exploration of Large Language Models in Drug-Induced Liver Injury Detection

This article introduces the open-source project DILIGENT-Clinical-Copilot, an LLM-based clinical assistant tool aimed at helping doctors detect and manage drug-induced liver injury (DILI). Positioned as "assisting rather than replacing" doctors, it addresses DILI diagnosis challenges through Retrieval-Augmented Generation (RAG) architecture, multi-dimensional risk assessment, and interactive decision support. It also faces challenges such as data privacy and insufficient clinical validation. In the future, it is expected to integrate multi-modal data and promote the popularization of medical AI through open-source collaboration.

Drug-induced liver injury (DILI) is a common and tricky adverse reaction in clinical medication, and one of the main causes of acute liver failure. Its diagnosis is extremely challenging: clinical manifestations mimic various liver diseases, and there is a lack of specific biomarkers; traditional diagnosis relies on the RUCAM scale, which requires rich experience from doctors and is time-consuming. Therefore, AI-assisted DILI detection has become a hot topic in clinical research.

DILIGENT-Clinical-Copilot is an open-source AI clinical assistant project developed by the CTCycle team, using LLM capabilities to provide doctors with real-time DILI risk assessment and diagnosis/treatment recommendations. The project's core positioning is "assisting rather than replacing" doctors; through intelligent information integration and analysis, it helps quickly and accurately identify potential DILI cases, representing the latest attempt of LLM in specialized disease management.

RAG-Based LLM Inference Engine: Integrates massive medical literature, drug instructions, guidelines, and real case data, deeply optimized for DILI scenarios, and retrieves the latest evidence in real time to avoid "hallucinations".

Multi-Dimensional Risk Assessment: Covers medication history analysis (identifying hepatotoxic drugs), time correlation assessment (complying with RUCAM time standards), clinical manifestation integration, and supportive exclusionary diagnosis.

Interactive Decision Support: Natural language interaction interface, returns structured assessment reports (risk level, evidence points, diagnosis/treatment recommendations), lowering the threshold for use.

Early Warning and Screening: Identifies easily overlooked DILI cases in outpatient/inpatient settings, provides drug interaction risk sorting and personalized monitoring recommendations for elderly/chronic disease patients taking multiple drugs.

Consultation for Difficult Cases: Serves as a virtual expert to provide evidence-based differential diagnosis ideas, citing research literature and guideline recommendations to support comprehensive judgment.

Medical Education: Provides a simulated case practice platform for residents/medical students, accelerating the cultivation of clinical thinking through feedback.

Data Privacy and Security: Processing real patient data requires strict compliance; how to train and improve the model while protecting privacy is an ongoing issue.

Insufficient Clinical Validation: Lacks large-scale prospective clinical trials to verify accuracy; strict regulatory approval is required for formal clinical use.

Model Interpretability: The "black box" nature of deep learning makes it difficult to meet doctors' needs for judgment basis; improving interpretability is a future direction.

Future Outlook: With the development of multi-modal technology, it is expected to integrate imaging and pathology data to achieve more comprehensive diagnostic support; the open-source feature supports global developer collaboration, accelerating technology maturity and popularization, and helping improve medical quality in resource-limited areas.

Conclusion: DILIGENT brings new possibilities to DILI detection. AI will not replace doctors, but such intelligent assistants are becoming indispensable tools in clinical practice, worthy of attention and participation from practitioners and researchers.