Aegion: Privacy-First Clinical Drug Safety Engine, Local LLM Enables Interpretable Medication Risk Analysis

Aegion is a privacy-focused clinical drug safety analysis system that combines local large language models (Ollama/Qwen 2.5) with a deterministic fallback mechanism to provide interpretable medication interaction analysis for healthcare institutions and patients. The project is maintained by Shyaam-04, hosted on GitHub with the link https://github.com/Shyaam-04/Aegion, and was released on June 16, 2026. Its core concept is "privacy first", aiming to solve the data privacy dilemma of medical AI.

Traditional cloud-based AI medication interaction analysis solutions face severe data privacy challenges: patients' sensitive data needs to be uploaded to third-party servers, violating medical data protection regulations such as HIPAA and GDPR, and raising concerns about privacy leakage. Aegion was developed to achieve data security while ensuring analysis accuracy through local processing.

Local LLM Inference Layer

It uses Ollama as the local model running platform, with Qwen 2.5 series models as the default. Reasons include open-source auditability, multilingual support (optimized for Chinese medical scenarios), and flexible hardware adaptation.

Deterministic Fallback Mechanism

When the LLM inference confidence is insufficient, high-risk drug combinations are involved, or knowledge blind spots are encountered, it automatically switches to a rule engine based on authoritative databases such as DrugBank and Lexicomp to ensure the reliability of key decisions.

Interpretability Design

Each analysis result is accompanied by a detailed reasoning chain: mechanism explanation, evidence level annotation, and alternative suggestion, helping doctors understand the AI's "thinking process".

• Outpatient Prescription Review: Real-time scanning of all medications used by patients (including prescription drugs, over-the-counter drugs, and health supplements) to warn of potential risks.
• Inpatient Management: Seamless integration with hospital intranets to provide real-time monitoring for patients with multiple coexisting diseases and polypharmacy.
• Pharmacy Medication Guidance: Provides professional consultation for community pharmacy pharmacists, especially suitable for elderly chronic drug users.
• Telemedicine Scenarios: Local processing allows patients to obtain professional medication assessments at home without uploading health data.

Privacy Computing Architecture

It follows the principle of "zero data out-of-domain": all inference is completed locally, no internet connection is required for core functions, and optional anonymized telemetry requires user authorization.

Modular Plugin System

Supports extensions such as drug database adapters, clinical guideline integration, and custom rule engines to adapt to the needs of different healthcare institutions.

Performance Optimization Strategies

Includes model INT8/INT4 quantization (reducing video memory usage), inference caching (millisecond-level response), and hot update of the drug knowledge base (no need to restart the service).

Aegion marks the evolution of medical AI from "cloud centralized" to "edge distributed":

• Compliance Breakthrough: Localized data eliminates legal risks of cross-border transmission and meets strict regulatory requirements.
• Cost Optimization: The marginal cost of local deployment is close to zero, saving high-frequency query costs for large hospitals.
• Technical Autonomy: The open-source architecture allows healthcare institutions to master core technologies and avoid vendor dependence.

The project demonstrates the potential of combining privacy computing with LLM, providing an example for privacy protection in medical AI, and is expected to become an important infrastructure in this field.