Generative AI for Medical Digital Twin Using RAG: A Clinical Decision Support Approach

Abstract

Potential issues in modern clinical decision-making are varied data formats such as physician notes and lab results, which conventional systems are unable to work together. This imposes a heavy cognitive burden on clinicians, and there is an urgent need to find an AI solution that would be capable of processing such information. In response to this, the project creates a new clinical decision support system, a medical digital twin with Retrieval-Augmented Generation (RAG) framework. The system has been built on the MIMIC-IV dataset and the patient data is processed into hourly snapshots, which are then translated to natural language and processed using a MiniLM model and translated into a vectorized format. The most pertinent clinical information is retrieved in response to natural language queries by a RAG pipeline which is driven by Gemini Flash. It is combined with a digital twin architecture, which processes patient trajectories via clustering to provide predictive information. The main result is the system that transforms disconnected data into actionable knowledge and makes the clinicians less cognitively loaded. It can predict and recommend proactive treatment of ICU patients with partial data by citing the similar cases of the past, improving the quality and speed of the decision-making process.

Ähnliche Arbeiten