Generative AI for clinical documentation
There is a consistent increase in the clinical documentation burden. In this project, we want to explore the potential of generative artificial intelligence (AI) to support the clinical documentation process.
Factsheet
- Schools involved School of Engineering and Computer Science
- Institute(s) Institute for Patient-centered Digital Health (PCDH)
- Research unit(s) PCDH / AI for Health
- Funding organisation Innosuisse
- Duration 04.02.2025 - 30.06.2026
- Head of project Prof. Dr. Kerstin Denecke
- Partner Cistec AG
- Keywords Large language model, Medical documentation, Artificial Intelligence
Situation
Time efforts spent into documentation tasks limit the time available for clinical decision making, and interaction with patients. This excessive focus on documentation often interferes with direct patient care, with approximately 80 % of physicians acknowledging that the time and effort required for these tasks interferes with their ability to provide quality care. On average, physicians spend nearly two hours per day on documentation outside of regular working hours, with those participating in value-based payment models reporting even higher burdens.
Course of action
The project developed and evaluated an LLM-based concept for the automated quality checking of clinical documentation, focusing primarily on identifying inconsistencies in clinical reports. We began by creating a systematic classification scheme that identifies seven different types of inconsistency. For the evaluation, we cleaned up and structured a dataset provided by the industry partner, and used an agent-based approach to specifically insert controlled inconsistencies. Subsequently, we implemented and tested various prompting and detection methods, which were then integrated into a modular software architecture. Finally, we developed evidence verification procedures to trace each identified inconsistency back to exact passages in the text.
Result
The project demonstrated the technical feasibility of performing an LLM-based system to check for inconsistencies in clinical documentation. It produced a prototype system that detects and localises inconsistencies in medical texts, linking anomalies that it identifies to verifiable source citations. Key results include the development of a scientifically sound inconsistency classification scheme with seven categories, an agent-based method for generating controlled test data, and an evidence loop that improves verifiability while reducing hallucinations.
Looking ahead
In future work, we will apply and systematically evaluate the concepts from the project on larger, more realistic datasets. We will place a special emphasis on expanding the framework to identify missing or incomplete information in clinical reports. While we already described this as a concept during the project, we were unable to validate it in full due to a lack of data. In addition, we will continue our efforts to integrate the methods developed into clinical documentation and quality assurance processes.
