New publication in the specialist journal „Somnologie“

A ZMI team consisting of our research associate Franz Ehrlich, who is the lead author, Dr. Miriam Goldammer, our Head of Data Science, and Prof. Dr. Martin Sedlmayr has published an article entitled „Evaluation of AI algorithms for clinical PSG evaluation using the example of apnoea detection“ in the specialist journal „Somnologie“.

The manual evaluation of polysomnographies (PSG) is time-consuming and subjective. Artificial intelligence (AI) offers promising opportunities to automate and support this process.

This article first explains the basics of machine learning in sleep analysis, including supervised learning and the importance of adequate data separation. This is followed by a presentation of the central requirements for data quality and data variety for practical application in the sleep laboratory. The article emphasises the importance of suitable evaluation metrics and the importance of investigating classification quality in different patient groups.

Based on these requirements, an overview of current algorithms for automated apnoea/hypopnoea detection is given. Despite advances in automated detection, there are still no fully evaluated algorithms that can replace manual assessment. This article aims to help evaluate current algorithms and formulate requirements for the future development of AI algorithms in sleep medicine.

This publication is part of our application focus on sleep research and sleep medicine and thus also of the MII research project Somnolink - Connected sleep data and decision support along the patient path for better care of Obstructive Sleep Apne, which aims to improve the diagnosis, treatment and adherence to therapy for obstructive sleep apnea (OSA). As a large amount of usable data from as many different clinics as possible is required for the fruitful creation of functional algorithms, multi-centre studies such as Somnolink, which provide data of high quantity and quality, are needed for development.

Click here for the full article: 

https://link.springer.com/article/10.1007/s11818-025-00505-7