Falls are among the most common and dangerous events in old age and are the second leading cause of unintentional injury-related deaths worldwide: between 28 and 35 percent of people over the age of 65 are affected each year. Early detection and prevention of such risks are therefore of central importance in nursing care, but staff shortages often make it difficult for professionals to assess the risk in a timely manner. A research team involving the Einstein Center Digital Future (ECDF) has now investigated the extent to which artificial intelligence (AI) can support existing fall risk assessment and enable nursing staff to take more targeted preventive measures. However, the development of fall risk AI is also relevant for other AI applications: Using this example, the researchers have developed processes that allow AI models to be developed even under the strict regulatory requirements in the healthcare context.

The study, which is based on data from a university hospital and a geriatric hospital in Germany, comprehensively evaluated the potential of AI models to predict fall risks. With over 940,000 patient records, it is one of the largest studies on this topic to date. The researchers pursued three central questions:

• Can AI improve the accuracy of predicting fall risks?
• How can AI models be trained securely across different hospitals?
• Are these models fair—i.e., independent of age or gender?

"To answer these questions, we trained AI models using various centralized and decentralized training paradigms and compared them with each other. In decentralized approaches, data from both hospitals was combined without exchanging sensitive patient information. To do this, we used federated learning and swarm learning, among other methods," explains Daniel Fürstenau, ECDF board member and part of the research team. The performance of the models was then compared with the rule-based systems used in clinical practice. In addition, analyses were carried out on the fairness of the models across different demographic groups.

The result: regardless of the training paradigm, the AI systems significantly outperformed the rule-based conventional methods in all scenarios of the retrospective study. For the geriatric hospital, the model achieved a prediction quality (AUC) of 0.735, and the model for the university hospital even achieved 0.926. However, in contrast to frequently cited studies, none of the decentralized learning approaches showed any improvement. This result is an important signal for the practical suitability of decentralized AI model development. In addition to the generally improved fall risk assessment of AI, another positive result is that fall risk assessment works equally well across gender groups and can be considered fair. However, minor differences occurred in the prediction quality for a few age groups.

“Our results show that AI-based models can make an important contribution to improving patient safety,” said Daniel Fürstenau. “At the same time, it is clear that issues of fairness and data representation are crucial for developing truly generalizable and equitable systems. We are pleased that the processes we have developed can also serve as a blueprint for other AI researchers for the development of AI technology in the healthcare sector—in compliance with strict regulatory requirements.” 

About the study: //here