German Clinical Trials Register

Acronym/abbreviation of the study

SURGICAL SWARM

URL of the study

No Entry

Brief summary in lay language

For the use of artificial intelligence (AI) in assistance systems, large amounts of training data are needed to train the systems for their use. In the field of medicine, these data are usually patient data such as age, image data or data on the clinical course. Such data could be used, for example, to train algorithms that identify patients at risk and thus lead to better care. Currently, only a few data sets are publicly available for the development of appropriate AI systems in surgery, and they are also of low diversity. They are also mostly generated in surgical procedures that are not very complex and have few complications. Furthermore, due to data protection regulations, direct exchange of sensitive patient data between centers is difficult. Decentralized AI systems enable collaboration between centers through indirect data transfer: instead of the actual data, only metadata of the AI models trained in a center are exchanged. In order for decentralized systems to make clinically relevant predictions about potential complications and surgical endpoints, these data must be collected from multiple centers. The study aims to investigate how precisely using decentralized AI methods, in particular swarm learning, a prediction of clinically relevant surgical endpoints (including complications) is possible when including multicenter (image) data (including minimally invasive surgery videos).

Brief summary in scientific language

The limited availability and diversity of training data is one of the major challenges in the development of clinically relevant artificial intelligence (AI) applications in medicine. This problem is particularly pronounced in the field of Surgical Data Science1 , where very few open access datasets exist for training and validation of AI models compared to other medical fields such as radiology, histopathology, or dermatology. Existing anonymized open access datasets of intraoperative surgical video data (e.g., the Cholec80 dataset2) are mostly from less complex and low-complication surgical procedures such as cholecystectomy. The amount and diversity of training data required to develop and validate generalizing and thus clinically meaningful AI systems can only be achieved through collaboration between multiple medical institutions, especially for complex and less frequently performed surgeries such as oncology procedures. However, the exchange of patient data between medical institutions is subject to data protection regulations, which in practice results in a variety of legal and logistical hurdles, especially when international cooperation partners are involved. In this project, these challenges will be addressed on a methodological level by applying decentralized machine learning methods. The goal is to predict surgical complications and clinical endpoints after surgical procedures based on categorical and numerical data as well as image data. image data. Methodologically, decentralized learning methods will be used, in particular Swarm Learning. This is a decentralized deep learning method that enables peer-to-peer collaboration between scientific partners without direct data exchange or data centralization. Models trained locally at collaborating institutions (Swarm Nodes), which use exclusively use data from the respective institution, are combined using blockchain-based communication within the Swarm network. Since only metadata is shared within this network, there is no exchange of sensitive patient data and also ensures equal collaboration among participating institutions. Previous studies have demonstrated that Swarm Learning outperforms locally trained models at individual sites in a variety of use cases, such as in the detection and classification of pulmonary nodules from chest radiographs, in the context of predicting spread patterns of COVID-19 or for predicting genetic aberrations in solid tumors based on routine histopathological sections. In the context of this study, it is planned to systematically apply Swarm Learning for the first time to medical (image) data in surgery.