Machine learning methods have become a key technology in a wide range of application areas and medical technology is no longer imaginable without them. These techniques enable the processing of complex multivariate signals directly at the bedside and will thus lead to completely new therapeutic and diagnostic approaches. Machine learning offers enormous potential and represents a decisive competitive advantage in the development of innovative medical products.
The Data Science and Artificial Intelligence department at Fraunhofer IMTE investigates issues related to model-based and AI-supported medical software solutions. The spectrum of services ranges from the development of digital patient:inside models to the design of efficient network architectures and standard-compliant validation to deployment on the target hardware. An important focus is on novel edge computing solutions, which can be used to solve inference on deep neural networks directly on embedded systems. Our team provides advice and guidance in all phases of AI-enabled medical device development. Our experienced team is also at your side for the acquisition of clinical datasets, their annotation and appropriate data cleaning. Machine learning methods have become a key technology in a variety of application areas and medical technology is no longer imaginable without them. These techniques enable the processing of complex multivariate signals directly at the bedside and will thus lead to completely new therapeutic and diagnostic approaches. Machine learning offers enormous potential and represents a decisive competitive advantage in the development of innovative medical products.
The physiological processes and control circuits in the body are highly complex and subject to multiple interactions. This complexity results in special challenges for the development of novel medical devices and forms of therapy. At Fraunhofer IMTE, the application of highly detailed digital patient:inside models for the reproduction and simulation of complex physiological processes is being investigated. The patient models can be used, for example, to generate artificial data, especially when the collection of these data is difficult to perform clinically or involves high costs and complex studies. In addition, physiological insight is addressed as well as the improvement of signal processing algorithms. In order to ensure a credible use of the patient:inside models, the focus is on verification and validation in all development steps.
The use of modern AI methods often still relies on large computing power and corresponding hardware support. This poses particular challenges for medical devices, where inference must be solved in real time and with high accuracy.
At Fraunhofer IMTE, we therefore offer innovative solutions to enable machine learning algorithms to be used on embedded systems as well. This paves the way for a new class of biosensors that can also process complex multimodal signals and enable new therapeutic and diagnostic approaches.
To this end, resource-efficient AI methods based on exploitation of structures and model compression are being investigated. At the same time, the focus is on the use of innovative hardware solutions (e.g. TPUs, FPGAs).
The continuous recording of electrophysiological signals plays an increasingly important role in the treatment of patients. EMG, ECG and EEG are used to non-invasively measure and assess the body's internal physiological and pathological processes. In combination with monitoring data, critical situations such as cardiogenic shock can be predicted at an early stage and used as early warning systems.
At Fraunhofer IMTE, AI methods for biosignal processing are being developed with the aim of improving the prevention, diagnosis and therapy of diseases.
The methodological competence of Fraunhofer IMTE lies in the areas of Deep Learning, Uncertainty Quantification and Probabilistic Graphical Models. A special focus is on the development of real-time capable and robust methods that meet the high regulatory requirements of medical device development.
Especially in the medical context, it is crucial that AI-supported software solutions can robustly deal with uncertainties and communicate them to the user:s in an appropriate way. However, many current machine learning methods provide inadequate quantification of prediction uncertainty (both aleatory and epistemic uncertainty). In particular, it has been shown that deep neural networks often have poor calibration of class confidences. This challenge can be addressed by new scalable Bayesian approaches. Fraunhofer IMTE is investigating in particular Gaussian process-based approaches and their combination with deep neural networks. Such methods represent an essential basis for safe and robust AI systems.