Prädiktive Analyse von zeitbasierten Gesundheitsdaten zur Früherkennung von Druckgeschwüren

On average, the worlds population is getting older due to steadily improving medical care. As a result, diseases with increased age-related risk are occurring more often and, therefore, older people are more frequently bedridden or in a wheelchair. For this reason, they are more likely to be exposed to prolonged pressure, which makes them more susceptible to pressure ulcers, also known as decubiti. At the same time, the incidence of diabetes is increasing in old age. This is accompanied by an increase in the number of diagnosed pressure wounds on the foot, known as foot ulcerations, which can be a consequence of diabetes. In the course of the KIPRODE project, a system is to be developed with the help of automated measurement methods based on artificial intelligence (AI) to detect the development of wounds at an early stage and ultimately help to prevent them. For this purpose, the patients pressure load, oxygen saturation, heart rate and skin temperature are recorded over time with sensors on the affected areas. The aim of this work is to develop and implement a software architecture that automates the processing of time-based aforementioned parameters and prepares it for anomaly detection and classification. In particular, errors in the recordings are corrected using different types of interpolation and the data is split into batches for later training. Both the scripts for the data pipeline and the neural networks are programmed in Python. As a result, a data structure was achieved that is suitable to store the data of the different studies and enables automated processing. Aspects of the sensors, such as sampling time or storage capacity, and the machine learning algorithms were taken into account in the development and implementation to enhance the performance of the software architecture. In addition, various convolutional neural network architectures are discussed that can be used for time-based anomaly detection and classification in the future. The ultimate goal is to build upon the developed architecture and to identify correlations between sensor readings and wound development that would allow prediction of the risk of pressure ulcer development. The foundation for this is laid in this paper through the implementation of the software system and the preliminary work for further analysis of the data on pressure ulcers is presented.