Abstract:
Given an aging population, shortage of nursing staff and a continuously increasing workload, automation in the medical sector is an important aspect of future intensive care. Although automation and machine learning are current research topics, progress is still very limited in comparison to other application areas. Probably one of the most serious problems is data shortage in a heterogeneous landscape of medical devices with limited interfaces and various protocols. In addition, the recording of data or, even more so, the evaluation of automation is limited by a complex legal framework. Given these complications and the sensitive legal nature of medical records, only very limited data is accessible for further analysis and development of automated systems. For this reason, within the context of this thesis various solutions for data acquisition and automation were developed and evaluated concomitant to two clinical studies utilizing a large animal model in a realistic intensive care setting at the University Hospital Tübingen. Foremost, to overcome the problems of data availability and interconnection of medical devices, a software framework for data collection and remote control using a client-server architecture was developed and significant amounts of research data could be collected in a central database. Furthermore, a closed-loop controller based on fuzzy logic was developed and used for management of end-tital CO2, glucose, and other parameters to stabilize the animal subjects during therapy and reduce caregivers’ workload. In addition to the fuzzy controller, closed-loop management for temperature and anticoagulation could be established by developing hardware interfaces for a forced-air warming unit and a point-of-care analysis device, respectively. Besides further reduction of caregivers’ workload, such systems can provide additional patient safety and allow management in settings where human supervision may not be present at all times. One general and encountered problem for closed-loop control in a medical setting is limited availability of measurements, especially if manual blood withdrawals are required. As an initial step to address this problem, measured parameters from other devices as potential surrogates were evaluated in a comparison between different regression approaches. The required training data, a matched set of blood gas and monitoring parameters, was obtained by utilizing a developed algorithm for automated detection of withdrawal events. Yet, besides any specific implementations and analysis, many general aspects regarding the physical implementation of such a system and interaction with caregivers could be evaluated in the experimental setting and might guide further development of clinical automation.