Embedded AI for Real-time Health State Assessment and Treatment Recommendation in Rescue Operations

Rescue emergencies are generally quite strenuous and challenging because they deal with human lives in a situation where it is difficult to apprehend precisely the health status of a distressed patient due to personal and work limitations. Often the rescue personnel have to deal with high levels of mental and physical stress, trauma, and emotional strain during rescue operations, especially when dealing with injured and vulnerable victims. This certainly impacts their decision-making abilities and overall well-being. Time, the main constraint in such rescue situations, also plays a vital role in decision-making. Therefore, it is necessary to recognize relevant situations i.e. health complications of the rescue patients on site, and to take appropriate first aid measures. The situation may change during the further course of initial treatment, with rescue workers judging this primarily by the visible condition of the emergency patient, the data from medical equipment (e.g., ventilator, ECG), and the mission description from the control room. Such changes are required to be responded to immediately, for example by resuscitation or appropriate medication. That's why, waiting for results from medical tests like MRI and ECG which are time-consuming and not suitable for emergency cases is not considered suitable for rescue cases. Considering these technical constraints, the doctoral thesis focuses on employing artificial intelligence (AI) models in two aspects that can expedite and improve the rescue process. Firstly, for the diagnosis of health complications in rescue situations and secondly to identify the correct medications as a part of initial treatment. A major part of this research is focused on advanced data analysis techniques that were used to extract information from 12 years of rescue records of 273,283 cases in the German city of Siegen-Wittgenstein. The initial data received from the rescue station was raw and in many cases contained incomprehensible information for which Natural Language Processing (NLP) techniques were applied to extract and interpolate relevant attributes. Subsequently, a detailed method for creating various AI models to promptly detect six key complications— Cardiovascular, Respiratory, Psychiatric, Neurological, Metabolic, and Abdominal—was conducted and is detailed in this dissertation. To develop the detection models for each complication, Artificial Intelligence(AI) algorithms like machine learning including both classical and deep learning approaches were used. To train these models attributes like patients' medical history, health diagnoses including neurological assessment, vital signs, initial impression of the rescue personnel, administered medications, and other treatment paths were used. During the course of development, one primary objective was to identify the model achieving the greatest accuracy and precision. Based on this research, Extreme Gradient Boosting (XGB) and Random Forest (RF) algorithms were found as the most promising, showcasing accuracy rates ranging from 80\% to 96\%. After recognizing health complications, further research was done to find out if AI can also be implemented to determine possible medications based on detected complications and patients' health vitals. The result achieved from it also was impressive with accuracy close to 80\%. AI models are further tested by deploying them into various accelerators, such as ARM processors, FPGAs and microcontrollers, to evaluate their performance based on inference time. The overall focus of this research is to overcome the rescue challenges in real-time by recognizing rescue situations and improving the quality of care and efficiency of rescue personnel.