SAFE-ICU: advancing AI-driven decision support in the intensive care units

Abstract

Integrating artificial intelligence into healthcare presents an unprecedented opportunity to address critical challenges in Intensive Care Units. The objective of this thesis is to create, validate, and deploy AI models to increase safety through the development of early warning systems for drug-drug interactions and predicting clinical decompensation of patients. Our first objective is creation of a data lake involving multimodal data, which combines vital signs, laboratory tests, treatment histories, patient files, and thermographic videos; not just images. The combination of structured and unstructured data contained in the datasets offers unmatched insight into intensive care unit patients, thereby enhancing the overall quality of research. The datasets used for prediction and diagnostics are pre-processed through various techniques like feature extraction, temporal alignment, etc. The second Objective aims at the creation of actionable, predictive, and management tools for sophisticated critical care of ICU patients. For example, ThermalShockNet, which was created for pediatric ICU settings and utilizes thermal videos and patient vitals to identify subtle physiological changes indicative of hemodynamic shock. This model uses data from thermal frames and vital signs to make predictions. When combined using deep learning, these predictions can reveal the onset of shock before it appears in a patient. ThermoGnose, a pipeline developed specifically for the early prediction of hypothermia, uses non-linear features drawn from patient monitoring data to estimate hypothermia onset and provide vital lead time for clinical intervention. SIgnose predicts shock risk by analyzing vital sign trends. When combined, this variety of approaches highlight how AI can address complex challenges in the ICU. The third Objective was to improve drug safety in the ICU. For this, the ContraIndicator tool was created. NLP processes the treatment chart and line listing of drugs to find probable adverse drug–drug interactions. The ContraIndicator insights assist clinicians in minimizing the chance of adverse reactions. The end Objective was SafeICU, a modular and explainable AI-driven real-time platform for the pediatric ICU. It brings together early warning systems along with safety and infection surveillance modules, which include vital signs, thermal imaging, medication, and microbiology results. At SafeICU, they help flag unusual Shock Index values in many ICUs. The ContraIndicator is on the lookout for drug interactions through the examination of ICU prescriptions. AMRtrack monitors the emergence of antimicrobial resistance by processing laboratory reports, mapping trend data, and informing infection control planning. When these modules are together, real-time alerts, risk scores, and infection monitoring are delivered into the ICU workflow. Overall, this thesis provides a clinically oriented framework for using AI in ICUs. It shows how models that can be multimodal, explainable, and deployable work together in real time to assist clinical decision making in high-acuity clinical settings. These advances create a pathway to the secure and scalable exploitation of AI in critical care and, eventually, precision medicine in health care that are rich in data.