Explainable Machine Learning based Early Warning System
- Reference number
- SM21-0060
- Start and end dates
- 220401-241231
- Amount granted
- 1 166 000 SEK
- Administrative organization
- KTH - Royal Institute of Technology
- Research area
- Life Science Technology
Summary
Analysis of healthcare data with explanations provides a unique opportunity to develop novel technologies for accurate predictions and diagnostics of severe outcomes in hospitals. The main objective of the grant is to explore this unique opportunity. At Karolinska University Hospital (KUH), the applicant will help to realise explainable machine learning (XML) based solutions for medical caregivers. In addition, the medical caregivers will learn the use, efficiency and limitations of machine learning. Overall, the mobility grant will help to strengthen long-term cooperations between KTH, KUH and Region Stockholm for medical data analysis. In the host group of KUH, the applicant's role will go beyond their ongoing joint project on Deep Learning based Novel Early Warning System (DeepNEWS). Using time-series patient data, the DeepNEWS will provide analysis in real time, and aid the medical caregivers for clinical decisions and therapeutic interventions in critical time. The applicant will help to scale DeepNEWS using explainable machine learning (XML) where medical caregivers will understand reasons for algorithmic decisions. As a second application, use of XML will be further explored for identifying best medication candidates to slow cognitive decline for dementia patients. In addition, the applicant and the host group will jointly arrange two data science workshops for medical caregivers and an appropriate hands-on machine learning course.
Popular science description
Artificial Intelligence (AI) can be used to detect infections earlier than current clinical practice. Infections do not always show clear and specific symptoms, why doctors and nurses can miss the diagnosis. Once infection starts, the immune system begins to fight bacteria and viruses. As an effect, body parameters, such as heart rate, blood pressure, breathing pattern, body temperature, start to change slowly. While humans can not recognise the slow subtle changes, AI can. AI can also predict infection type and risk of future patient deterioration. Doctors and nurses then have precious time to initiate life-saving interventions early. This heralds use of AI for designing early warning systems in hospitals. Each patient at modern hospitals generates a huge amount of data. Such data includes time-series waveform data from physiological sensors, vital signs, electronic health records (EHR), lab results, radiology imaging, ventilator states, fluid and medication injections. Use of new non-invasive wearable sensors is on the increase for hospital patients. All the data contain clues to the patient's medical state and can be used for early detection of deterioration, such as the infection buildup, development of sepsis, respiratory failures, or cardiac arrest. Our long term vision is to use modern machine learning (ML), mainly deep learning - a core for data analysis in AI - to analyse such data in real time for early warning. Real time early prediction will save critical time for therapeutic interventions, much ahead of clearly visible clinical symptoms. Deep learning based novel early warning systems (DeepNEWS) will aid accurate medical decisions, saving lives. The automatic systems will perform real time data acquisition, storage, analysis, and provide automatic monitoring and appropriate warning notifications. To develop the DeepNews, it is imperative that machine learning experts spend quality research time in hospitals for performing a full scale system study. Machine learning experts need to clearly understand hospital data collection and management systems, mainly the patient database management systems and the appropriate IT infrastructures of hospitals. Further the experts need to clearly understand the needs of doctors and nurses, and how they analyze data and decide medical diagnostics. This inter-disciplinary research collaboration with active engagement of all partners will pave the path of DeepNEWS at Swedish hospitals in future.