MICCAI 2021 Daily - Tuesday

failure will occur using autoregressive modeling . This opens the door for continuous learning because I may need labels and I can’t ask a clinician to go through thousands of patient data and label them for me. The system implicitly performs auto labeling because it’s trying to understand what the representation is that is being learnt in this particular patient and how that representation differs from other patients. ” The idea for designing the best model architecture to create that representation came from a field known as AutoML, or meta-learning , where another neural net is used to design the neural net, so AI designing better AI, which is in itself an optimization process. Not only does this give the optimal AI for the data it is running on, but it gives a unique AI model for each ICU it will be deployed on because each unit has its own internal biases. But will busy clinicians, who have little time or desire to adapt to something new, accept an additional system being added to something that is already so complex? “ From our discussions with users, we know that clinicians and ICU staff don’t have time to change their current workflow and adhere to a new one, ” Samyakh replies. “ We can’t change it completely, but we can augment it to make it better and faster and ensure that anything we do works with current systems. We found that the ICU is one of the only places in the hospital that has standardized their data storage and data communication protocols . Every ICU uses a protocol called FHIR HL7 , and that means any device or system that is recording data or monitoring patients has to adhere to this protocol. That is a huge benefit for us because it means that all the 22 DAILY MICCAI Tuesday Startup Village