Dräger produces technologies for medicine and security. Using the VitalDB dataset where there is a great count of sensor readings of thousands of patients available online, we modelled and trained an AI model that can predict whether there'll be a 'shock' in the coming 30 seconds / 2.5 minutes / 5 minutes. We reached test accuracies ranging from 70% to 90% using the methods we developed, and demonstrated them live using the aforementioned monitor. We had a doctor in the team who guided us towards measuring shock index (SI) as HRT / ART_SBP. For predictions we experimented with different models, but within the limited time and without access to GPUs, we could only prepare our LSTM model in time (even on that model we had made simple training mistakes, which we could only -to some extent- fix in the last minutes of the event), with which we reached the aforementioned accuracies. We also had working code to extract frequency information from some sensor readings using Mexican hat wavelet transformations, but even though we made use of as much multithreading as we could think, we couldn't transform all the necessary data within the given time. We also had trouble finding pretrained autoencoders/embedders with which we could embed the data in order to put it into more explainable models. My role in the team was mainly participating in AI brainstorming sessions and trying to implement respective transformations and hopefully run explainable models. The jury and other Dräger employers especially praised our team for demonstrating 'order in chaos' and shared expertise on the domain. The source code is confidential and thus cannot be shared here.

Project media ↗