Karine Zeitouni, Yehia Taher (Laboratoire DAVID, Université Versailles Saint-Quentin-en-Yvelines) and Cédric Gouy-Pailler (Laboratory of Data Analysis and Systems Intelligence,CEA List) have decided to combine their skills to offer the scientific community a simple new tool to develop powerful algorithms capable of managing data flow problems. This tool will be applied in particular in the medical field in collaboration with Philippe Aegerter (Inserm UMR 1168) and Marc Fischler (Hôpital Foch, Université Versailles Saint-Quentin-en-Yvelines).
A generic but cutting-edge streaming platform
The scientific community is developing algorithms to manage data flows. Industrialists, on the other hand, are seeking to analyse the subject in a more applicative way. In IT, extremely powerful tools in terms of data throughput and robustness are being developed. "With the StreamOps project, we want to position ourselves at the interface of algorithmic, business and software aspects to offer all players a generic streaming platform that is at the forefront of algorithms," explains Cédric.
Indeed, StreamOps' ambition is to simultaneously meet the following objectives:
- Detection performance (reactivity, accuracy), information compression performance, consideration of data confidentiality
- Consideration of problems related to real data (missing data, sensor problems)
- Easy integration of new algorithms
- Operational robustness (high data rate, node failure robustness)
Data flows from environmental and health sensors
Karine has been working with Yehia for several years on the Polluscope project (ANR) and with Philippe as part of ACE-ICSEN, an IRS project at the University of Paris-Saclay. They met Cedric at the Center for Data Science and decided to raise this issue of data management, data analysis and Machine Learning on IoT data together.
On the one hand, StreamOps will be based on a sample of Polluscope data collected by portable multi-sensor boxes as part of a participatory collection. The objective of Polluscope is to analyze, in all dimensions, all pollution information to characterize an individual's exposure to air pollution. The SteamOps project will contribute to the analysis and Machine Learning of these data flows.
On the other hand, a connected multi-signal physiological sensor (patch) is in parallel about to be tested by Philippe and Marc on medical monitoring in preoperative and especially post-operative care. The idea is to stick a multi-sensor patch on the thorax of the operated patient to monitor him remotely and permanently during the days following the operation, in order to anticipate the risks of complications and to trigger the alert in good faith thanks to an intelligent decision support system, without having to block the patient in a specialised medical unit.
The objective in StreamOps is to use these two types of data to create a generic application.
Position yourself at the interface
Karine adds: "We will develop new algorithms that will interface between a community that sees data as time series and analyzes it from a historical perspective, and another that sees the IoT as a data flow and analyzes all of this data dynamically as it is being recorded. "The objective of StreamOps is to develop methods and algorithms to consider the temporal order in data flows.
Cédric regularly works with manufacturers who are also interested in the possibility of having automatic tools to process the data that come in flow.
"It's not about offering one more platform," explains Karine, "but an integrating platform. "The StreamOps team plans to collaborate with Albert Bifet, Telecom ParisTech, who developed the MOA (Massive Online Analysis) platform to create a compatible platform and achieve synergies.