HEART project: industry-driven & interdisciplinary R&D for secure smart health applications

By Koustabh Dolui, Chetanya Puri, Hans Hallez, Dimitri Van Landuyt, Sam Michiels, Bart Vanrumste, Greet Bilsen, Heereen Shim, Oleksandr Tomashchuk, KU Leuven (Contact author: Sam Michiels <sam.michiels@cs.kuleuven.be>)

Preventing excessive weight gain during pregnancy is critical because the risks for both mother and child are severe. Yet, doing so in an accurate way, from the early phase of pregnancy and without harming fundamental privacy rights is far from trivial.

Interdisciplinary research applied in a smart health business case

Excessive gestational weight gain during pregnancy can pose considerable risks. For the mother, it can lead to fetal macrosomia, maternal obesity or caesarean delivery. For the child, it can result in large-for-gestational-age infants or complications during labor and delivery. Gaining too little weight increases the risk of preterm birth or small-for-gestational-age infants.  The Institute of Medicine (IOM) provides guidelines on how much weight women in different BMI categories should gain during their pregnancy to deliver at a normal weight. Yet, without accurate predictions as from the early phase of pregnancy it becomes difficult to adjust before it is too late to intervene.

Within the H2020/Marie-Sklodowska-Curie project HEART[1], KU Leuven addresses this challenge in close collaboration with Philips Electronics Netherlands and the University of Macerata (Italy). HEART brings together an international team of six researchers with an interdisciplinary background in ICT, law and business modeling.

Promising R&D results

The first R&D results after 1 year are promising. First of all, PREgDICT[2] – a personalized prediction model and software prototype designed and developed at KU Leuven – can accurately predict the weight gain of pregnant women with only a few weight gain measurements available (e.g. early on in the pregnancy). The PREgDICT model builds upon pregnancy data from real life measurements and fine-tunes the data with individual weight gain information. First results show that weight category classification via PREgDICT is 10% more accurate than when state-of-the-art data processing techniques are used. Having available accurate predictions in an early phase of pregnancy can help medical professionals to intervene early and provide better pre-natal care. In addition, PREgDICT enables self-reporting by the pregnant individual, allowing for irregular data collection over time while still being able to combine individual samples with a large set of reference data collected from other people.

Secondly, the software architecture of PREgDICT is designed with privacy in mind that avoids uploading all weight gain data from a person’s smartphone to a central data analytics server in the cloud[3]. PREgDICT applies a federated learning approach where training data is not accessed directly by the application provider; data is locally used (i.e. on a person’s smartphone) to update the global prediction model, and only the model update is uploaded to the cloud (not the data itself). By collecting all model updates from individual application users, the model is improved centrally and again distributed to all users. Meanwhile, application users are assured that their personal data does not leave their smartphone.

Meanwhile, we are investigating how to lower the threshold for applying these technologies, both for application developers and end users. To support application developers, we have defined a data utility-driven benchmark[4] that enables them to select the most appropriate combination of data de-identification methods that satisfies privacy requirements while minimizing losses of data utility. To increase the added value for end users, we have defined a model[5] that uses direct feedback from the user via free text or voice input while being able to handle unusual expressions and misspelled words more robustly than traditional rule-based approaches.

The road ahead

The PREgDICT personalized prediction model, its privacy-aware distributed software architecture, the utility-driven benchmark and the accurate model to process free text or voice input are only some of the results delivered by the HEART team at KU Leuven, Philips, and the University of Macerata. Ongoing collaborations are investigating how to, amongst others, translate GDPR restrictions to software design, dynamically enforce the most appropriate privacy-enhancing technologies throughout the distributed health platform, and integrate results into a coherent solution for smart health applications.

[1] HEalth related Activity Recognition system based on IoT (HEART): This project has received funding from the European Union`s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 766139. This publication reflects only the authors’ views and the REA is not responsible for any use that may be made of the information it contains.

[2] Puri C., Kooijmann G., Masculo F., Sambeek SV., Boer SD., Luca S., Vanrumste B. (2019). PREgDICT : Early prediction of gestational weight gain for pregnancy care. In: 41st International Engineering in Medicine and Biology Conference Presented at the IEEE Engineering in Medicine and Biology Conference, Berlin, Germany, 23 Jul 2019-27 Sep 2019.

[3] Dolui K., Cuba Gyllensten I., Lowet D., Michiels S., Hallez H., Hughes D. (2019). Poster: Towards privacy preserving mobile applications with federated learning: the case of matrix factorization. Presented at the The 17th Annual International Conference on Mobile Systems, Applications, and Services, Seoul, Korea, 17-21 Jun 2019. doi: 10.1145/3307334.3328657.

[4] Tomashchuk O., Van Landuyt D., Pletea D., Wuyts K., Joosen W. (2019). A data utility-driven benchmark for de-identification methods. In Proceedings of the 16th International Conference on Trust, Privacy and Security in Digital Business (TrustBus 2019, http://www.dexa.org/trustbus2019), Linz, Austria, 26-29 Aug 2019.

[5] Shim H., Luca S., Lowet L., Vanrumste B. Position paper: From Free Text to Structured Information in Sleep Description Analysis Using Deep Learning Networks. In Book of Abstracts of the 7th IEEE Dutch Bio-Medical Engineering Conference, Egmond aan Zee, The Netherlands, 24-25 January 2019.