The project

Sentinel, a project within the Penta-Euripides² Eureka Clusters, was a collaborative research initiative dedicated to innovating patient monitoring. While vital signs are continuously tracked, biomarkers such as glucose and lactate are still measured manually through invasive blood draws. Sentinel aimed to close these gaps in hospitals’ early warning systems (EWS) by developing a wearable device capable of semi-continuous hybrid sensing, supported by integrated electronics, algorithms, manufacturing technology and verified prototypes.

By ensuring semi-continuous, non-invasive monitoring in an autonomous and hybrid manner, Sentinel has the potential to decrease complications and mortality, reduce the number of permanent disabilities, and thereby improve the quality of life for patients while reducing healthcare costs. 

The consortium brought together a diverse group of academic institutions, technology companies and hospitals, including Philips Electronics Nederland BV, AZ Turnhout, Catharina Ziekenhuis, Eindhoven University of Technology, Etteplan, Jobst Technologies GmbH, Micronit BV, Sapienza University – Dept. Mech. Aerosp. Eng and Verhaert. Verhaert’s multidisciplinary expertise in adhesives, electronics, and especially Verhaert Digital’s experience in data acquisition and AI-powered data labeling played a key role in helping achieve this ambitious goal.

The challenge

The Sentinel project faced a multi-faceted challenge. While Verhaert’s Strategic and Product teams focused on developing a reliable physical clinical data collection device, including a Pressure Sensitive Adhesive (PSA) Human Factors Engineering (HFE) selection tool to find the right adhesive for long-term skin contact, the Digital team developed a system to support clinical decision-making based on this clinical data.

A critical component of this digital challenge was the development of robust AI algorithms to interpret sensor data to provide early warnings. However, the reliability of any AI model is fundamentally dependent on the quality of the data it’s trained on. A significant challenge was determining whether a given measurement from the wearable device was of good enough quality or if it was too corrupted by noise or other artifacts. This is a common problem for wearable devices, as data integrity is often compromised by patient motion in uncontrolled, real-world environments.

Traditionally, this quality assessment requires a huge effort and time of medical experts to manually review and label vast amounts of data. This process is not only slow and expensive but also diverts valuable time away from patient care. The core challenge for Verhaert was to find ways to automate this data quality assessment and labeling process as much as possible. We needed to build a system that could intelligently assess data quality in real-time and continuously learn and improve, all while minimizing the time investment required from medical professionals.

The solution

Verhaert Digital’s solution for effective clinical data assessment is the Quality Index Toolbox (QIT), a sophisticated platform designed to streamline and automate data quality assessment and labeling. The QIT is built into a flexible and configurable tool for labeling data and integrates two key machine learning concepts: Automated Machine Learning (AutoML) and Active Learning.

  1. Automated machine learning: To shield end-users from the technical complexities of AI, we integrated an AutoML framework. This system automates the entire machine learning pipeline—from data preparation and model selection to training and evaluation. It allows for the rapid generation and deployment of new quality-assessment machine learning models without requiring deep AI expertise from the user.
  2. Active learning: To minimize the burden on medical experts, the QIT incorporates an Active Learning loop. Instead of requiring experts to label all data, the system intelligently identifies only the most uncertain or informative samples that the current model struggles with. These select samples are then presented to the expert for labeling via a user-friendly interface. This targeted approach significantly reduces the manual workload while maximizing the model’s learning efficiency.

The machine learning models created by this modular and generic framework can be exported, deployed, and also automatically generate a quality index for clinical time-series data, effectively labeling it as ‘good’, ‘medium’ or ‘bad’.

“Project Sentinel wins the PENTA-EURIPIDES² Innovation Award 2023 with a solution to improve patient’s care”

Read the full article

Impact & results

Verhaert’s multidisciplinary team successfully delivered on all its work packages. Key achievements include the PSA HFE selection tool, the design and realization of 2 integrated devices for clinical data collection, and, most importantly, the Quality Index Toolbox (QIT).

The QIT was successfully validated using public clinical (ECG) datasets, proving its ability to automate model generation and data quality assessment/labeling, and to save significant time. The project achieved its goals of minimizing expert input and abstracting away the technical complexity of AI model creation.

Because the QIT was designed as a generic and modular framework, its utility extends far beyond the Sentinel project. It can be adapted for other data types (like images) and integrated into future projects that require fast and reliable data quality assessment. The Sentinel project has thus produced not only valuable research but also a versatile, powerful tool for accelerating the development of data-driven products.

Any questions?
Tell me all about them.

Jente Somers
Manager Innovation Acceleration

Discover

Fishway

Optimizing foodtech research and insights through a tailored AI roadmap

DYMO

Empowering customers with smart and connected label printing

Menu