黑料福利网

Every year, hundreds of thousands of patients are admitted to intensive care units (ICUs). Many of them experience serious problems with blood circulation. In these patients, the heart and blood vessels cannot adequately deliver blood and oxygen throughout the body. This condition is known as shock and is one of the leading causes of death in the ICU.

Timely treatment with fluids, medications, or other interventions can save lives. However, too much treatment can also be harmful. ICU physicians therefore must continuously balance the benefits and risks of treatment while making decisions under significant time pressure and with large amounts of information available to them.

Looking beyond the numbers

Modern ICUs continuously monitor physiological signals such as heart rhythm through an electrocardiogram (ECG) and blood pressure through an arterial line. These measurements generate enormous amounts of data.

In daily practice, physicians mainly focus on single values displayed on monitors, such as average heart rate or blood pressure. As a result, much of the information contained in the underlying waveforms remains unused. Yet these patterns can sometimes provide early signs that a patient's condition is worsening.

Van Mierlo investigated how AI could transform this rich stream of ICU data into information that supports better and more timely clinical decision-making.

AI uncovered hidden information in patient signals

The dissertation consisted of two parts. The first focused on analyzing waveform signals from ECG and blood pressure measurements.

The research showed that these waveforms contain substantially more clinically relevant information than the waveform features traditionally used in patient monitoring. To uncover this information, Van Mierlo used autoencoders, a type of AI that automatically learns patterns from large amounts of data.

This approach enabled complex physiological signals to be summarized into useful features without manual interpretation. It outperformed traditional techniques and could be applied to different types of physiological signals.

An additional advantage of this approach is that the model only needs to be trained once on a large dataset and can then be reused efficiently for new applications.

Local models performed better

The second part of the dissertation focused on models that predict hemodynamic deterioration, meaning a decline in a patient's ability to maintain adequate blood circulation.

Van Mierlo evaluated these models using data from in Eindhoven. One of the most important findings was that a model developed abroad did not automatically perform well in this local clinical setting.

Differences in treatment protocols, measurement methods, and clinical decision-making can affect how well a prediction model performs. As a result, a model trained on local data consistently outperformed the externally developed model, even when less data were available.

These findings showed that AI models cannot simply be transferred from one clinical setting to another. They need to be adapted to local clinical practice.

How a model is trained matters

The research also demonstrated that the way an AI model is trained is at least as important as the model itself.

When a model mainly learned from differences between patients, its performance often appeared better than it actually was. Models that learned from changes within the same patient before treatment provided a more realistic estimate of their clinical value.

This finding highlighted the importance of carefully designing and evaluating medical AI systems before they are used in clinical practice.

Combining waveforms and clinical data

In a final step, Van Mierlo combined the waveform features from the first part of the research with the clinical information used in the second part.

This multimodal approach brought together detailed physiological signals and the data physicians already use when making treatment decisions. At this stage, the waveform features added only limited value beyond the available clinical information.

Nevertheless, the results pointed toward a promising direction for future research. The researchers suggested that these signals may play a larger role when prediction models are linked more closely to a patient's physiological condition rather than to the moment a physician decides to intervene.

Toward AI-supported decision-making in intensive care

The dissertation laid an important foundation for AI-supported decision-making in the ICU. It showed that patient signals contain far more information than is currently being used and that intelligent algorithms can unlock this information in an interpretable and reusable way.

At the same time, the research demonstrated that predictive models depend strongly on how prediction targets are defined and on the data used for training. Models developed in one clinical setting are therefore not automatically suitable for another. Locally trained models may provide better performance because they are better aligned with local clinical practice.

By combining waveform information with the already available clinical data, Van Mierlo took an important step toward locally applicable, multimodal decision-support tools for the early recognition of hemodynamic deterioration. The findings also underscored that successful implementation requires careful validation, local adaptation, and close collaboration between clinicians, engineers, and AI researchers.

AI is unlikely to replace intensive care physicians. Instead, it may become a useful tool that helps them recognize deterioration earlier, interpret complex patient data more effectively, and support clinical decision-making.

For more on Roy van Mierlo's research and PhD journey, read the interview published by ..