Smart infection diagnostics: UTI detection with AI

Traditional detection for Urinary tract infections (UTIs), involving manual microscope examination of urine samples, is slow and time consuming. To improve efficiency, urine screening machines analyze various parameters like bacteria and white blood cells. However, interpreting this data is complex.

Researchers at Fraunhofer Austria have developed an AI-based decision tree to address this issue. It analyzes multiple urine sample parameters, improving UTI detection by identifying both established and new diagnostic indicators, enhancing accuracy and efficiency. In this interview with MEDICA-tradefair.com, Giacomo Da Col explains how it works and how it can be used.

Mr. da Col, can you tell us about the traditional methods used for UTI detection and their limitations?

Giacomo da Col: Urinary tract infections (UTIs) are frequently tested in medical labs, with one of our partner hospitals processing around 15,000 urine samples yearly for UTIs and other diseases. Traditionally, a small sample is examined under a microscope over one to two days to detect bacterial growth, but this manual method is inefficient, as two-thirds of samples are typically negative.

To enhance efficiency, urine screening machines now pre-filter samples. These machines analyze 20 to 40 parameters, including bacteria and white blood cell counts, which are key infection indicators, along with other elements like presence of sperm and mucus. While these machines can detect a wide range of factors, the challenge for doctors is to effectively analyze and integrate this extensive data into regular medical practice, as they usually focus on a select few indicators.

How does the new AI method improve upon these traditional approaches?

da Col: Our method uses an automated system with a decision tree algorithm to analyze a wide range of parameters in urine samples, identifying those most relevant to diseases. This provides doctors with clear explanations for diagnoses and enables more informed decisions. The decision tree, simple and easy to understand, ranks parameters by importance. It is user-friendly for both professionals and non-trained individuals.

Our approach not only offers comprehensive analysis beyond just bacteria count and white blood cells, but also aligns with significant parameters identified in medical literature.

What are the key advantages of this AI-based method?

da Col: Our method, which goes beyond the traditional focus on parameters like bacterial levels and white blood cells as identified in medical literature, shows a significant improvement in efficiency. Specifically, we achieved an additional 16% improvement in sample reduction, building on the initial 45% effectiveness of the previous rule. This means an estimated financial impact of €40,000 per year, given a yearly volume of 15,000 samples.

Crucially, we maintained a high sensitivity level of about 95%. In medical testing, a 95% sensitivity level is a standard benchmark, indicating that the test is expected to have only one error in every 20 samples. Any test with more than this error rate is typically considered not sensitive enough. Achieving or exceeding this sensitivity level is vital, as it ensures the reliability and safety of the medical test.

What challenges did you face during the development of this AI method, and how were they addressed?

da Col: The decision tree algorithm, known for its simplicity and interpretability, has been around for a long time. Our goal was to maintain these key qualities of the decision tree, ensuring that anyone can easily read and understand it without additional training. However, we also aimed to address some of its limitations. Despite its simplicity, the decision tree is not widely used in all applications due to certain drawbacks, prompting ongoing research for improvement.

One significant challenge with decision trees is their sensitivity to outliers and slight changes in data distribution. To enhance our method, we focused on increasing the accuracy of the decision tree while preserving its high interpretability. Our objective was to design an algorithm that strikes a balance: improving precision in handling data variability and outliers, yet still being straightforward and clear for users to interpret.

da Col: Our next steps involve two primary goals. First, we aim to develop this system into a service. During the study, we utilized common tools from the scientific community, such as standard Python libraries for machine learning, and added custom code where necessary to enhance the algorithm. However, our goal now is to create a more sophisticated system that allows operators not just to view the decision tree once it is trained, but also to interact with it.

Secondly, we aim to validate our system, which was initially tested with a single hospital, by partnering with other hospitals for external validation. This is crucial for verifying the system's applicability and reliability in different medical settings. While we have not had the chance to collaborate with additional hospitals yet, such partnerships would be invaluable. They would allow us to tailor the system to each hospital's specific data and test the decision tree's generalizability, creating a mutually beneficial situation that enhances the system's utility in the medical field.

How do you envision the role of AI evolving in medical diagnostics, particularly in relation to administrative burdens and diagnostic processes?

da Col: In the past year, there have been remarkable advancements in AI, both in terms of technological capabilities and its integration into daily life. One notable example is ChatGPT, which has experienced unprecedented growth. It has become widely recognized, even among those who have not used it, as a system capable of near-human level conversations. This achievement is akin to passing the Turing test, a milestone that seemed incredible just a year ago, yet it has occurred almost unnoticed by the public.

Looking ahead, predicting the future of AI, especially in the short term, is challenging because we are at the beginning of a steep exponential curve in AI applications. In the field of medicine, there is naturally a lot of caution. Our algorithm's design emphasizes interpretability, a key aspect that may not be the primary focus in many other AI approaches. As AI continues to evolve rapidly, its impact and applications in various fields, including medicine, will likely expand, but it is crucial to balance technological advancement with considerations like interpretability and ethical use.