Kriba’s (former Newborn Solutions) senior data scientist Beatrice Jobst presented early results of ongoing R&D at the 2nd Taulí Health Artificial Intelligence Symposium (THAIS), held at the Parc Taulí Hospital Universitari in Sabadell, (Barcelona) June 19th and 20th 2023.
The focus of the presentation was on the potential of non-invasively detecting an elevated number of white blood cells in the peritoneal fluid during peritoneal dialysis with Kribas’ first-in-class non-invasive white blood cell counting device, Neosonics®. The device offers an innovative application of ultrasounds and AI for the detection of serous fluid infections through white blood cell counting, by using deep learning models to classify images and provide accurate results.
Results of a deep learning model trained on in-vitro data of peritoneal dialysis bags presents exceptional accuracy in image classification, with an impressive 99.14% success rate. The primary objective of the study was to determine the device’s ability to classify images into infection and non-infection classes. The results are highly promising, with the model demonstrating a sensitivity of 100% and a specificity of 98.80%. This signifies the device’s ability to effectively distinguish between infected and non-infected states, showcasing its potential in detecting peritonitis during home-based peritoneal dialysis.
We also presented a complementary proof of concept research that explores the use of digitally generated images in conjunction with the deep learning model. By training a Generative Adversarial Network (GAN) model for data augmentation and subsequently integrating these digital images into the classification model, we aim to accelerate the device’s validation while maintaining accuracy and reliability. By generating images that closely resemble real data, we will optimize clinical study resources and potentially expedite the product’s approval process. Moreover, this approach holds promise for other applications of the Neosonics device that are being simultaneously developed, such as screening and monitoring meningitis in newborns and infants with an open fontanelle, and uveitis in adult patients.
A multi-centric clinical study now involving Bellvitge University Hospital in Barcelona, and Navarra University Hospital in Pamplona by October 2023, is currently underway to evaluate the device’s ability to detect peritonitis in home-based peritoneal dialysis with high accuracy.
In her presentation titled “Can Artificial Intelligence Surpass the Clinical Eye Today?” at the symposium, Paula Petrone – associate research professor at ISGlobal and data science advisor to the Newborn Solutions team-, highlighted our focus on ultrasound images of cerebrospinal fluid in newborn fontanelle, where AI algorithms are employed to classify and interpret intricate patterns. She defended that by leveraging deep learning algorithms, AI can effectively classify and interpret subtle patterns in these images that may go unnoticed by the human eye alone.
In addition, Paula emphasized the need for explainable machine learning, a concept also known as AI explainability or XAI. Trustworthy AI algorithms should not only make predictions but also provide explanations for those predictions, shedding light on the underlying factors and potential biases present in the data. XAI enables us to interpret and understand the reasoning behind AI’s predictions, promoting transparency and accountability in healthcare. A manuscript, cowritten by Kriba data scientists, Beatrice Jobst and Francesc Carandell, and ISGlobal Data Science team, is in preparation. The paper will detail thoroughly on the above mentioned XAI we use in the deep learning models we create and train for our application for infant meningitis. Paula concluded her presentation by invoking the famous de Saint-Exupéry‘s Little Prince quote “What is essential is invisible to the eye,” suggesting that AI has the potential to enhance our vision and provide us with invaluable insights in medical diagnostics.