The mission of Artificial Intelligence in Medicine Laboratory (AIMLab.) is to develop artificial intelligence algorithms for basic medical & clinical research which will lead to significantly improved patient care.
The billions of mobile devices in use worldwide, along with numerous affordable medical sensors, have simplified the recording and transmission of medical data. However, harnessing this data to improve patient care remains a challenge. Existing algorithms often overlook individual variability and fail to extract actionable clinical insights from vast databases of physiological time-series and medical images. To address this, AIMLab. focuses on developing advanced pattern-recognition algorithms that can unlock the valuable information within such datasets. Our goal is to create intelligent patient monitoring systems that enhance healthcare delivery.
Early cardiovascular disease diagnosis and risk prediction from the raw electrocardiogram.
Current diagnostic methods for cardiac arrhythmias, which mainly rely on clinician interpretation of ECGs, are challenged by variability, time consumption, and the potential for human error. Deep learning, trained on extensive datasets of ECG recordings, can help overcome these limitations. These models are capable of detecting subtle ECG patterns that are often overlooked but indicative of cardiac arrhythmias, leading to earlier and more accurate diagnoses. This is especially critical considering that the early detection of certain arrhythmias, such as atrial fibrillation, can significantly reduce the risk of serious complications like stroke or heart failure.
The image was created by Brimer S. using BioRender.com
Looking through the eyes to diagnose ophthalmic and cardiovascular diseases.
The development of deep learning models for retinal image analysis marks a groundbreaking stride in research, carrying significant implications for diagnosing ophthalmological and cardiovascular diseases. The retina offers a unique glimpse into the body's vascular health, where changes in retinal vessels can signal systemic diseases such as hypertension, diabetes, and cardiovascular disorders. These deep learning models may enable capability to identify subtle abnormalities and patterns that are indicative of disease, even before clinical symptoms become apparent. Such early detection is crucial for timely intervention, which could potentially alter the progression of the disease.
Automatic blood arterioles and venules segmentation from digital retinal fundus images.
Using sleep as a test bed for diagnosing and monitor diseases using digital health technology.
Sleep breathing disorders are highly relevant and can lead to serious health issues if untreated. Traditional diagnostic methods, such as polysomnography, are effective but limited by their accessibility, cost, and complexity. Deep learning provides an innovative solution by efficiently analyzing large and complex physiological time series data from sleep studies, including respiratory patterns, heart rate, and oxygen levels. These models can identify subtle physiological changes indicative of sleep breathing disorders, enabling earlier and more accurate diagnoses, as well as the potential for remote health systems with fewer physiological channels.
Generalization performance in medical artificial intelligence.
Deep learning has emerged as a powerful technique for achieving state-of-the-art results in various domains and applications. However, one of the key assumptions underlying traditional supervised learning approaches is that training and testing data are drawn from the same distribution. In real-world scenarios, this assumption may not hold true. There can be a significant distribution shift between the source domain, on which the data has been trained, and the target domains where the model is expected to generalize. Even a slight departure from a network’s training domain can cause it to make spurious predictions and significantly hurt its performance. As part of this new research pillar we develop and/or evaluate novel algorithms for deep unsupervised domain adaptation or domain generalization and demonstrate their validity on a variety of tasks harnessing physiological time series and medical images.