This session provides two examples of how AI and machine learning can enhance human health through better diagnostics, prognoses, and treatment. In the first case example, we will look into different methods that enable Electronic Health Records to provide tailored prognosis and treatment recommendations to individual patients, when we apply deep learning models and include domain knowledge from medical guidelines and ontologies. In the second case example, we will focus on how AI and machine learning have redefined analysis and interpretation of ECGs with promising outlook for e.g. low income countries where skilled cardiologists are not readily available.
 

ABSTRACTS

Augmenting Patient Records for Machine-Learning by Tomer Sagi, Aalborg University
The application of deep learning technologies to Electronic Health Record (EHR) data has the potential to expand the precision medicine revolution by providing tailored prognosis and treatment recommendation to individual patients. However, previous work based on HER data does not utilize the wealth of domain knowledge encoded in medical guidelines and ontologies. In this presentation, we will present several methods to incorporate medical domain knowledge into patient records.

Our first exploration entails the use of hierarchical medical taxonomies in tailoring loss functions, boosting diagnostic prediction accuracy based on prescription history. In our second example, we present our work in examining the applicability of our model to Danish EHR data, investigating its transferability and generalizability.

During the presentation we will discuss some of the challenges and inherent complexities arising from variations in data quality and collection methodologies across different countries. We then present a graph-based patient representation which can be used by graph-convolution-based methods in deep learning. The graph-based representation can subsequently be enhanced by adding medical knowledge from guidelines. We show how this can be done using large language models in the absence of structured data.

We then show how medical taxonomies can be used to modify the node embeddings in the graph, considerably improving the performance of graph convolution neural networks.

AI-Powered Cardiology: Revolutionizing Healthcare, One Beat at a Time by Sadasivan Puthusserypady, DTU
Artificial Intelligence (AI) has redefined the analysis and interpretation of electrocardiograms (ECG), which is a fundamental diagnostic tool in cardiology. ECGs, which record the heart's electrical activity, have traditionally demanded skilled cardiologists for accurate interpretation, a resource often unavailable globally, particularly in lower-income nations.

AI, leveraging deep and machine learning, has shown promise in detecting intricate patterns and irregularities in ECG data, including arrhythmias and ischemic events, often outperforming humans in accuracy and speed. These advancements not only facilitate swift analyses in pressing clinical settings but also enhance diagnostic accuracy, curtailing oversights and fostering improved patient care in cardiovascular health. Furthermore, AI extends beyond diagnostics to the prognosis of heart diseases, thereby pivoting healthcare from a reactive stance to a preventive approach.

An exciting frontier that AI opens up is that of personalized medicine, where vast amount of ECG, genomics and behavioral data can be analyzed to formulate highly personalized treatment plans that leads to better outcomes and fewer side effects. The integration of AI and ECG signals a transformative, more preventative direction in global healthcare, promising enhanced precision and timely interventions.