Automatic etiological diagnosis of cardiac diseases

Motivations

At present, many medical studies have been carried out to investigate the capacity of machine learning methods to detect cardiac pathologies from patient data with standard echocardiographic measurements or with the content of ultrasound images. Since echocardiography is the imaging of choice for establishing a first cardiac diagnosis, this modality has naturally been the subject of important research.

Most of the underlying studies are focused on situations where the expert cardiologist is already able to establish a diagnosis from the data at its disposal, restricting the interest of using such methods to two main scenarios: i) the reduction of the analysis time by automatic computation of a first diagnostic hypothesis to be confirmed by an expert; ii) the deployment of automatic tools for the screening of a large population.

Ambition and novelty

In this project, we intend to move a step forward by studying the possibility of using patient data complemented by high-quality information extracted from echocardiographic acquisitions to make etiological diagnoses. Our solution will be developed on the basis of a unique dataset that will be implemented in the particular context of this study. It will include a total of 1500 patients from a multi-center (HCL and CHU of Caen) and multi-vendor (GE and Philips systems) study for which additional examinations were carried out in order to establish an etiological diagnosis.

 

Targeted pathologies

We will focus on two clinical situations that commonly occur and contribute to the overcrowding of hospitals and an increase in the cost of patient care.

Hypokinetic cardiomyopathy

The first concerns the diagnosis of a hypokinetic cardiomyopathy for which the echocardiographic data do not easily allow to distinguish a cause related to a coronary artery disease or related to a primary myocardial dysfunction, imposing the realization of an invasive intervention, i.e. a coronary angiogram.

Left ventricular hypertrophy

The second concerns the diagnosis of a left ventricular hypertrophy, the etiology of which can be diverse and whose assessment is particularly exhaustive and costly. In this project, we will study the two most common causes: arterial hypertension and infiltrative myocardial disease. The cohort will be composed in a balanced way of the four pathologies mentioned above with data from healthy subjects.

 

AI method with high level of interpretability

Motivations

Among the solutions studied in the literature to predict cardiac pathologies, machine learning methods produce the most advanced results. These techniques are used either to classify targeted pathologies or to identify and analyse phenotype-based groups (also named phenogroups). Deep learning methods are the current solutions of choice to classify cardiac diseases from echocardiography.

Although these methods can achieve high performance for some pathologies, the underlying decision mechanisms does not allow a thorough analysis and interpretation of the results, making it difficult to deploy such solutions in clinical routine.

Ambition and novelty

We will exploit the AI formalism based on the transformer (attention-based neural network) paradigm to perform etiological diagnosis of cardiac diseases. Transformers are a natural and convenient model for combining multimodal inputs. We will therefore format the clinical data in several distinct modalities to effectively exploit the underlying formalism. We will then develop a transformer architecture to extract representations for each modality based on self-attention, but also (and more importantly) to exploit specific cross-attention mechanisms to efficiently fuse the multimodal data of the project. We will also use the transformer paradigm to develop solutions to explain model decision making by selecting the most salient interactions between the key elements for each modality.