Cardiovascular diseases (CVDs) are the leading cause of death globally, with approximately 17.9 million people dying each year from CVDs, according to the World Health Organization.
The rapid development of technology has generated vast amounts of patient data, ranging from clinical information such as genome sequencing data to individual-level biometric data such as heart rates measured by smart bracelets.
In addition, vital bioimaging measurements are essential for treatment selection by doctors. Patients are increasingly looking for faster, more personalized care, driving demand for automated solutions in healthcare.
Deep learning, known for its ability to handle large and complex sets of data, has gained popularity in clinical diagnosis due to advances in computational power, especially using graphics processing units (GPUs).
Deep learning has achieved success in a variety of fields, including image classification, image segmentation, natural language processing, voice recognition, and genomics, all of which hold great potential to improve the diagnosis, prediction, and prevention of CVDs.
As of June 2019, only seven AI-based algorithms in cardiology have received FDA approval. These algorithms mainly focus on analyzing medical imaging data, such as echocardiography, MRI, and computer imaging. However, many of these algorithms focus on unimodal data.
With the increasing availability of multimodal health data from biobanks, it is becoming possible to predict CVDs using integrated multimodal data, including demographic data, multi-omics biodata, and vital symptom markers. This integrated approach can enhance the accuracy and effectiveness of diagnosis and management of CVDs.
How to integrate various image modalities?
The initial use of deep learning in medicine and biosciences focused on the detection and segmentation of medical images. Deep learning algorithms have been shown to outperform established doctors in some tasks. However, most current applications of machine learning in medical imaging focus on one method at a time, with limited integration of information from diverse sources.
There are relatively few studies that use machine learning algorithms to integrate data from different modalities in the context of cardiovascular diseases (CVDs).
One such study, conducted by Nakanishi and colleagues, demonstrated the integration of temporal CT data with coronary atherosclerosis score (CAC) estimates, CAC volume estimates, extracardiac coronary atherosclerosis score, and cardiac fat volume to predict coronary heart disease.
The study found that machine learning algorithms that integrated all of this data to predict coronary heart disease events outperformed those using risk indicators alone, with an average area under the curve (AUC) of 0.765.
This research highlights the possibility of predicting cardiovascular disease more accurately by analyzing clinical data from diverse sources using machine learning algorithms.
Furthermore, integration of clinical data with omics data (e.g., genomics, proteomics) and demographic information can provide valuable insights into the genetic inheritance and environmental factors related to cardiovascular disease.
This could be a comprehensive approach to understand the complex factors contributing to cardiovascular disease and improve prediction and prevention strategies.
How to better the diagnosis by linking images with electrocardiograms?
Electrocardiography (ECG) is a valuable diagnostic tool that measures electrical activity passing through the heart. It is widely used to detect abnormal rhythm disturbances, which can indicate potential heart problems.
Automated interpretation of ECG recordings has been in use since the 1960s with the introduction of digital ECG machines, and it has played a major role in simplifying care in hospitals and reducing costs.
However, many current methods of ECG analysis and interpretation do not fully exploit data from other clinics that could enhance their accuracy and predictive power. One promising approach is the integration of ECG data with photopulse profiling (PPG) data, which can provide additional information about cardiac health.
Shashikumar and colleagues conducted research in which they used convolutional deep learning networks (CNNs) to analyze PPG data recorded using a multi-channel wrist-worn device and single-channel ECG data (used to verify rhythm) for the purpose of predicting atrial fibrillation (AF), a common disorder in heart beat.
Their innovative integration approach showed robust and accurate algorithm development for AF detection using PPG data. This approach has potential to be expanded and may continue to improve in accuracy as larger datasets become available.
Integrating multiple types of physiological data, such as ECG and PPG, can provide a more comprehensive view of a patient’s heart health and can improve the detection and management of heart-related conditions such as atrial fibrillation.