The influence of artificial intelligence (AI) on cardiovascular medicine and imaging is growing, and showing increasing promise. That’s the conclusion drawn by researchers from London’s Imperial College and Ohio’s Cleveland Clinic, who have produced a review of AI in personalized cardiovascular medicine and cardiovascular imaging.
In the review, a range of studies into AI’s applications in cardiovascular medicine and imaging were examined and conceptualized. The researchers considered the advantages, limitations, and potential of AI in this sphere. Here we will summarise the review’s major findings.
Why cardiovascular medicine needs AI
The review tracked the increase of cardiovascular disease, which is the world’s leading cause of mortality, and the concurrent shift in disease management from “population-based care to more patient-centered approaches”. With this development, precision medicine has emerged; a healthcare model that can account for individual variability in genes, environment, and lifestyles.
It was explained that the high volumes of heterogeneous data currently being produced in cardiovascular medicine – from electronic health records to omics data, high-resolution medical imaging data, and biosensors – lends itself to “sophisticated analysis” by AI.
Three aspects of cardiovascular medicine were highlighted as particularly suited to the introduction of AI and machine learning (ML):
- Clinicians can benefit from improved image interpretation, diagnosis, and risk prediction, as well as better guidance on the best treatment route at the various disease stages
- Health systems can benefit from increased efficiency, reduced errors, and better patient outcomes
- Patients can benefit from insights into their own data, and increased awareness of primary and secondary cardiovascular health prevention
The application of AI to cardiovascular imaging in cardiovascular medicine
Electrocardiography
Electrocardiography has been earmarked as an area with enormous potential for AI applications. ML studies on an electrocardiogram (ECG) date back to the 1960s, and in recent years, significant studies have included:
- Convolutional neural network (CNN)-based screening for cardiac contractile dysfunction with an AI-enabled ECG
- A CNN to recognize patients with atrial fibrillation while still in sinus rhythm
- Eliciting a sex and age estimate from a 12-lead ECG
Several studies were able to demonstrate algorithmic performance exceeding a cardiologist. A 34-layer CNN trained to detect 12 arrhythmia classification was tested against six cardiologists and yielded an aggregate positive predictive value of 0.80 compared to the cardiologists’ 0.72.
Another study detected 12 arrhythmia classifications from a single-lead ECG using a deep neural network – the application outperformed cardiologists with a 0.91 AUC. Among the limitations of the studies cited by the review were subjectiveness (the neural network for atrial fibrillation was trained on a population with higher incidences of atrial fibrillation), some investigations being confined to single-lead ECGs, and data having been labeled by a cardiologist and technician beforehand, potentially affecting reproducibility.
Computed tomography coronary angiography (CTCA)
Studies suggest that a combination of cardiac computed tomography coronary angiography (CTCA) and ML could help in “improving diagnostic accuracy and predicting prognostic events”.
ML was used to predict 5-year all-cause mortality in over 10,000 patients and compared to the CTCA severity score and Framingham risk score (FRS).
ML performed well, with an AUC of 0.79; higher than CTCA severity scores (0.63) and FRS (0.61). Another study created an ML risk score for coronary artery disease (CAD) using data gathered by the CONFIRM registry, including information from CTCA readings. The ML algorithm achieved an AUC of 0.77; higher than the 0.69–0.70 of other scores.
Among the limitations reported for using ML with CTCA was the manual selection of variables for ML performance and the potential for selection bias.
Myocardial perfusion imaging
ML can help with the interpretation of myocardial perfusion single-photon emission computed tomography (SPECT), a type of nuclear imaging that is used to diagnose CAD.
CNNs have been developed to predict CAD, achieving an AUC of 0.80 compared to 0.78 in total perfusion deficit; a score that is commonly used for SPECT analysis.
Another study used an ML algorithm to combine quantitative perfusion and functional variables from SPECT, in order to improve the detection of CAD. Compared to visual segmentation scoring, the algorithm achieved a diagnostic accuracy of 86% and AUC of 0.92, performing better than TPD (81.0% and 0.90).
Limitations of studies using ML to interpret myocardial perfusion imaging included the size of the dataset, the use of some degree of visual interpretation, and a limited number of feature classifications which impacted the diagnostic accuracy in different patient populations.
Cardiac magnetic resonance imaging
ML has also been tried with cardiac MRI, being applied to functional indices. Deep learning (DL) CNN was utilized for left ventricle volume, end-diastolic volume (EDV), end-systolic volume (ESV), and ejection fraction (EF) estimation.
Accuracies of y=0.91x+11.7 for ESV, y=0.97x+9.5 for EF, and y=0.87+0.2 for EDV were found, revealing promise in the clinical use of ML-powered left ventricle volume prediction.
Using the Automatic Cardiac Diagnosis Challenge (ACDC) , a fully annotated CMR dataset, a study assessed the performance of DL algorithms in carrying out CMR assessment via segmentation and classification of pathologies. DL achieved a mean correlation score of 0.97 for extraction of clinical indices and an accuracy of 0.96 for automatic diagnosis, suggesting effectiveness in clinical practice.
Obstacles to clinical use could be the relatively low datasets used in the studies, and uncertainty in exactly how algorithms make decisions which could prevent diagnostic summaries from being accompanied by clinical reasoning.
Electronic health records
The review concluded that electronic health records offer large patient datasets, but they are complex and potentially disorganized, which can make the creation of a predictive algorithm difficult. Nevertheless, the potential of ML as a means to further understanding the complex variables within EHR datasets was highlighted.
AI in cardiovascular medicine and imaging: The future
There are many challenges which AI must overcome if it is to achieve clinical viability in cardiovascular medicine and imaging. Further studies with larger datasets involving cross-industry collaboration must be conducted, security concerns relating to algorithm development should be addressed, and the cost-effectiveness of the technology needs to be proven in hospitals.
However, it is clear that there is promise in areas such as classification, predicting, improving diagnostic accuracy, reporting, and screening. With the introduction of multi-center randomized controlled trials, the review concluded that “the utility and reproducibility of AI in cardiovascular medicine” can be further validated.
How Circle CVI’s new AI software is paving the way for cardiovascular imaging and medicine
Leading the way in cardiac imaging solutions, Circle CVI’s latest cardiac imaging software, cvi42, is equipped with fully embedded AI to help with reading and reporting for:
- Cardiac CT
- Cardiac MRI
- Quantitative perfusion
- Myocardial strain
- Interventional planning
- 4D flow
- Electrophysiology
calculation algorithms and automated import functions of cvi42 software generate a seamless transition from quantitative data to an organized, coherent report. Straightforward reporting tools and interactive graphics augment your workflow with no client configuration required. Download your cvi42 free 42 trial today to find out more.
Sources:
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8261749/
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8261749/#r1
https://www.ncbi.nlm.nih.gov/pubmed/30617318
https://pubmed.ncbi.nlm.nih.gov/31378392/
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7661045/
https://stanfordmlgroup.github.io/projects/ecg/
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6784839/
https://pubmed.ncbi.nlm.nih.gov/27252451/
https://pubmed.ncbi.nlm.nih.gov/29753765/