A proof of concept study for machine learning application to stenosis detection

A data-driven computational methodology towards a pre-hospital Acute Ischaemic Stroke screening tool using haemodynamics waveforms

BACKGROUND AND OBJECTIVE

Acute Ischaemic Stroke (AIS), a significant global health concern, results from occlusions in cerebral arteries, causing irreversible brain damage. Different type of treatments exist depending on the size and location of the occlusion. Challenges persist in achieving faster diagnosis and treatment, which needs to happen in the first hours after the onset of symptoms to maximize the chances of patient recovery. The current diagnostic pipeline, i.e. "drip and ship", involves diagnostic via advanced imaging tools, only available in large clinical facilities, which poses important delays. This study investigates the feasibility of developing a machine learning model to diagnose and locate occluding blood clots from velocity waveforms, which can be easily be obtained with portable devices such as Doppler Ultrasound. The goal is to explore this approach as a cost-effective and time-efficient alternative to advanced imaging techniques typically available only in large hospitals.

METHODS

Simulated haemodynamic data is used to conduct blood flow simulations representing healthy and different AIS scenarios using a population-based database. A Machine Learning classification model is trained to solve the inverse problem, this is, detect and locate a potentially occluding thrombus from measured waveforms. The classification process involves two steps. First, the region where the thrombus is located is classified into nine groups, including healthy, left or right large vessel occlusion, left or right anterior cerebral artery, and left or right posterior cerebral artery. In a second step, the bifurcation generation of the thrombus location is classified as small, medium, or large vessel occlusion.

RESULTS

The proposed methodology is evaluated for data without noise, achieving a true prediction rate exceeding 95% for both classification steps mentioned above. The inclusion of up to 20% noise reduces the true prediction rate to 80% for region detection and 70% for bifurcation generation detection.

CONCLUSIONS

This study demonstrates the potential effectiveness and efficiency of using haemodynamic data and machine learning to detect and locate occluding thrombi in AIS patients. Although the geometric and topological data used in this study are idealized, the results suggest that this approach could be applicable in real-world situations with appropriate adjustments. Source code is available in https://github.com/ahmetsenemse/Acute-Ischaemic-Stroke-screening-tool-.

Peripheral artery disease diagnosis based on deep learning-enabled analysis of non-invasive arterial pulse waveforms

This paper intends to investigate the feasibility of peripheral artery disease (PAD) diagnosis based on the analysis of non-invasive arterial pulse waveforms. We generated realistic synthetic arterial blood pressure (BP) and pulse volume recording (PVR) waveform signals pertaining to PAD present at the abdominal aorta with a wide range of severity levels using a mathematical model that simulates arterial blood circulation and arterial BP-PVR relationships. We developed a deep learning (DL)-enabled algorithm that can diagnose PAD by analyzing brachial and tibial PVR waveforms, and evaluated its efficacy in comparison with the same DL-enabled algorithm based on brachial and tibial arterial BP waveforms as well as the ankle-brachial index (ABI). The results suggested that it is possible to detect PAD based on DL-enabled PVR waveform analysis with adequate accuracy, and its detection efficacy is close to when arterial BP is used (positive and negative predictive values at 40 % abdominal aorta occlusion: 0.78 vs 0.89 and 0.85 vs 0.94; area under the ROC curve (AUC): 0.90 vs 0.97). On the other hand, its efficacy in estimating PAD severity level is not as good as when arterial BP is used (r value: 0.77 vs 0.93; Bland-Altman limits of agreement: -32%-+32 % vs -20%-+19 %). In addition, DL-enabled PVR waveform analysis significantly outperformed ABI in both detection and severity estimation. In sum, the findings from this paper suggest the potential of DL-enabled non-invasive arterial pulse waveform analysis as an affordable and non-invasive means for PAD diagnosis.

Classification and regression of stenosis using an in-vitro pulse wave data set: Dependence on heart rate, waveform and location

BACKGROUND

Data-based approaches promise to use the information in cardiovascular signals to diagnose cardiovascular diseases. Considerable effort has been undertaken in the field of pulse-wave analysis to harness this information. However, the inverse problem, inferring arterial properties from waveform measurements, is not well understood today. Consequently, uncertainties within the estimation hinder the diagnostic application of such methods.

METHOD

This work contributes a publicly available data set measured at an in-vitro cardiovascular simulator, focusing on a set of input conditions (heart rate, waveform) and stenosis locations. Furthermore, a first attempt is undertaken to perform classification and regression on this data set using standard machine learning methods on features extracted from four peripheral pressure signals.

RESULTS

The locations of six different stenoses could be distinguished at high accuracy of 93%, where transfer function-based features outperformed features based solely on signal shape in almost all cases. Furthermore, regression on the stenosis position could be performed with a root mean square error of 2.4 cm along a 20 cm section of the arterial system using a shallow neural network. However, the performance difference between shape and transfer function features was not clear for this task.

CONCLUSION

The data set contains 800 measurements and allows investigating the influence of different heart boundary conditions, such as heart rate and waveform shape, on classification and regression tasks. Extracting features that minimise this influence is a promising way of improving the performance of these tasks.

CNN-KCL: Automatic myocarditis diagnosis using convolutional neural network combined with k-means clustering

Myocarditis is the form of an inflammation of the middle layer of the heart wall which is caused by a viral infection and can affect the heart muscle and its electrical system. It has remained one of the most challenging diagnoses in cardiology. Myocardial is the prime cause of unexpected death in approximately 20% of adults less than 40 years of age. Cardiac MRI (CMR) has been considered a noninvasive and golden standard diagnostic tool for suspected myocarditis and plays an indispensable role in diagnosing various cardiac diseases. However, the performance of CMR depends heavily on the clinical presentation and features such as chest pain, arrhythmia, and heart failure. Besides, other imaging factors like artifacts, technical errors, pulse sequence, acquisition parameters, contrast agent dose, and more importantly qualitatively visual interpretation can affect the result of the diagnosis. This paper introduces a new deep learning-based model called Convolutional Neural Network-Clustering (CNN-KCL) to diagnose Myocarditis. In this study, we used 47 subjects with a total number of 98,898 images to diagnose myocarditis disease. Our results demonstrate that the proposed method achieves an accuracy of 97.41% based on 10 fold-cross validation technique with 4 clusters for diagnosis of Myocarditis. To the best of our knowledge, this research is the first to use deep learning algorithms for the diagnosis of myocarditis.

Machine learning for detection of stenoses and aneurysms: application in a physiologically realistic virtual patient database

This study presents an application of machine learning (ML) methods for detecting the presence of stenoses and aneurysms in the human arterial system. Four major forms of arterial disease-carotid artery stenosis (CAS), subclavian artery stenosis (SAS), peripheral arterial disease (PAD), and abdominal aortic aneurysms (AAA)-are considered. The ML methods are trained and tested on a physiologically realistic virtual patient database (VPD) containing 28,868 healthy subjects, adapted from the authors previous work and augmented to include disease. It is found that the tree-based methods of Random Forest and Gradient Boosting outperform other approaches. The performance of ML methods is quantified through the [Formula: see text] score and computation of sensitivities and specificities. When using six haemodynamic measurements (pressure in the common carotid, brachial, and radial arteries; and flow-rate in the common carotid, brachial, and femoral arteries), it is found that maximum [Formula: see text] scores larger than 0.9 are achieved for CAS and PAD, larger than 0.85 for SAS, and larger than 0.98 for both low- and high-severity AAAs. Corresponding sensitivities and specificities are larger than 90% for CAS and PAD, larger than 85% for SAS, and larger than 98% for both low- and high-severity AAAs. When reducing the number of measurements, performance is degraded by less than 5% when three measurements are used, and less than 10% when only two measurements are used for classification. For AAA, it is shown that [Formula: see text] scores larger than 0.85 and corresponding sensitivities and specificities larger than 85% are achievable when using only a single measurement. The results are encouraging to pursue AAA monitoring and screening through wearable devices which can reliably measure pressure or flow-rates.