Machine Learning through FMTVDM Proprietary QCA Equations

Several investigators have previously raised the issue of what constitutes a significant stenosis and how to determine the normal region of an artery from the abnormal regions given the potential of diffusely diseased vessels. Attempts to visualize the normal region by semi-automated Quantitative Coronary Arteriography (QCA) have been one of several approaches employed to try to reduce the error associated with interpretation of coronary arteriograms.

However, this semi-automated approach is still associated with the need for human intervention, calibration and estimations, including determination of the normal proximal region of the artery in question, the location of vessel walls, and determination of the most stenotic region as well as determination of appropriate biplane images when orthogonal views are used.

Previously, phantom images with known concentric diameter reductions have been used to examine the reliability of QCA systems and have demonstrated QCA accuracy to within 0.1 millimeter. There are currently no phantoms available with eccentric stenoses and curvatures such as those seen in human coronaries, and the expense of reproducing human coronary arteries from autopsies after previous coronary arteriograms without intervening changes provides not only an economic problem but possibly an ethical one.

For this reason, the clinical results from these 1040 coronary arteries, which the primary author has previously analyzed provides the data base against which quantitatively conclusions can be drawn with regard to SFR, % DS results and their associated standard deviations. From this, it was possible to provide data, proprietary equations, and a graphic record of the experience to date and its potential use by clinicians. The theoretical and clinical relationship of SFR to % DS as briefly revealed here shows a curvilinear function where no apparent effect upon stenosis flow reserve until there is at least a 35-40 percent reduction in coronary lumen diameter. This plateaus off at approximately 85-90 % DS. No lesions were quantitatively reported between the range of 91-100 % DS, leading one to suspect that stenoses in this range tend to totally occlude or were not detected on arteriography.

These two variables can also be compared as average values obtained over the stenoses studied, with a strong (R = 0.98) linear correlation. This relationship allowed for the development of proprietary equations, which can be used to calculate stenosis flow reserve from percent diameter stenosis, and percent diameter stenosis from stenosis flow reserve, respectively as shown in figure 2.

The development and use of Figure one allows clinicians to estimate the stenosis flow reserve for a given lesion based upon % DS. However, the use of visual estimates of % DS cannot be improved by the simple use of Figure one and likewise, Figure one is only as useful as the accuracy of the method used to determine % DS. Based upon the results obtained by analyzing this SFR data and comparing it with the FMTVDM data, several points can be made. First, a coronary lumen narrowing alone, which has less than a 40 percent reduction in diameter stenosis due to the development of an underlying coronary artery inflammatory plaque [27-32], cannot be appreciably detected using arteriography or QCA analysis. These plaques can be detected using Intravascular Coronary Ultrasound (IVUS) once the clinician knows where to look for them. Such inflammatory plaques are extremely important clinically as they account for the majority of sudden cardiac death seen in individuals with no or few prior anginal symptoms. These plaques are measureable using FMTVDM to unmask the underlying inflammatory plaques and expose their effect upon CFR as detailed elsewhere [1-2,18-24] and provide the necessary information need by the interventional cardiologist.

Second, lesions with a greater than 40 percent reduction in diameter are associated with linearly decreasing SFR. Third, while lesions with a greater than 85 % DS have the widest variability in SFR, they all are associated with a SFR of < 1.0, revealing impaired coronary flow at rest and a further impaired ability to increase coronary blood flow when needed, thus representing significant and potentially life-threatening CAD.

The key to improved diagnostic and therapeutic CAD intervention is contingent upon accuracy. Accuracy is determined by the ability to measure and not qualitatively look for the presence or absence of disease. The human eye and brain is unable to define CAD at the level where coronary artery plaques are impairing coronary artery function. Furthermore, the reliability of qualitative imaging is associated with sensitivity and specificity errors already well established in the medical and lay literature; independent of whether we are talking about coronary arteriography or Myocardial Perfusion Imaging (MPI).
Quantification of coronary arteriography requires QCA and quantification of MPI requires FMTVDM. The development of proprietary equations (FMTVDM) provides the Artificial Intelligence (AI) language for these machines to learn from each other as shown in figure 5. The acquired information from QCA teaches FMTVDM and FMTVDM teaches QCA. The consequential refinement and ML that occurs will further enhance this quantification of CAD both for diagnostic purposes and for modification of treatment based upon measureable treatment outcomes. Systems, which do not directly measure [19-20,22-23,26] cannot provide true AI.

 

Figure 5: Development of true artificial intelligence through machine learning removing human error from qualitative interpretation of results.
The birth of true AI for diagnosing and treating coronary artery disease.

 

Through the use of the FMTVDM proprietary equations, QCA and FMTVDM can learn from each other without human intervention, marking the evolution to true AI.