Page 161 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 4 Data-driven reduction of cardiac models 133

Table 4.3 Diagnostic statistics for patient-specific dataset (N=125); a positive event, i.e.
functionally significant lesion, is defined by invasive FFR ≤ 0.80. cFFR CFD diagnostic statistics
are based on the CFD algorithm, cFFR ML diagnostic statistics are based on the ML algorithm.

Measure cFFR CFD cFFR ML
True positive 31 31
False positive 14 14
True negative 73 73
False negative 7 7
Sensitivity 81.6% (66.6%–90.8%) 81.6% (66.6%–90.8%)
Specificity 83.9% (74.8%–%90.1) 83.9% (74.8%–%90.1)
PPV 68.9% (54.3%–%80.2) 68.9% (54.3%–%80.2)
NPV 91.2% (83.2%–%95.7) 91.2% (83.2%–%95.7)
Accuracy 83.2% (75.6%–%88.7) 83.2% (75.6%–%88.7)
Correlation 0.725 0.729
Mean ± St. dev. 0.814 ± 0.135 0.815 ± 0.135

for non-invasive FFR computation methods: 73% to 85% [322–324,
326,328,346,347]. All of these past studies have demonstrated an
incremental value of non-invasively determined FFR over quanti-
tative or visual stenosis grading on CCTA data. Thus, such a solu-
tion can further enhance the gatekeeper role played by CCTA for
excluding patients from invasive catheterization. Past approaches
assessing the hemodynamic significance of CAD from anatomi-
cal features alone had limited success [351,352]. The herein pro-
posed approach though relies on a comprehensive set of anatom-
ical characteristics describing not only the lesion, but the entire
coronary tree. Further, relying on a state-of-the-art deep learning
based approach, and using solely anatomical features for predict-
ing patient-specific hemodynamics can lead to diagnostic statis-
tics that are comparable to those obtained by solving complex
hemodynamics equations.

4.1.4.1 Use of synthetic data
Specifically, the key ingredients of the approach are the genera-
tion of a very large training database, and the definition of features
that are more relevant for coronary hemodynamics. Ideally, the
training should be performed based on patient-specific models
reconstructed from CCTA images, thus taking into consideration
the natural variation in the coronary anatomy across a large num-
ber of individuals, and the resulting range of invasive FFR values