SPC365

Non-invasive derivation of iFR from Invasive Coronary Angiography using a new Deep Learning AI model and comparison with Human operators’ performance

Session: Sessão Melhores Comunicações Orais
Speaker: Catarina Oliveira

Congress: CPC 2024
Topic: P. Other
Theme: 37. Miscellanea
Subtheme: 12.3 Coronary Artery Disease – Diagnostic Methods
Session Type: Comunicações Orais
FP Number: ---
Authors: Catarina Simões De Oliveira; Miguel Nobre Menezes; João Lourenço Silva; João Silva Marques; Cláudia Moreira Jorge; Ana Rita Francisco; Beatriz Silva; Marta Vilela; Rita Marante de Oliveira; Tiago Rodrigues; Arlindo L. Oliveira; Fausto J. Pinto

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:&quot;Times New Roman&quot;,serif"><strong><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif"><span style="color:black">Background</span></span></span></strong><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif"><span style="color:black">: Coronary angiography (CAG) derived physiology methods have been developed to simplify and increase the usage of coronary physiology, by removing or reducing its invasive nature. Almost all studies have focused on FFR (only a single, yet unpublished study, focused on iFR), based mostly on dynamic fluid computational algorithms. </span></span></span><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif">We aimed to develop a different approach based solely on artificial intelligence (AI) methods, to fully automatically derive iFR from CAG images alone.</span></span></span></span></p>

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:&quot;Times New Roman&quot;,serif"><strong><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif"><span style="color:black">Methods</span></span></span></strong><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif"><span style="color:black">: Consecutive patients undergoing invasive iFR measurements were included. We developed an AI model capable of classifying target lesions as positive (iFR &le; 0,89) or negative (iFR &gt; 0,89) based solely on standard CAG images. Three Interventional Cardiologists were also asked to classify the target lesions binarily as well. The predictions of both AI and operators were then compared to the true invasive measurements, by calculating </span></span></span><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif">accuracy, negative predictive value (NPV), positive predictive value (PPV), sensitivity and specificity, as well as Area Under the Curve (AUC) by ROC curve analysis.</span></span></span></span></p>

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:&quot;Times New Roman&quot;,serif"><strong><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif"><span style="color:black">Results</span></span></span></strong><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif"><span style="color:black">: 250 measurements, from 223 patients (age 68&plusmn;11 years old, 66.37% male), were included - <strong>Table 1</strong>. iFR was performed predominantly in a chronic setting (66.37%), and in acute coronary syndromes (33.18%) functional assessment was only performed in non-culprit vessels &ge; 48h after the index event. Left descending coronary artery (LAD) was the most evaluated vessel (51.6%), followed by right coronary artery (RCA) and circumflex (Cx) &ndash; <strong>Table 2</strong>. </span></span></span></span></span></p>

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:&quot;Times New Roman&quot;,serif"><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif"><span style="color:black">For AI, the </span></span></span><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif">accuracy, NPV, PPV, sensitivity and specificity were 72%, 90%%, 48%, 77% and 71%, respectively. All three operators had an inferior performance to AI, with an accuracy ranging from 24 to 46%, NPV 47 to 68%, PPV 11 to 19%, sensitivity 26 to 43% and specificity 23 to 50% - <strong>Image 1 and Graphic 1</strong>. </span></span></span></span></p>

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:&quot;Times New Roman&quot;,serif"><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif">Performance differed per target vessel, but the superiority of AI persisted when individual vessels were considered. The best performance of AI was for the RCA (accuracy 86%, NPV 97%), followed by Cx (accuracy 69%, NPV 96%) and LAD (accuracy 66%, NPV 78%). For operators, the best performance was for the LAD (accuracy ranging from 33 to 50%), RCA (accuracy ranging from 17 to 50%) and Cx (accuracy ranging from 9 to 27%).</span></span></span></span></p>

<p><strong><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif">Conclusion</span></span></strong><span style="font-size:9.0pt"><span style="font-family:&quot;Arial&quot;,sans-serif">: We developed an AI model capable of binary iFR estimation from CAG images, with superior performance to human operators&rsquo; predictive capability for all metrics. Despite modest accuracy, the consistently high NPV is of potential clinical significance, as it would enable avoiding further invasive manoeuvres after CAG. This is especially relevant because most iFR measurements in this dataset were negative (&gt;0,89), as in previously published iFR studies.</span></span></p>