SPC365

Radiomics-based artificial intelligence model allows for personalized prediction of ventricular arrhythmias in patients with hypertrophic cardiomyopathy

Session: SESSÃO DE COMUNICAÇÕES ORAIS 20 - PRÉMIO MELHOR COMUNICAÇÃO ORAL
Speaker: Miguel Marques Antunes

Congress: CPC 2025
Topic: F. Valvular, Myocardial, Pericardial, Pulmonary, Congenital Heart Disease
Theme: 17. Myocardial Disease
Subtheme: 17.2 Myocardial Disease – Epidemiology, Prognosis, Outcome
Session Type: Comunicações Orais
FP Number: ---
Authors: Miguel Marques Antunes; Ricardo Carvalheiro; João Santinha; Vera Ferreira; Isabel Cardoso; Boban Thomas; Mário Martins Oliveira; António Fiarresga; Nuno Cardim; Rui Cruz Ferreira; João Bicho Augusto; Sílvia Aguiar Rosa

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:Aptos,sans-serif"><strong><span style="font-size:11.0pt"><span style="color:black">Background</span></span></strong><span style="font-size:11.0pt"><span style="color:black">: Ventricular arrythmias (VA) are potentially life-threatening in hypertrophic cardiomyopathy (HCM) </span></span><span style="font-size:11.0pt">patients (P). Traditional risk models have limited accuracy to predict these events. Artificial intelligence models allow for deep quantitative phenotyping of high-dimensional radiomics data derived from cardiac magnetic resonance (CMR), which may enhance VA risk stratification in HCM.</span></span></span></p>

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:Aptos,sans-serif"><strong><span style="font-size:11.0pt">Aim: </span></strong><span style="font-size:11.0pt">To develop a model able to predict VA events derived from CMR left ventricular (LV) late gadolinium enhancement (LGE) imaging.</span></span></span></p>

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:Aptos,sans-serif"><strong><span style="font-size:11.0pt"><span style="color:black">Methods</span></span></strong><span style="font-size:11.0pt"><span style="color:black">: CMR images from 63 HCM P (median age 54 [43-66] years, 35% female), prospectively followed at a Cardiomyopathy Clinic were analyzed. The LV wall was manually segmented using 3D Slicer 5.2.2. We extracted 1223 features using PyRadiomics (v3.1.0), covering shape, first order and textural features from original and filtered images, which were z-score normalized for intensity discretization. The outcome was a time-to-event analysis of a composite of VA &ndash; ventricular fibrillation (VF), ventricular tachycardia, (VT), and non-sustained VT (NSVT) - with T0 being the day of the CMR. Sixty-three P were randomly split in a 75%:25% ratio into a 47P (training) and 16P (held-out testing) sets. A</span></span><span style="font-size:11.0pt"><span style="color:#222222"> Random Survival Forest (RSF) was optimized using a 5-fold cross-validation</span></span><span style="font-size:11.0pt"> and performance was assessed with the concordance index (c-index).</span></span></span></p>

<p style="text-align:justify"><span style="font-size:12pt"><span style="font-family:Aptos,sans-serif"><strong><span style="font-size:11.0pt"><span style="color:black">Results</span></span></strong><span style="font-size:11.0pt"><span style="color:black">: The studied cohort had a median 12% [6 &ndash; 18] of LV mass LGE. The primary outcome occurred in 14 (22%) P &ndash; 1 VF, 2 VT and 11 NSVT &ndash; over a median follow-up of 3.6 [1.5-4.2] years. </span></span><span style="font-size:11.0pt"><span style="color:#222222">The RSF with 30 estimators (2 samples/split;10 samples/leaf) yielded a c-index of 0.872&nbsp;+/- 0.146.</span></span> <span style="font-size:11.0pt"><span style="color:#222222">A c-index of 0.761 was achieved when tested in P not used for training (held-out set). </span></span><span style="font-size:11.0pt"><span style="color:black">Permutation importance analysis identified the features that were key to the model (Figure 1). Patients who suffered events had a higher presence&nbsp;of heterogeneity-related features in their myocardium (Figure 2). Finally, the model allowed for the creation of unique risk curves for each individual P, with&nbsp;which a physician can capture a personalized evolution of the arrhythmic risk of each P throughout time (Figure 3). </span></span><br />
<strong><span style="font-size:11.0pt"><span style="color:black">Conclusion</span></span></strong><span style="font-size:11.0pt"><span style="color:black">: For the first time, a LV LGE radiomic-based model performed a time-to-event analysis of </span></span><span style="font-size:11.0pt">VA</span><span style="font-size:11.0pt"><span style="color:black">. This approach </span></span><span style="font-size:11.0pt">showed strong internal and external validation performance, enabling the development of individualized risk profiles. &nbsp;Further model refinement and validation could allow clinicians to predict individual arrhythmic risk since the day a CMR is performed.</span></span></span></p>