AI random survival forest predicted major arrhythmic events in dilated cardiomyopathy with AUC up to 0.80 at 28 months, outperforming LVEF-based models (AUC 0.54-0.70).
Does an artificial intelligence model improve the prediction of major adverse arrhythmic cardiac events in patients with non-ischemic dilated cardiomyopathy compared to guideline-based LVEF and LGE models?
An AI-based random survival forest model incorporating clinical, echocardiographic, and CMR data outperforms traditional LVEF and LGE-based models in predicting major arrhythmic events in patients with non-ischemic dilated cardiomyopathy.
Absolute Event Rate: 0% vs 0%
Abstract Background Sudden cardiac death is a significant cause of mortality in patients with dilated cardiomyopathy (DCM). Existing risk stratification guidelines primarily rely on left ventricular ejection fraction (LVEF), which has limited predictive accuracy for arrhythmic events. This study evaluates the performance of an artificial intelligence model in predicting major adverse arrhythmic cardiac events (MAACE) in patients with DCM and compares it to guideline-based Cox regression models. Methods The study analysed data from the DERIVATE-NICM international registry, a multicenter prospective cohort of patients with non-ischemic DCM. The dataset was randomly split into training (75%) and test (25%) cohorts. Preprocessing involved removing low-variance variables, handling missing data, and standardizing inputs. Prediction models included four Cox regression models based on LVEF and late gadolinium enhancement (LGE) variables and a random survival forest (RSF) model incorporating clinical, echocardiographic, and CMR data. Model performance was assessed using area under the receiver operating characteristic curve (AUC) at multiple time points, with validation in the test cohort. Results The final cohort consisted of 1346 patients. After a median follow-up of 32 months (32-18), 84 patients died. Overall, MAACE occurred in 74 (5.5%) patients, with an incidence rate of 2.81 events per 1000person/years. The LVEF≤35% model presented the lowest performance, with a mean AUC ranging between 0.54 and 0.57 from 8 to 32 months. The addition of the LGE variables progressively improved performance: the LVEF ≤35% + LGE presence model achieving a 28-month AUC of 0.63 (0.73–0.52), while the LVEF ≤35% + LGE extension and LVEF ≤35% + LGE extension + LGE pattern models reached AUCs of 0.66 (0.77–0.55) and 0.70 (0.81–0.59), respectively. The RF showed the highest performance, with a 0.78(0.87-0.69) and 0.80(0.87-0.71) at 20 and 28 months, respectively. Conclusion This study underscores the superiority of an AI model in predicting MAACE in DCM compared to guideline-based decision making. The findings highlight the limitations of LVEF as a standalone predictor and support the integration of AI models to improve risk stratification. Future studies should validate these models in external cohorts and explore their integration with advanced deep learning frameworks to enhance predictive capabilities further.Methods Results
Coriano' et al. (Sat,) reported a other. AI random survival forest predicted major arrhythmic events in dilated cardiomyopathy with AUC up to 0.80 at 28 months, outperforming LVEF-based models (AUC 0.54-0.70).