Personalized Risk Prediction of Sudden Cardiac Death

Background / State of the art Electrocardiography, imaging biomarkers (LVEF measured by 3D echocardiography, LV global longitudinal strain, and mechanical dispersion by speckle-tracking 2D echocardiography, the extent of myocardial fibrosis at CMR, reduced 123ImIBG uptake (or accelerated 123I-mIBG washout rate from the heart), and genetic testing, such as selected DNA variants as Desmoplakin, Lamin A/C, PLN and FLMNC) have been associated with cardiac arrest events in pts with LV dysfunction.

Although circulating biomarkers of myocardial stress and fibrosis have been reported to predict prognosis, these biomarkers generally reflect the severity of cardiac dysfunction rather than the specific risk of SCD. Accordingly, they may be used to identify pts who are unlikely to benefit from ICD therapy because of the high risk of death resulting from the progression of HF. However, the predictive power of all these biomarkers of arrhythmic risk has been tested individually in different studies. Because of the complexity of the substrates that underlie SCD, it is improbable that any single marker/ test will achieve significantly better predictive accuracy than LVEF. To overcome this limitation, a combination of markers could be used to identify potential mechanisms associated with an increased risk of SCD in individual patients independent of LVEF.

Description and distribution of activities of each operating unit Prof Badano, in Milan, will be responsible for the whole research project, coordinating the research team and collaborating with the partners. IRCCS Istituto Auxologico in Milan has a dedicated research unit with data managers, statisticians, dedicated research nurses, and research technicians who will collaborate to run the RESPECT project. The partners in Operative Units 1,2,3 and 4 will enroll the patients (WP1) and follow them for a minimum of 12 months (WP2). The central database will be developed in UO 3 using REDCap (Research Electronic Data Capture powered by Vanderbilt, WP3). The 4 UOs will set up an independent clinical event committee to review and classify the events reported during follow-up. Specific core-labs will be organized to centralize the quali-quantitative analyses scheduled within the project: electrocardiography (UO3, WP4), quantitative analysis and radiomics of echocardiography data (UO1, WP5) genetic testing (UO1, WP6), quantitative analysis and radiomics of cardiac magnetic resonance (UO2, WP7); quantitative analysis of nuclear imaging data on myocardial sympathetic innervation(UO2, WP 8). Images and ECG tracings will be transferred from the recruiting partners to a dedicated PACS system in UO1 using DICOM connectivity, and the various core labs will hook to the MILAN PACS to download images on the workstations for reading. Data will then be transferred to UO4 to develop the Machine Learning algorithm (WP9) and the App (WP10). The predictive accuracy of the App will be validated in UO4 using 50% of the enrolled patients. The evaluation of the economic impact on national health systems of a strategy that implants ICDs only in high arrhythmic risk patients identified by App, independently on patients' LVEF, versus the implantation of ICDs guided by current clinical guidelines will be performed in UO3. Finally, all the RESPECT project partners will collaborate to interpret, write, and disseminate the results. The publication record of the partners and their national and international visibility will guarantee the wide and quality dissemination of the results of the RESPECT project. In summary, the RESPECT project proposes combining several partners with cutting-edge complementary competencies to contribute to the project jointly. The proposed network will progress in an area of medical research that requires interdisciplinary interaction to develop a web-based platform that leverages artificial intelligence, data visualization, and mobile health technologies to empower physicians to personalize the risk SCD of their patients independent of the value of LVEF and pursue personalized prevention programs to reduce the number of patients who die suddenly or who survive a CA with residual post-ischemic cognitive impairment. The RESPECT consortium is a unique marriage of different but highly complementary scientific fields and expertise that will allow a comprehensive evaluation of the potential clinical and cardiac predictors of SCD. The partners of the RESPECT Project are worldwide renowned experts in their different fields, know each other very well, and have worked together to run scientific projects. Moreover, all of them are working in large academic organizations able to recruit many patients, ensure strict adherence to the protocol, and ensure the high quality of the clinical data, imaging datasets, and electrocardiography tracings.

Specific aim 1

1. To use statistical and Artificial Intelligence techniques to integrate demographic, clinical, humoral, genetic, electrocardiography, echocardiography, nuclear imaging, and cardiac magnetic resonance data from 1250 patients with ischemic and non-ischemic LV dysfunction to obtain a multiparametric predictive model to personalize the risk of occurrence of SCD based on phenotyping individual patients based on complex information similarity. 2. To implement this approach in a web-based app and test its accuracy to predict the individual risk of SCD To reach these aims, the research partners will enroll consecutive patients of both genders with either ischemic or nonischemic LV dysfunction defined as LVEF< 50% measured by two-dimensional echocardiography) and NYHA class II-III.

A recruitment period of 12 months to enroll 250 patients/UO and a minimum follow-up of 12 months are planned. In the sample, according to inclusion criteria, a cumulative incidence of hard events (SCD, resuscitated cardiac arrest, appropriate ICD intervention (i.e., both anti-tachycardia pacing and device shock), and from 10 to 15% in a median follow-up of 12 months is expected. The expected power to detect a difference between a predictive model with an area under the ROC curve (AUC) higher than 80% (e.g., 82%), compared with a reference AUC of 0.75, will be between 86% and 92%.

The development of the predicting model will employ and integrate a variety of statistical and Artificial Intelligence. The various steps of the development of the predicting model are summarized as follows:

1a) feature selection and construction will be achieved by investigating filters, wrappers, and embedded approaches, together with unsupervised deep learning embedding methods, to reduce the model complexity and prevent the overfitting of the available data set. Bayesian Networks (BN) will be used to model the probabilistic relationships between the selected variables and the constructed features. Furthermore, domain knowledge through knowledge elicitation will also be exploited.

Then, both the score-based and the constraint-based structural learning methods for BN will be used to build the predicting model;

1b) Semiparametric or parametric regression models with time-dependent outcomes (e.g., Cox or parametric models as appropriate) will be used to build a statistical predictive model by cross-validation. Variable selection will be performed with LASSO (the least absolute shrinkage and selection operator) models. Selected interaction terms between relevant new parameters and traditional parameters (i.e., LVFE value, NYHA class, ischemic vs. non-ischemic etiology) will be tested to detect if these new parameters can improve prediction for the individual patients; 2) image analysis will be carried out by exploiting Convolutional Neural Networks for classification tasks; 3) an expert system based on fuzzy reasoning will automatically integrate the output provided by steps 1a, 1b, and 2. The reasoner will be based on a (minimal) set of probabilistic fuzzy rules to train the clinical data and provide the final suggestion to the medical doctor. This solution will implicitly define a transparent, interpretable, and easily extensible decision-support AI system thanks to human-comprehensible statements.

Altogether, these methods are robust concerning the noise intrinsic in the clinical data and are effective in (i) learning, although the number of patients is relatively limited, (ii) leveraging and integrating the physicians¿ knowledge, and (iii) mitigating the overfitting.

To evaluate the economic impact on national health systems of a strategy that implants ICDs only in high arrhythmic risk patients identified by our predictive model versus the implantation of ICDs guided by current clinical guidelines The costs related to each implant of ICD will be calculated based on the average price of the device implant (17500€) and the mean annual per capita costs after ICD implantation (4136€, 95%CI: 4004-4262) estimated in the Lomb Cost-effectiveness of the personalized arrhythmic risk stratification model proposed will also be assessed about the potential implications for the National Health System related to the reduction of ICD implants in patients with LVEF<35% and low likelihood of clinical benefit, the number of additional implants of life-saving ICDS in patients with LVEF > 35%, and the number of ICD-related complications (i.e. infections, inappropriate discharges. Monte Carlo simulations will be used to evaluate the economic impact of the model, applying the model to the demographics, sanitary characteristics, and healthcare-related costs of the Lombardia region. This model will build two scenarios: the first one, in which the implant of ICDs is according to current guidelines, and the second one, in which ICDs are implanted only in patients identified by our predictive model to be at high risk of SCD. The expected specificity of the model will be 90-95%.

Specific aim 3 To assess the fundamental cellular arrhythmogenic mechanisms associated with specific cardiomyopathy mutations using in vitro cellular models based on patient-specific induced pluripotent stem cells (iPSC) To reach this goal, 4-5 patients with a genetic diagnosis and a documented outcome event during follow-up (cardiac arrest, appropriate ICD firing) will be recruited for a translational substudy. Mononuclear cells will be extracted from peripheral blood and subjected to reprogramming to generate patient-specific lines of induced pluripotent stem cells (iPSC). iPSC will be differentiated into cardiomyocytes (iPSC-CMs) using appropriate protocols for optimal cell maturation. iSPC-CMs from specific patient lines will be used to explore the main cellular arrhythmogenic mechanisms in vitro, aiming to correlate the disease-causing mutations with the primary arrhythmic triggers occurring at a single cardiomyocyte level. Cells will be studied with patch-clamp measurements and ion fluorescence assays. Molecular biology studies on relevant molecular targets will also be performed.