Identification of Digital Twins to Guide Interpretable AI for Diagnosis and Prognosis in Heart Failure

Summary Background Heart failure (HF) is a highly heterogeneous and complex condition. Although patient care generates vast amounts of clinical data, robust methods to synthesize available data for individualized management are lacking. Methods A mechanistic computational model of cardiac and cardiovascular system mechanics was identified for each individual in a cohort of 343 patients with HF. The identified digital twins — comprising optimized sets of parameters and corresponding simulations of cardiovascular system function—for patients with HF in the cohort is used to inform both supervised and unsupervised approaches in identifying phenogroups and novel mechanistic drivers of cardiovascular death risk. Findings The integration of digital twins into AI-based analyses of patient data enhances the performance and interpretability of prognostics AI models. Prognostics AI models trained with digital twin features are more generalizable than models trained with only clinical variables, as evaluated using an independent prospective cohort. In addition, the digital twin-based approach to phenomapping and predictive AI helps address inconsistencies and inaccuracies in clinical measurements, enables imputation of missing data, and estimates functional parameters that are otherwise unmeasurable directly. This approach provides a more comprehensive and accurate representation of the patient’s disease state than raw clinical data alone. Interpretation The developed and validated digital twin-based AI framework has the potential to simulate patient-specific pathophysiologic parameters, thereby informing prognosis and guiding therapeutic options.Ultimately, this approach has the potential to enhance the ability to focus on the most critical aspects of a patient’s condition, leading to individualized care and management.