Machine learning for the generation of personalized image analysis protocol in echocardiography: a pilot study in arterial hypertension

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Agence National de la Recherche Europian Union-NextGenerationEU. Background Current workflow in cardiac imaging is based on protocols meant to capture a set of predefined images for a certain condition [1]. Although standardized, such approaches are not personalized, and are often time- and resource-ineffective. Reinforcement learning (RL), a subfield of machine learning that addresses decision making [2], can be used to incorporate complex, whole cardiac cycle imaging data, while optimizing costs associated to data acquisition and processing – including economy, patient safety/comfort and examination time. Our aim is to demonstrate the viability of RL to create a personalized, resource-effective data analysis routine from echocardiography to separate remodelling in arterial hypertension from normal cardiac function. Methods An echo study was performed in 189 clinically managed hypertensive (HTN) patients, and 60 non-hypertensive healthy control subjects [3]. Speckle-tracking analysis of the left ventricle was performed and deformation curves extracted. Aortic and mitral flow pulsed-wave (PW) Doppler and septal basal tissue PW Doppler velocity profiles were obtained. These whole cardiac cycle deformation and velocity curves were used as input data to the RL algorithm. RL optimized the required image analysis sequence to maximize the probability of obtaining a correct diagnosis while reducing the number of analysed images. Results The RL-generated protocol identified three separate image analysis groups to determine cardiac remodelling (Figure 1). The proposed protocol starts with the analysis of the mitral PW, and based on its velocity pattern, assigns each individual to one of the three groups. In Group 1 (G1, n=158) analysing the mitral PW acquisition was sufficient to determine disease status. HTN patients were separated based on a pattern of an inversed E/A ratio and EA fusion reflecting signs of diastolic dysfunction. In G2 (n=56), mitral flow showed a slightly decreased E/A ratio and longer E deacceleration time, and adding aortic PW was required to determine remodelling: HTN had higher aortic peak velocity and shorter peak acceleration time due to higher systemic afterload. Finally, G3 (n=19) required global longitudinal strain, which showed pathological early systolic stretching, reflecting the change in the timing of cardiac events in arterial hypertension. Fifteen (6%) participants did not match any of the groups, requiring the acquisition of remaining sequences (TDI, regional strain) to secure a correct diagnosis. In average, the protocol analysed 1.44 images per patient, out of the four available. Conclusion Through a pilot study utilizing a limited set of imaging sequences, we demonstrate a template for automated generation of efficient, tailored image analysis protocols, that minimimize the number of analyzed images to determinine patient phenotypes. RL is a promising tool to reduce analysis time, potentially saving costs in preventive screening programs.