Use of artificial intelligence for the automatic assessment of left ventricular ejection fraction by oncology staff in chemotherapy patients

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Background The calculation of LV ejection fraction (LVEF) by transthoracic echocardiography is pivotal in detecting cancer therapy–related cardiac dysfunction. Referrals for LVEF estimation pre- and post-chemotherapy occupy significant amount of resources of echocardiography laboratories and increase service deliverance. Novel handheld ultrasound devices (HUDs) can provide echocardiographic images at the point of care with diagnostic image quality. Recently, artificial intelligence (AI) technology enabled the development of algorithms for the real-time guidance of ultrasound probe to acquire optimal images of the heart and calculate LVEF automatically. Purpose To evaluate the feasibility and accuracy of LVEF calculation by oncology staff using an AI enabled HUD. Methods We studied 115 oncology patients referred for echocardiographic LVEF estimation. All patients were scanned by a cardiologist using standard echocardiography (SE) systems and biplane Simpson’s rule was used as reference standard. A brief training on echocardiography basics and use of HUD was provided to the oncology staff before the study. Then, each patient was scanned independently by a cardiologist, a senior oncologist, an oncology resident, and an oncology nurse using the AUTO-GUIDANCE and AUTO-GRADING AI applications of the HUD (Figure 1) to acquire apical 4-chamber and 2-chamber views of the heart. The LVEF was automatically calculated by the device autoEF algorithm. Method agreement was assessed using Pearson’s correlation and Bland-Altman analysis. The diagnostic accuracy for detection of impaired LVEF<50%, a commonly used cut-off point for deferring chemotherapy, was calculated. Results Diagnostic images acquisition was possible in 96% of cases by the cardiologist, in 94% of cases by the senior oncologist, in 93% by the junior oncologist and in 89% by the nurse. The correlation between autoEF and SE-EF (Figure 2A) was excellent for the cardiologist (r=0.90), good for the senior oncologist (r=0.79), excellent for the junior oncologist (r=0.82) and excellent for the nurse (r=0.84), p<0.001 for all. The Bland-Altman plots (Figure 2B) showed a small underestimation of LVEF by the HUD autoEF algorithm compared to SE-EF for all the operators. There was bias −2.1% for the cardiologist, bias −3.5% for the senior oncologist, bias −2.2% for the junior oncologist, and bias −2.3% for the nurse (p<0.001 for all). Detection of impaired LVEF by autoEF algorithm was feasible with sensitivity 95% and specificity 94% for the cardiologist; sensitivity 86% and specificity 93% for the senior oncologist; sensitivity 95% and specificity 91% for the junior oncologist; sensitivity 94% and specificity 87% for the nurse. Conclusions Calculation of automated LVEF by oncology staff was feasible using AI enabled HUD in a selected patient population. Detection of impaired LVEF was possible with good accuracy. These findings show clinical potential to improve the quality of care for oncology patients.