Objective:
To evaluate the diagnostic accuracy and categorical agreement of an AI-based echocardiography solution against expert-annotated ground truth (GT). Because most echocardiographic categorical findings derive from numerical measurements applied through standardized threshold tables, this study aimed to validate the AI by comparing its underlying numerical inputs to confirm clinical diagnostic agreement across standard structural and functional cardiac parameters.
Materials and Methods:
A retrospective validation study analyzed 20 clinical echocardiograms. We extracted 22 standard numerical measurements from GT reports and AI outputs. To improve measurement consistency, the AI systematically averages values across multiple cardiac cycles to minimize beat-to-beat variability. These numerics served as direct proxies to evaluate categorical agreement across eight domains, including Left Ventricular (LV) Size, LV Systolic Function, LV Hypertrophy Pattern, and Diastolic Function Grade. Regional Wall Motion Abnormality was assessed visually as it cannot be derived from standard numerics. Mean Absolute Error (MAE) was also calculated for all findings to quantify numerical deviation.
Results:
The AI demonstrated robust numerical agreement and highly proficient categorical equivalence in primary functional assessments. It achieved excellent agreement with GT for LV Systolic Function (100%), Diastolic Function Grade (100%), Filling Pressures (95%), LV Size (95%), and Pulmonary Hypertension Probability (91%). Moderate agreement occurred for LA Size (77%) and RV Systolic Function (70%). LV Hypertrophy Pattern agreement was 41%, primarily driven by unavailable Body Surface Area (BSA) data required for indexing LV Mass rather than an intrinsic AI measurement limitation. MAE was optimal for derived ratios and small measurements, yielding a low 0.11 error for the Diastolic Grade E/A ratio.
Conclusion:
The AI echocardiography solution is highly proficient at categorizing primary systolic and diastolic functional states. By averaging measurements across multiple cardiac cycles, the system delivers outstanding accuracy in core echo categorical, competing with manual reproducibility despite minor variances in specific 2D-based structural measurements.
Authors: Chhabra T , Thakkar D , Amanullah M , Poh Li Choo R 2 , Surve A , Anand S , Mahajan V , Rao S , Ang Y , Heng X , Lee Q
