Division of Cardiology, University of British Columbia, Vancouver, BC, Canada.
Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, BC, Canada.
Int J Cardiovasc Imaging. 2021 Jan;37(1):229-239. doi: 10.1007/s10554-020-01981-8. Epub 2020 Nov 19.
We developed a machine learning model for efficient analysis of echocardiographic image quality in hospitalized patients. This study applied a machine learning model for automated transthoracic echo (TTE) image quality scoring in three inpatient groups. Our objectives were: (1) Assess the feasibility of a machine learning model for echo image quality analysis, (2) Establish the comprehensiveness of real-world TTE reporting by clinical group, and (3) Determine the relationship between machine learning image quality and comprehensiveness of TTE reporting. A machine learning model was developed and applied to TTEs from three matched cohorts for image quality of nine standard views. Case TTEs were comprehensive studies in mechanically ventilated patients between 01/01/2010 and 12/31/2015. For each case TTE, there were two matched spontaneously breathing controls (Control 1: Inpatients scanned in the lab and Control 2: Portable studies). We report the overall mean maximum and view specific quality scores for each TTE. The comprehensiveness of an echo report was calculated as the documented proportion of 12 standard parameters. An inverse probability weighted regression model was fit to determine the relationship between machine learning quality score and the completeness of a TTE report. 175 mechanically ventilated TTEs were included with 350 non-intubated samples (175 Control 1: Lab and 175 Control 2: Portable). In total, the machine learning model analyzed 14,086 echo video clips for quality. The overall accuracy of the model with regard to the expert ground truth for the view classification was 87.0%. The overall mean maximum quality score was lower for mechanically ventilated TTEs (0.55 [95% CI 0.54, 0.56]) versus 0.61 (95% CI 0.59, 0.62) for Control 1: Lab and 0.64 (95% CI 0.63, 0.66) for Control 2: Portable; p = 0.002. Furthermore, mechanically ventilated TTE reports were the least comprehensive, with fewer reported parameters. The regression model demonstrated the correlation of echo image quality and completeness of TTE reporting regardless of the clinical group. Mechanically ventilated TTEs were of inferior quality and clinical utility compared to spontaneously breathing controls and machine learning derived image quality correlates with completeness of TTE reporting regardless of the clinical group.
我们开发了一种用于高效分析住院患者超声心动图图像质量的机器学习模型。本研究应用机器学习模型对三组住院患者的经胸超声心动图（TTE）图像质量进行自动评分。我们的目标是：（1）评估机器学习模型用于超声心动图图像质量分析的可行性，（2）通过临床组确定真实世界 TTE 报告的全面性，以及（3）确定机器学习图像质量与 TTE 报告全面性之间的关系。开发了一种机器学习模型，并将其应用于三组匹配队列的 TTE 中，以评估九个标准切面的图像质量。病例 TTE 是在 2010 年 1 月 1 日至 2015 年 12 月 31 日期间接受机械通气的患者的全面研究。对于每个病例 TTE，都有两个匹配的自主呼吸对照（对照 1：在实验室接受扫描的住院患者；对照 2：便携式研究）。我们报告了每个 TTE 的整体平均最大和切面特定质量评分。超声心动图报告的全面性通过记录的 12 个标准参数的比例来计算。应用逆概率加权回归模型来确定机器学习质量评分与 TTE 报告完整性之间的关系。共纳入 175 例机械通气 TTE，175 例非插管样本（对照 1：实验室；对照 2：便携式）。总共，机器学习模型分析了 14086 个视频片段的质量。该模型在视图分类方面的整体准确性相对于专家真实数据为 87.0%。机械通气 TTE 的整体平均最大质量评分（0.55[95%CI 0.54,0.56]）低于对照 1：实验室（0.61[95%CI 0.59,0.62]）和对照 2：便携式（0.64[95%CI 0.63,0.66]），p=0.002。此外，机械通气 TTE 报告的全面性最低，报告的参数较少。回归模型证明了超声心动图图像质量与 TTE 报告完整性之间的相关性，无论临床组如何。与自主呼吸对照相比，机械通气 TTE 的质量和临床应用较差，无论临床组如何，机器学习衍生的图像质量都与 TTE 报告的完整性相关。
