ECG-X: Making ECGs explainable with colour to support early detection of life-threatening heart conditions

在一家繁忙的医院里，一位初级急诊医生正在学习解读心电图。以前，她会在纸上计算小正方形，并尝试用三角函数来手动确定波的开始和结束——例如绘制切线——而现在，她正在iPad上探索数据，改变ECG的轴和方向，在曲线下的区域涂上颜色，并放大以了解信号的细节。她用来理解数据的可视化工具也以另一种方式支持她：提供她正在调查的条件的可能性的自动估计，用自然语言解释它是如何得出判断的。因为自动解释使用了“认知契合”，即人类和机器共享相同的数据表示，所以它直观地可以理解，并且易于检查和探索。在这种情况下，拟合是由心电图数据的可视化呈现提供的。支持人工解释的技术，将预先注意处理与临床知识相结合，突出信号中的异常，以确保可靠性和准确性，也被用作机器解释的基础。”一位病人开始服用一种治疗癌症的新药。他没有注意到它会产生任何副作用，但他的智能手表提醒他，他的心脏电活动可能发生了变化，并建议他紧急咨询临床医生。他在一次紧急会诊中向医生展示了这些数据，并改变了他的药物。他的心脏活动很快恢复正常。“上述场景显示了我们研究目标的转变——从专家手动解读心电图的困难转变为在家自我监测，以检测可能导致心源性猝死的情况。心电图（ECG）是心脏电活动的图形表示，在临床实践中广泛用于检测心脏病理。众所周知，ECG解释是复杂的，对人类和机器都具有挑战性。本研究将开发一套用于询问心电图数据的可视化技术，将视觉感知原理与临床知识相结合，以创建决策支持工具，使人类和机器共享相同的数据表示。该研究将分为两部分：首先，可视化技术将与临床医生共同开发可靠的工具，这些工具将在临床实践中得到信任，并支撑理论，可作为操纵ECG数据以辅助解释的基础，并支持自动解释算法。第二，将对特定情况的选定数据呈现形式进行试验，并与公众一起进一步发展，以确定这些技术是否有可能被非专业人员用于监测自己的心脏健康。我们的长期愿景是设计临床可靠且可解释的类人人工智能，使患者或其护理人员能够在临床环境之外直观地自我监测他们的心电图，以发现可能危及生命的心脏病。这一愿景的最终目标是促进心源性猝死的一级和二级预防。心源性猝死是一种灾难性事件，占心血管死亡的50%，每年在美国造成30万人死亡，在英国造成6万人死亡。导致心源性猝死的心脏电问题通常只能在心电图上检测到，而缺血性心脏病的早期症状可以在其他主要症状出现之前在心电图上检测到。因此，改善心电图解释对于早期发现潜在致命性心脏病至关重要。更快的诊断和自我监测的能力将特别有利于女性，因为她们经常因为与男性不同的症状表现而延迟治疗。

"In a busy hospital, a junior A&E doctor is learning to interpret ECGs. Where once she would have been counting tiny squares and trying to manually determine the beginning and end of waves using trigonometry - for example drawing tangents - on a paper printout, instead she is exploring the data on an iPad, changing the axes and orientation of the ECG, applying colour to the area under the curve and zooming in to understand details of the signal. The visualisation tool she uses to understand the data also supports her in another way: providing an automated estimation of the likelihood of the conditions she is investigating, explaining in natural language how it has come to its judgement. Because the automated interpretation uses 'cognitive fit' - where the human and the machine share the same representation of the data - it is intuitively understandable, and easy to check and explore. In this case, the fit is provided by the visual presentation of the ECG data. The techniques that support human interpretation, which combine pre-attentive processing to highlight anomalies in the signal with clinical knowledge to ensure reliability and accuracy, are also used as the basis for the machine interpretation.""A patient has started taking a new medication for cancer treatment. He hasn't noticed it causing any side effects, but his smart watch has alerted him that the electrical activity of his heart may have changed, and advises him to consult a clinician urgently. He shows the data to his doctor in an emergency consultation, and has his medication changed. His heart activity soon returns to normal."The scenarios above show the transformation our research aims to achieve - moving from difficult manual interpretation of ECGs by experts, to self-monitoring at home to detect conditions that may lead to sudden cardiac death.The Electrocardiogram (ECG) is a graphical representation of the heart's electrical activity that is widely used in clinical practice for detecting cardiac pathologies. ECG interpretation is known to be complex, challenging both humans and machines. This research will develop a suite of visualisation techniques for interrogating ECG data, combining principles of visual perception with clinical knowledge to create decision support tools where humans and machines share the same representation of the data. The research will have two strands:In the first, visualisation techniques will be co-created with clinicians to develop reliable tools that will be trusted in clinical practice, and underpinning theory that can be used as a foundation for manipulating ECG data to assist interpretation, and support automated interpretation algorithms.In the second, selected forms of data presentation for specific conditions will be trialled and further developed with members of the public, to determine whether the techniques could potentially be used by lay people for monitoring their own cardiac health. Our long-term vision is to engineer clinically reliable and explainable human-like AI that will empower patients or their caregivers to intuitively self-monitor their ECGs for potentially life-threatening cardiac conditions outside the clinical setting. This vision ultimately aims to promote primary and secondary prevention of sudden cardiac death - a catastrophic event accounting for 50% of cardiovascular mortality, causing an estimated 300,000 deaths in the US and 60,000 deaths in the UK annually. Electrical problems with the heart leading to sudden cardiac death are often detectable only on an ECG, and the early signs of ischaemic heart disease can be detected on an ECG before other major symptoms occur. Improving ECG interpretation is thus essential for the earlier detection of potentially lethal heart conditions. Faster diagnosis and the ability to self-monitor will particularly benefit women, who often experience a delay in treatment due to different symptom presentation to men.