Machine learning and the diagnosis of myocardial infarction

胸痛是急诊科常见的症状。当病人出现这种症状时，医生必须排除心脏病发作的可能性。这是一种严重的疾病，需要及时治疗。为了诊断心脏病发作，医生必须与患者交谈，以了解他们的症状和病史，进行描述心脏电活动的心电图检查，并对肌钙蛋白进行血液检查。肌钙蛋白是一种在心肌受损时释放到血液中的蛋白质。心脏病发作通常会对心肌造成损伤，但也有许多其他情况会导致心肌损伤，并伴有胸痛。这些情况对心脏病发作需要不同的治疗方法，对医生和病人来说，区分这种情况来指导护理是很重要的。在过去的十年里，我们对心脏病发作的了解有了很大的提高。研究人员已经开发出更灵敏的血液检测方法，以高精度检测非常低浓度的肌钙蛋白。这改进了医生评估疑似心脏病患者的方式，并允许识别低风险患者并迅速排除心脏病发作，避免不必要的住院治疗。虽然非常低的肌钙蛋白水平可以让医生有信心安全地排除心脏病发作，但需要进一步的工作来帮助我们在肌钙蛋白水平升高的情况下诊断心脏病发作。在我的研究中，我建议使用数学、统计和计算方法来解决这一诊断困境。根据我们之前的研究，我们已经获得批准，可以访问来自苏格兰医院的45,000多名疑似急性心脏病发作的患者的医疗数据。这包括患者的年龄、性别、既往心脏状况、高血压史或吸烟史以及许多其他有用的变量。我们知道，其中一些信息与患者心脏病发作的风险有关。我们将使用这些数据来训练机器学习算法，以帮助我们了解除了肌钙蛋白测量外，在考虑心脏病发作诊断时哪些其他患者信息是重要的。机器学习是人工智能的一种形式，计算机软件被用来发现数据中的模式，并计算概率或“可能性”得分。一旦计算机系统熟悉了这种模式，它就会应用同样的方法来解决类似的数学问题(如：计算概率)，当出现类似的变量。例如，我们有一位54岁的女性病人，她的症状是胸痛。她是一个吸烟者，有高血压，心电图正常，血液肌钙蛋白水平略有升高。一旦我们在计算机模型中输入变量，它就会计算出一个分数，反映心脏病发作的概率，分数从1到100。在这个病人中，尽管肌钙蛋白检测结果为临界值，心电图正常，但由于她很年轻，肌钙蛋白水平应该很低，再加上她的其他危险因素，这种升高被认为是非常显著的。她的分数是75，这表明她很可能是心脏病发作。我们相信，与仅使用肌钙蛋白相比，这种评分将提供一种更可靠的诊断心脏病发作的方法。一旦我们开发并测试了我们提出的模型，我们计划评估这种新工具在患者中的表现。我们将对700名以胸痛为主诉的急诊科患者进行研究。一半的患者将使用该工具进行诊断，另一半患者将根据血液检查进行诊断。我们将比较这两种方法的性能。我们预计，包含重要患者特征的综合模型将帮助医生更准确地诊断心脏病发作。

Chest pain is a common presenting symptom in Emergency Departments. When patients present with this complaint doctors have to exclude the possibility of a heart attack. This is a serious medical condition, which requires prompt medical treatment. To diagnose a heart attack doctors have to speak to patients to find out more about their symptoms and medical history, to perform an electrocardiogram which describes the electrical activity of the heart and to take a blood test for troponin. Troponin is a protein which gets released into the blood stream when the heart muscle is damaged. Heart attacks commonly cause damage to the heart muscle, however there are many other conditions that can cause heart muscle injury and present with chest pain. These conditions require different treatment to a heart attack, and it is important for doctors and patients to make this distinction to guide care. Over the last decade our understanding of heart attacks has greatly improved. Researchers have developed more sensitive blood tests to detect very low concentrations of troponin with high precision. This has improved the way doctors assess patients with suspected heart attacks and allows identification of low risk patients and the rapid exclusion of a heart attack, avoiding unnecessary hospital admissions. Although very low levels of troponin can give doctors the confidence to safely exclude a heart attack, further work is needed to help us rule in the diagnosis of a heart attack in those where troponin levels are raised. In my fellowship I propose to use mathematical, statistical and computing methods to address this diagnostic dilemma. From our previous research we have approval to access medical data from over 45,000 patients who presented to hospitals in Scotland with a suspected acute heart attack. This includes information on the patients' age, gender, previous heart conditions, history of high blood pressure or smoking and many other useful variables. We know that some of this information is linked to a patient's risk of a heart attack. We will use this data to train a machine learning algorithm to help us understand which other patient information is important when considering a diagnosis of a heart attack in addition to the troponin measurement. Machine learning is form of artificial intelligence, where computer software is used to find patterns in data and to calculate a probability or "likelihood" score. Once the computer system is familiar with the pattern it will apply the same method to solve a similar mathematical problem (ie. calculate probability) when presented with similar variables. For example, we have a female patient who is 54 years old and presents with chest pain. She is a smoker and has high blood pressure with a normal electrocardiogram and a small increase in the blood troponin level. Once we input the variables in the computer model it will calculate a score that reflects the probability of a heart attack on a scale from 1 to 100. In this patient despite the borderline troponin result and normal electrocardiogram, because she is young and should have very low troponin levels, and because of her other risk factors the increase is considered to be highly significant. Her score is 75, suggesting she is very likely to have had a heart attack. We believe that this score will provide a more reliable way of diagnosing heart attacks compared to using troponin only. Once we have developed and tested our proposed model we plan to evaluate the performance of this new tool in our patients. We will conduct a study in 700 patients attending the Emergency Department with complaints of chest pain. Half of the patients will be diagnosed using the tool and half of them will be diagnosed on the basis of the blood test. We will compare the performance of the two methods. We anticipate that the integrated model that includes important patient characteristics will help doctors to more accurately diagnose heart attacks.