Excerpted from Thinking, Hard and Soft
More people die of heart-related disease than from any other cause. The World Health Organization estimates that 17.5 million people died from cardiovascular diseases in 2012, about thirty percent of all global deaths.
A cardiovascular disease becomes an acute problem when it leads to a reduction in blood supply to the heart. For example, one of the coronary arteries supplying the heart with blood may be blocked by a blood clot. The part of heart muscle supplied by this artery will die without a blood supply, a process called myocardial infarction. “Heart attack” is the more familiar umbrella term.
Twenty years apart, here are two examples of data analysis in the treatment of acute heart disease: a data-based decision tool for patient classification and a “computationally generated biomarker”, which is a risk indicator derived from objectively observable clinical data using advanced data processing. The two examples reveal the progress that has been made and the challenges that lie ahead in using data analytics to improve healthcare.
To introduce the first example, suppose a patient arrives at a hospital with chest pain or other symptoms indicative of acute coronary trouble. Hospital staff – a triage nurse or a doctor – must make a decision: is this an emergency or not? It’s a crucial classification that can go wrong in two ways. Most catastrophically, a doctor may predict the patient is not in immediate danger when he is. In data reporting, predicting “no” when the actual answer is “yes” is called a “false negative”. The opposite case, a “false positive” is a patient sent to emergency care unnecessarily.
Racking up false positives can lead to serious problems. An emergency unit may be so clogged up with minor cases that a “true positive” cannot be processed. Unnecessary emergency care also increases a hospital’s running costs. However, because the likely consequence of a false negative is death, it is natural for doctors to be cautious and send patients into emergency care, just in case.
Defensive decision making can get out of control. Studies in the 80s and 90s reported that in some hospitals more than half of patients sent to coronary care units were identified as “false positives”. The Heart Disease Predictive Instrument was first developed in 1984 to improve patient classification. Emergency staff was taught to check patients for combinations of seven symptoms. A statistical table was used to convert the presence or absence of these symptoms into probabilities, which were tallied in a programmable calculator to reveal a predicted probability that the patient had an acute heart problem. The probabilities in the table were generated by a logistic regression trained against known cases in an early application of what might be called “machine learning” today.
The results were unexpected and counter-intuitive. Though unpopular with doctors, the instrument was a success: having been trained with the tool, doctors began to allocate patients optimally without using it. The “predictive machine” did not replace human intuition. Instead, it became useful as a training tool to improve intuition, focusing attention on the presence or absence of key symptoms: an example of machine-aided human learning at its best.
The quantity of medical data available per patient has increased significantly over the last twenty years: it has become “big”. Over the same period, computing power has increased and statistical methods have developed ever greater sophistication and have increasingly merged with computer science. While the essential heuristics used to classify emergency heart patients have been remarkably stable over time, data analytics may transform treatment for heart attack survivors over the next decade.
To introduce our second example, suppose that our patient has survived a first heart attack. Electrical signals traverse a healthy heart from end to end, triggering a synchronized pattern of contractions and expansions, but the tissue in our patient’s heart is scarred from myocardial infarction. The scar tissue can disrupt electrical impulses, causing his heart to become an unproductive pump. To guard against fatal electrical irregularities, an implantable cardioverter defibrillator (ICD) can be implanted into patients who have survived a heart attack. When an ICD detects an irregular rhythm (arrhythmia), it emits pulses to correct it.
An ICD reduces the risk of dying from arrhythmia, but not necessarily the overall risk of dying. The vast majority of implanted defibrillators are never activated, which means that most patients with an ICD don’t benefit from having one. On the contrary, some studies have shown that the overall risk of dying can increase with an ICD.
Traditionally, a recommendation to implant an ICD is based on a handful of objectively measurable indicators, called biomarkers. Examples of relevant biomarkers include the body mass index, the concentration of a small protein called BNP and, most importantly, the left ventricular ejection fraction (LVEF) – the proportion of the blood which, after entering the left ventricle of the heart, also exits there – usually above 55% in a healthy heart.
Patient classification for ICD treatment based on these biomarkers is unreliable. Up to nine in ten implanted devices are never activated. At the same time, many deaths could be prevented if more devices were implanted into patients who “actually are” at risk of fatal arrhythmia.
Recent research attempts to derive new and improved biomarkers based on readings from electrocardiograms, which measure a heart’s electrical activity. The basic idea is to compare the pattern of electrical activity between heart beats. A pattern that varies significantly from one pair of beats to the next indicates dangerous electrical instability.
A central technique used in the analysis is called dynamic time warping and is borrowed from the field of speech recognition. To recognize words in spoken text, you have to align identical words spoken at different speeds. The dynamic time warping algorithm was developed to achieve such alignment. Heart beats can be aligned using the same technique, making it possible to compare patterns of electrical activity between pairs of beats.
To test the new biomarker, researchers re-purposed a dataset originally collected to test a drug for angina (chest pain). In this dataset of 4500 patients who had suffered an acute coronary syndrome, 10% of the patients identified as high-risk using the new biomarker died within one year, while 3% of the patients identified as low-risk died during the same period. This means that patients in the high risk group were 3.3 times more likely to die in a year than those patients identified as low risk (this ratio is known as the hazard ratio). The claim is that these figures are significantly better than those derived using traditional indicators to classify patients. One of the next steps is to develop wearable monitors that can generate alerts based on real time analysis of beat-to-beat pattern variability.
Compared to a classification tool based on probability charts and a hand held calculator, modern analyses rely on more data, more computing power and greater mathematical sophistication. Yet the fundamental challenge is unchanged. Medical practitioners were reluctant to use probability charts because a probability chart is not intuitive. What do those numbers mean, and why should I trust them over my gut instinct? Similarly, it is unclear what the new computationally generated biomarker means. Why should a pragmatic cardiologist trust a black box? Experience with machine-aided learning for emergency triage twenty years ago suggests that it is crucial to establish a link between machine output and expert intuition before medical outcomes can be improved in practice.
Pozen, Michael W., et al. “A predictive instrument to improve coronary-care-unit admission practices in acute ischemic heart disease: a prospective multicenter clinical trial.” New England Journal of Medicine 310.20 (1984): 1273-1278.
Green, Lee, and David R. Mehr. “What alters physicians’ decisions to admit to the coronary care unit?” Journal of Family Practice 45.3 (1997): 219-226.
Syed, Zeeshan, et al. “Computationally generated cardiac biomarkers for risk stratification after acute coronary syndrome.” Science translational medicine 3.102 (2011).