ЭКГ-ийн бичлэгт гүн сургалтын загвар ашиглахад одоогийн аргаар илрүүлэхгүй өнгөрдөг өндөр эрсдэлтэй хүмүүсийг ялгаж болохыг судалгаа харуулжээ.
АНУ-д жил бүр 350,000 гаруй хүн гэнэтийн зүрхний зогсолтод өртдөг бөгөөд ихэнх нь эмнэлэгт хүрч чаддаггүй. Гэнэтийн зүрхний шалтгаант нас баралт тус улсад жилд 300,000 хүртэл тохиолдож, дэлхийн нийт нас баралтын 15 хүртэлх хувийг эзэлдэг гэж тооцоолдог.
Аюултай хэм алдагдлыг засах суулгац дефибриллятор байдаг ч түүнийг хэнд суулгах шаардлагатайг урьдчилан тогтоох амаргүй. Одоогоор зүрхний цус шахах үйл ажиллагааг хэт авиан шинжилгээгээр үнэлж, зүүн ховдлын шахалтын фракц (LVEF)-ийг ашиглан эрсдэлийг тодорхойлдог боловч энэ арга гэнэтийн нас баралт тохиолдох зарим хүнийг илрүүлэхгүй, харин дефибриллятор шаардлагагүй хүмүүсийг өндөр эрсдэлтэй гэж ангилж болдог.
Зиад Обермейер тэргүүтэй судлаачид Шведийн 440,000 гаруй ЭКГ бичлэгээр гүн сургалтын загвар сургаж, АНУ-ын бараг 140,000 хүний болон Тайванийн цөөн тооны өвчтөний бичлэгт туршжээ. ЭКГ нь арьсан дээр байрлуулсан электродоор зүрхний цахилгаан идэвхийг бүртгэдэг, харьцангуй бага зардалтай, инвазив бус шинжилгээ юм. Шведийн түүврийн 2.2 хувийг өндөр эрсдэлтэй гэж ангилсан загварын бүлэгт нэг жилийн дотор гэнэтийн зүрхний нас баралт 7 хувь байв; харин LVEF 1.9 хувийг өндөр эрсдэлтэй гэж тодорхойлж, тэдний 4.6 хувь нь тухайн хугацаанд нас баржээ. LVEF нь хиймэл оюуны илрүүлсэн өндөр эрсдэлтэй хүмүүсийн 86.1 хувийг таньж чадаагүй байна.
Хиймэл оюунаар өндөр эрсдэлтэй гэж тодорхойлогдсон, суулгац дефибриллятортой хүмүүсийн гэнэтийн зүрхний нас баралтын түвшин статистикийн хувьд хүлээгдэж байснаас 54.4 хувиар бага байсан нь төхөөрөмж зарим нас баралтаас сэргийлэх боломжтойг харуулж болох ч ажиглалтын судалгаа тул шалтгаант холбоог батлахгүй. Судлаачид мөн өндөр эрсдэлтэй хүмүүст ажиглагдсан, өмнө нь тодорхойлоогүй “налуу нь бүдэг” ЭКГ долгионы шинжийг илрүүлжээ. Энэ нь зүрхний булчингийн фиброзоос үүдэж, цахилгаан дохионы дамжуулалтыг алдагдуулж болзошгүй гэж үзсэн бөгөөд судалгааны үр дүн Nature сэтгүүлд нийтлэгджээ.
Дэлгэрэнгүйг эх сурвалжаас харах
↓Эх сурвалжийг нээх ↓
The human heart beats about 100,000 times a day. It’s the pounding cadence of life that makes you possible.
Until it doesn’t.
Every year, over 350,000 people in the US experience sudden and unexpected cardiac arrest. Most never make it to a hospital.
The fatalities are categorized as sudden cardiac death, generally defined as sudden, unexpected death within an hour due to a heart-related cause.
Sudden cardiac death is estimated to claim up to 300,000 American lives each year, and is thought to be responsible for up to 15 percent of all deaths worldwide.
But there’s a way to predict and potentially prevent many of these fatalities, a recent study in Nature reveals.
While implantable defibrillators can detect and correct dangerous cardiac arrhythmias before they develop into fatal cases of SCD, it’s difficult for doctors to know who should receive them.
Our best current system tries to identify the highest-risk patients by using ultrasound to assess blood flow, using a measurement called left ventricular ejection fraction (LVEF).
But LVEF fails to identify many patients who go on to experience SCD, and it also flags many people who never end up needing defibrillators to save their lives.
As it turns out, there could be a powerful new way to identify patients that LVEF overlooks.
“We have the cure for this problem,” says physician and AI researcher Ziad Obermeyer from the University of California, Berkeley.
“If you knew you were one of the people who was going to drop dead, you would go to a cardiologist and you’d get a defibrillator implanted. The problem is that doctors can’t figure out who needs one before it’s too late.”
In their study, Obermeyer and fellow researchers trained a deep-learning model to recognize electrical patterns associated with SCD, based on a collection of over 440,000 electrocardiogram (ECG) recordings sourced from Sweden, and also tested their AI system on ECG data from almost 140,000 patients in the US, plus a smaller group of patients from Taiwan.
The benefit of ECGs is that they’re a relatively low-cost and non-invasive testing system, capable of recording the heart’s electrical activity from electrodes placed on the patient’s skin.
“Deep learning applied to ECGs could provide the best of both worlds: a costless, widely available, standardized image that can be fed directly into a classifier, predicting risk without human input,” the researchers write in their paper.
As the results show, the concept holds promise.
From studying the trove of patterns in the ECG data, the AI system identified a high-risk group of patients that comprised 2.2 percent of the Swedish sample.
This high-risk group had a 7 percent rate of sudden cardiac death within the following year. By contrast, LVEF flagged a smaller group as high-risk (1.9 percent of the sample), fewer of whom went on to develop SCD (4.6 percent within a year).
But only a small fraction of the high-risk groups overlapped between the two measurement systems, with the researchers noting LVEF failed to identify 86.1 percent of the high-risk patients detected by the AI.
Those findings could have life-saving implications. Among high-risk patients identified by the AI, those with implanted defibrillators had a 54.4 percent lower sudden cardiac death rate than statistically expected, suggesting the devices could prevent some deaths in this group.
While the findings are observational and don’t establish that the defibrillators caused the reduction, it could be an important new lead for future research.
“Practically, high-risk patients seem to represent a new, previously unsuspected population with frequent, preventable death,” the study authors write.
But that’s not all the team found.
When the researchers used a second AI model to visualize the electrical patterns associated with higher risk, they uncovered something else: a previously undescribed waveform feature that independently predicts SCD.
According to the scientists, a “slurred downstroke” seen in the waveform data of high-risk patients could be a sign of impending heart trouble, possibly stemming from fibrosis (scarring) in heart muscle, which may disrupt the transmission of electrical impulses in the heart.
“The fact that a machine-learning approach can reveal a previously unseen pattern in such familiar data demonstrates a broader lesson: deep learning, when paired with appropriate complementary tools, can move beyond prediction, towards scientific discovery,” writes Changxin Lai, a biomedical engineer from Johns Hopkins University, who wasn’t involved with the study, but is the author of a commentary on the research.
“ECGs have been read clinically for more than a century, yet this work demonstrates that even the most extensively studied data could contain unread messages.”
The findings are reported in Nature.
This article was fact-checked by Fiona MacDonald and edited by Fiona MacDonald. While we pride ourselves on our process, we are only human. If you spot a mistake, please let us know.

