Researchers have developed a deep-learning model capable of identifying individuals at high risk for sudden cardiac death using standard electrocardiogram (ECG) data, surpassing the predictive accuracy of current clinical standards.

研究人員開發了一款深度學習模型，能利用標準心電圖（ECG）數據識別心臟猝死高風險個體，其預測準確度超越了目前的臨床標準。

The prevailing clinical methodology for risk stratification relies upon the measurement of left ventricular ejection fraction (LVEF) via cardiac ultrasound. However, the authors note that LVEF is characterized by significant diagnostic insufficiency, frequently failing to identify high-risk patients while simultaneously flagging low-risk individuals for unnecessary defibrillator implantation. To address this gap, a 64-layer ResNet model was trained on a comprehensive dataset from a Swedish region, linking ECG waveforms to death certificates and electronic health records.

目前的臨床風險分層方法依賴於透過心臟超音波測量左心室射血分數（LVEF）。然而，作者指出 LVEF 存在顯著的診斷不足，經常無法識別高風險患者，同時又將低風險個體標記為需要植入除顫器。為了填補這一空白，研究團隊利用一個來自瑞典地區的綜合數據集，將心電圖波形與死亡證明及電子健康記錄相結合，訓練了一個 64 層的 ResNet 模型。

Quantitative analysis of a sequestered data 'lockbox' demonstrated that the model isolated a high-risk cohort (2.2% of the sample) with an annual sudden cardiac death rate of 7.0%, which exceeds the 4.6% rate observed in patients with reduced LVEF. Furthermore, 86.1% of the model's high-risk group were not identified by LVEF metrics. The model's generalizability was confirmed through 'zero-shot' validation in a US health system and a Taiwanese hospital registry, where it successfully predicted ventricular arrhythmias and arrhythmic cardiac arrests, respectively.

對一個獨立數據「鎖箱」的定量分析顯示，該模型分離出的一個高風險群體（佔樣本 2.2%）之年度心臟猝死率為 7.0%，高於 LVEF 降低患者中觀察到的 4.6%。此外，模型高風險組中 86.1% 的個體未能被 LVEF 指標識別。該模型的通用性透過在美國醫療系統和台灣醫院登記冊中的「零樣本」驗證得到證實，分別成功預測了心室心律不整和心律不整導致的心臟驟停。

To elucidate the specific morphology identified by the predictive model, the researchers employed a generative variational auto-encoder to create synthetic high-risk waveforms. This process revealed a previously undescribed biomarker in lead aVL, characterized by a slurred terminal aspect of the R wave. The researchers hypothesize that this morphology is indicative of diffuse myocardial fibrosis—the deposition of electrically inert collagen—which creates obstacles to conduction and precipitates electrical instability. This hypothesis was supported by blinded reviews of cardiac MRIs, which indicated a higher prevalence of subtle, diffuse late gadolinium enhancement in the highest-risk patients.

為了闡明預測模型識別的具體形態，研究人員採用生成變分自動編碼器來創建合成的高風險波形。這一過程揭示了 aVL 導極中一個先前未被描述的生物標誌物，其特徵為 R 波末端模糊。研究人員假設這種形態代表瀰漫性心肌纖維化——即電中性膠原蛋白的沉積——這會造成傳導障礙並導致電不穩定。這一假設得到了心臟 MRI 盲審的支持，結果顯示最高風險患者中瀰漫性晚期釓強化（LGE）的盛行率較高。

The study concludes that deep learning can identify a significant population at risk for sudden cardiac death who are currently overlooked by standard diagnostics, potentially expanding the candidate pool for life-saving defibrillator interventions.

研究結論指出，深度學習可以識別出一大部分目前被標準診斷忽略、具有心臟猝死風險的人群，潛在地下擴大了可接受救命除顫干預的候選對象範圍。