以機器學習演算法偵測致命性心律不整

Abstract

突發性院外心跳停止是引發成人死亡的首要原因之一,經常由心 室顫動 (VF) 造成。即時偵測這些致命性的心律不整,並且盡早以自動 體外電擊器 (AED) 施予去顫,是治療關鍵。過去的研究提出了各種偵 測心室顫動的演算法,但是大部分並未遵循現行由美國心臟病協會所 制定的醫學標準。本論文呈現了一個基於支撐向量機的機器學習演算 法,並在演算法的發展和測試過程當中,謹慎的依循美國心臟病協會 的醫學標準。整體而言,此演算法滿足美國心臟病協會標準要求的性 能,達到 93.21 % 的敏感度、99.88 % 的特異性、以及 89.28 % 的精確 度。此外,本研究使用的測試資料,比起過去研究更為全面,並且由 內科醫師檢視過確保正確性。因此,對於未來自動體外去顫器演算法 的研究,本資料集或許可作為一個更好的測試標準。

Sudden out-of-hospital cardiac arrest, one of the leading causes of death among adults, is frequently caused by ventricular fibrillation (VF). Prompt recognition of these life-threatening arrhythmias and early defibrillation treat- ment using an automated external defibrillator (AED) are crucial. Previous researchers proposed various VF detection algorithms, but most of them did not comply with the existing medical standards for AED development set by the American Heart Association (AHA). This thesis presents a machine- learning AED algorithm based on support vector machine. The development and evaluation processes of the algorithm carefully followed the AHA med- ical standards. With an overall sensitivity of 93.21 %, specificity 99.88 %, and precision of 89.28 %, the proposed algorithm satisfied all of the perfor- mance goals required by the AHA guideline. In addition, the dataset used in our study was more comprehensive then that used in previous studies and was reviewed by a physician to ensure its correctness. Therefore, it might be a better benchmark for future researches of AED algorithms.