人工智慧在心電圖的心律不整分類

Abstract

人工智慧在提升心電圖於心律不整分類中的應用已變得至關重要。本研究利用2018年中國生理信號挑戰賽提供的包含6,877筆記錄的大型12導聯心電圖資料集，結合卷積神經網路和循環神經網路(CNN+RNN)開發出一個人工智慧模型。該模型在2,954筆心電圖隱藏測試集上的九類心律不整分類中，達到了0.84的整體F1-score中位數，展現了其在心律不整檢測中的卓越效能。進一步分析顯示，該模型能充分預測具有多種心律不整診斷的患者。值得注意的是，使用單導聯資料時，模型性能僅輕微下降，其中aVR和V1導聯最具資訊價值。 為解決便攜式/穿戴式裝置在長期心電圖監測中，因電池壽命和傳輸頻寬限制所使用的低解析度信號帶來的準確性下降，我提出了一個基於深度學習的心電圖信號超解析框架（SRECG）。SRECG透過在應用於高解析度訊號多分類器（HMC）的心律不整分類時提升低解析度心電圖信號的準確性。實驗結果顯示，與傳統的插值方法相比，SRECG顯著提高了HMC的心律不整分類準確性，證實了SRECG在增強來自便攜式/穿戴式裝置的低解析度心電圖信號以改善基於雲端的HMC性能方面的可行性。 此外，我研究了最初為心電圖心律不整分類設計的模型在其他生理信號中的多模態應用潛力，尤其是結合語音信號、人口統計資料和結構化醫療記錄以檢測聲門腫瘤。在從良性聲音障礙檢測聲門腫瘤的應用中，我的模型取得了顯著的準確性，展現了人工智慧方法學在跨不同生理信號診斷上的潛力。 總結而言，人工智慧技術，尤其是深度學習模型，在心電圖分析和心律不整分類方面取得了重大進展。強大的CNN+RNN模型、SRECG等信號增強技術，以及人工智慧模型在其他生理信號中的多模態適應性，展示了人工智慧在廣泛的醫學診斷領域中的應用潛力。

Artificial intelligence (AI) has become crucial in enhancing electrocardiography (ECG) classification for cardiac arrhythmia (CA). Leveraging a large 12-lead ECG dataset with 6,877 records from the 2018 China Physiological Signal Challenge (CPSC2018), I developed an AI model combining convolutional neural network and recurrent neural network (CNN+RNN). This model achieved a median overall F1-score of 0.84 across nine CA categories on a hidden test set of 2,954 ECG data, demonstrating high efficiency in CA detection. Further analysis showed that the model effectively predicted coexisting CAs in patients with multiple CA diagnoses. Notably, performance only slightly decreased when using single-lead data compared to the full 12-lead dataset, with the aVR and V1 leads providing the most informative signals. To address challenges in long-term ECG monitoring with portable/wearable (P/W) devices, such as battery constraints and limited transmission bandwidth resulting in low-resolution signals, I proposed a deep learning (DL)-based ECG super-resolution framework (SRECG). SRECG enhances low-resolution ECG signals, focusing on accuracy for CA classification when applied to high-resolution multiclass classifiers (HMC). Experimental results showed that SRECG significantly improved HMC accuracy over traditional interpolation methods, confirming the feasibility of using SRECG to enhance low-resolution ECG signals from P/W devices, improving cloud-based HMC performance. Additionally, I investigated the adaptability of the model, initially designed for ECG CA classification, to other physiological signals. Specifically, I explored its multimodal potential to detect glottic tumors in benign vocal disorders by integrating voice signals, demographic data, and structured medical records. My model achieved remarkable accuracy, highlighting the potential of AI methodologies to transition across different physiological signals for diagnostic purposes. In summary, AI, particularly DL models, has achieved significant advances in ECG analysis and CA classification. The integration of powerful CNN+RNN models, signal enhancement techniques such as SRECG, and adaptability to other physiological signals underscores the transformative impact of AI’s potential for broader medical diagnostics.