利用多模態深度學習模型結合胸部 X 光 和電子健康紀錄以篩檢急性心臟衰竭

Abstract

本研究旨在使用結合患者胸腔 X 光片和電子健康紀錄的多模態深度學習模型 於急診室進行急性心臟衰竭之篩檢。如果病人 N 端前腦利鈉肽的血液濃度高於 300ng/L，則被定義為陽性。對於每張胸腔 X 光片，肺心遮罩生成器最初通過肺心 分割模型識別肺和心臟區域。然後通過預定義的演算法手動提取三個估計心臟大 小的比例值和 306 個放射組學特徵。最終，多模態深度學習模型融合來自電子健 康紀錄和胸腔 X 光片的資訊並輸出最終預測。研究群體是從公開資料集 Medical Information Mart for Intensive Care (MIMIC) IV 中所提取。研究群體包括 1,432 名患 者和 1,833 對胸腔 X 光片和電子健康紀錄。其中，71% 的樣本為陽性。53% 的患 者為男性，47% 為女性。研究群體的年齡分佈最小為 20 歲最大為 91 歲，年齡平 均為 65 歲。此回溯性實驗在此資料集上顯示，當多模態深度學習模型整合來自每 個單模態模型的預測結果時，模型預測表性最高可達 AUROC 值 0.89。

This study aims to screen suspected acute heart failure (AHF) in the emergency department using a multimodal deep learning model combining patients'chest X-rays (CXRs) and electronic health records (EHRs). The binary label for AHF is defined as positive if a patient’s value of N terminal pro B type natriuretic peptide (NT-proBNP) is higher than 300 ng/L. For each CXR, the lung-heart mask generator initially identified the lung and heart region by the lung-heart segmentation model. Then three heart-size ratios and 306 radiomic features were extracted manually by predefined formulas. Eventually, the information from EHRs and CXRs was fused to output the final prediction. The study population was extracted from the Medical Information Mart for Intensive Care (MIMIC) IV open-source dataset. The study population includes 1,432 patients and 1,833 pairs of CXR and EHR. 71% of the samples are positive. 53% of the patient are male, and 47% are female. The age of the study population range from 20 to 91, with a mean of 65. The retrospective experiments illustrated that the proposed method achieved the highest AUROC of 0.89 when fusing all predictions from every single-modality model.