融合靜態與動態特徵之深度學習模型以預測急診再回 診：開發與驗證研究

Abstract

急診室(ED)再回診是急診醫學領域中一個重要的議題。若能準確預測患者再次於急診室就診的可能性，可以為主治醫師在評估患者出院與否時提供有價值的決策支援。在所有再回診病例中，識別高風險再回診尤為重要，因其可能表示患者的初始緊急護理不足，也是急診室評估醫療量能的一個指標。 儘管目前的研究已經提出了幾種用於預測急診室再回診的機器學習模型，但迄今為止，還沒有專門為此任務開發的基於深度學習之模型。且過去研究多著重於使用靜態特徵做預測，較少有效利用動態特徵。 在此研究中，我們使用了台大醫院急診室(NTUH)所提供之資料集，包含靜態與動態特徵。靜態特徵為病患進入急診室後紀錄之基本資料，如性別、年齡、檢傷級數等。動態特徵為患者在急診室暫留期間所測量之生命徵象，如血氧、心率等等。我們設計了一種資料預處理策略來解決動態特徵時間序列的不規則性，並提出了一個由TCN和FT-Transformer組成的混合模型。透過混合架構，我們的模型可以整合靜態和短期時間特徵以更全面地學習患者狀態，增強對潛在急診再回診病例的識別與預測，並克服僅依賴靜態特徵的模型的局限性。 首先，我們使用2016年至2019年的NTUH資料集來評估該模型。我們的模型在高風險回診預測任務中達到了0.8453的AUROC和 0.0935 的AUPRC；在一般回診預測任務中達到了0.7250的 AUROC和0.2005的AUPRC。接著，我們也利用NTUH 2020年至2022年的急診資料集作為我們的內部驗證集，並且在高風險再回診預測任務中實現了0.7976的AUROC和0.0103的AUPRC；在一般再回診預測任務中實現了0.7300的AUROC和0.1537的AUPRC。這兩個結果都明顯優於僅靜態的邏輯迴歸，AUPRC提高了約3.24倍，precision從0.0281提高到0.0428，提高了約1.52倍。此外，我們進行了特徵重要性分析，顯示心率對預測結果的影響最大。我們還進行了案例研究，以確定有助於模型預測的關鍵患者特徵。結果表明，我們的模型有效地整合了動態時間序列和靜態表格特徵，在預測急診再回診方面具有可靠的性能。這凸顯了融合不同類型的臨床資料來源以改善早期風險識別的潛力。

Emergency Department (ED) revisits represent a critical issue in emergency medicine. The ability to accurately predict whether a patient will return to the ED after discharge can provide valuable decision support for attending physicians. Identifying high-risk revisit cases is particularly important, as such events may reflect inadequate initial emergency care and serve as an indicator of the ED's medical capacity. Although prior studies have developed several machine learning models to predict ED revisits, there remains a lack of deep learning models specifically designed for this task. In addition, previous works have focused primarily on static features, while dynamic features have been underutilized. In this study, we utilize a dataset provided by National Taiwan University Hospital (NTUH) and incorporate both static and dynamic features for model training and evaluation. Static features include demographic and triage data collected upon ED admission, such as age, sex, and triage level. Dynamic features consist of vital signs measured throughout the ED stay, including blood oxygen saturation and heart rate. We design a preprocessing strategy to handle the irregularity of time series data and propose a hybrid deep learning model combining Temporal Convolutional Network(TCN) and FT-Transformer. This architecture enables integration of both static and short-term dynamic features to more comprehensively represent patient states, thereby overcoming the limitations of models that rely solely on static features and in turn improving the prediction of potential ED revisit cases. We first evaluate our model using NTUH data from 2016 to 2019. The model achieves AUROC of 0.8453 and AUPRC of 0.0935 in the high-risk revisit task, and AUROC of 0.7250 and AUPRC of 0.2005 in the general revisit task. We then validate the model using internal NTUH data from 2020 to 2022, where it achieves AUROC of 0.7976 and AUPRC of 0.0103 in the high-risk task, and AUROC of 0.7300 and AUPRC of 0.1537 in the general task. These results significantly outperform the static-only logistic regression baseline, with AUPRC being improved to approximately 3.24 times and precision being increased from 0.0281 to 0.0428 (about 1.52 times). Feature importance analysis indicates that the heart rate is the most influential dynamic factor. Furthermore, we conduct case studies to identify key patient features that contribute to the model's predictions. Our results demonstrate that the proposed model effectively integrates dynamic time series data and static tabular data, producing reliable performance in predicting ED revisits. This highlights the potential of multimodal clinical data fusion in improving revisit prediction and supporting clinical decision-making.