利用機器學習方法和MIMIC資料庫早期預測加護病房中的急性腎衰竭

Abstract

背景：急性腎衰竭是一種可能導致預後不良的疾病，並且與加護病房患者的發病率和死亡率風險增加有關。它的發展非常迅速，通常在幾天內。因此，早期診斷和治療是一項重大挑戰。 目的：由於高發病率和死亡率，建立急性腎衰竭的預測模型至關重要。此研究的目標是要開發一個更全面的預測模型，能夠更早、更準確且更廣泛地預測急性腎衰竭。 方法：此研究的預測模型是用MIMIC-IV數據庫建立的，並使用MIMIC-III 數據庫評估模型的通用性。我使用了特徵重要性和t檢定來選擇具有預測力和通用性的特徵。在建構模型方面，我使用XGBoost和隨機森林模型以及增量式學習的技巧。AUROC為評估模型的指標。 結果：在提前 24、48和72小時預測急性腎衰竭的表現上，XGBoost模型的AUROC分別可以達到0.9489、0.9479和0.9466。當使用的選擇特徵對不同的數據集進行測試時，採取增量式學習後的模型可以提前24、48和72小時預測急性腎衰竭，其AUROC分別為0.7988、0.8074和0.7957。這表示我的模型和選擇的特徵具有良好的預測能力和通用性。本研究中選擇的變數可以廣泛應用於不同的數據。此外，本研究中建立的模型也有助於在不同的患者中早期預測急性腎衰竭。

Background: Acute kidney injury, also known as AKI, is a condition that can result in a poor prognosis and is linked to an increased risk of morbidity and mortality in intensive care unit (ICU) patients. It develops very quickly, typically within a few days. Therefore, early diagnosis and treatment pose a significant challenge. Objective: Creating a prediction model for acute kidney injury is essential due to the high morbidity and mortality rates associated with the condition. My objective was to develop a more comprehensive prediction model that is capable of predicting AKI earlier and more accurately. Methods: My prediction models were constructed with the Medical Information Mart for Intensive Care (MIMIC) IV database, and the generalizability of the models was evaluated using the MIMIC-III database. I employed feature importance and two sample t-tests in order to select features that were predictive and general. The construction of prediction models was handled by XGBoost and random forest models, both of which are tree-based algorithms. Also, I implemented the incremental learning technique when constucting the models. The area under the receiver operating characteristic (AUROC) served as the metric by which I evaluated our models. Results: The AUROCs of our XGBoost models can reach 0.9399, 0.9372, and 0.9373 when predicting AKI 24 hours, 48 hours, and 72 hours before onset, respectively. When testing on a different dataset using the selected features, the models after implementing incremental learning can predict AKI 24 hours, 48 hours, and 72 hours before onset with an AUROC of 0.8418, 0.7649, and 0.7479, respectively. This indicates that my models and selected features have good predictive ability and generalizability. The variables selected in this study can be widely applied to different data. Also, the models developed in this study have the potential to aid in the early prediction of AKI in a wide range of patients.