- 期刊
運用深度學習建置護理健康問題推薦系統
Applying Deep Learning to Construct a Nursing Health Problem Recommendation System
摘要
護理計畫是護理照護的專業核心，其訂定包括健康問題的確立，隨著臨床疾病的多變性和複雜性增加，新進護理人員在短時間內完成全面評估並確立健康問題面臨挑戰，因此，運用人工智慧深度學習技術，建構護理健康問題推薦系統至關重要。本研究採用觀察型回溯設計，利用結構性及非結構性資料，分兩個階段比較深度學習、遞歸神經網絡與自然語言處理的Big Bird模型，第一階段評估不同模型和資料組合的性能，第二階段則重點測試Big Bird模型在大規模非結構性資料上的應用，從而選擇最佳的預測模型來進行健康問題推薦。研究結果顯示，Big Bird模型在自然語言處理方面展現出卓越的預測性能，相當於結合結構性深度學習和非結構性遞歸神經網絡的模型，具備更高的準確率和預測力，即使僅利用病人入院前24小時的護理記錄，Big Bird模型也能達到0.791的召回率及0.871的AUC（area under the ROC curve），這反映出護理記錄中包含了豐富而關鍵的特徵信息，比結構化資料更能提供精確的健康問題推薦。未來，我們將持續優化模型，進一步提升其準確性與應用範圍，並進行實地驗證，以提供護理人員精準的健康問題推薦工具，輔助臨床決策，節省文書作業時間，提升醫療服務的質量與效率。
並列摘要
Solving health problems is central to nursing care. However, with the increasing complexity of clinical conditions and shorter hospitalization periods, nursing time is becoming more compressed, which is posing major challenges for new nursing staff and increasing the risk of omissions in assessment processes. To address these challenges, a nursing health problem recommendation system can be developed using artificial intelligence and deep learning technologies. This study adopted an observational retrospective design, incorporating both structured and unstructured data to provide a comprehensive analysis. The study was conducted in two phases: the first phrase compared the performance of deep learning, recurrent neural networks, and the Big Bird model for natural language processing across different model-data combinations. The second phase focused on applying the Big Bird model to large-scale unstructured data to determine the best predictive model for making health problem recommendations. The results are promising, indicating that the Big Bird model has excellent predictive performance, on par with that of models combining structured deep learning with an unstructured recurrent neural network. Significantly, even when utilizing only nursing records from the first 24 hours of patient admission, the Big Bird model achieved 0.791recall and an AUC (area under the ROC curve) of 0.871. These findings indicate that nursing records contain critical and nuanced information, making them more effective for health problem recommendations than structured data alone. Looking ahead, we will continue to optimize the model to enhance its accuracy and expand its application scope. Additionally, field validation will be undertaken to equip nursing staff with precise health problem recommendation tools, which will assist in clinical decision-making, reduce documentation time, and ultimately improve the quality and efficiency of health-care services.