| 摘要: | 背景：譫妄是一種在住院患者中普遍觀察到的神經性精神障礙。雖然現有臨床評估工具可識別譫妄，但其實際應用在照護中經常受到時間限制、評估一致性問題及專業醫護人員不足等因素的影響。面對這些挑戰，本研究探索使用自然語言處理（NLP）技術來增強譫妄檢測的效率和準確性。
方法：本研究包含了56位參與者共31,557份護理記錄，採用了護理知識增強型預訓練語言模型（NKLM），該模型融合了基於變換器的雙向編碼器（BERT）技術，專注於從護理記錄中識別譫妄症狀。NKLM不僅結合了時間特徵，如輪班類型、週末和傳統中國節日，還整合了BERT的嵌入功能。模型的可解釋性，得益於局部可理解的模型無關解釋法（LIME），LIME可使模型注意力機制能夠被清晰地展示出在識別譫妄時的關鍵特徵。此外，研究透過卡方檢定和費雪精確檢定的統計分析，進一步探討了時間變數對譫妄檢測的影響。
結果：本研究全面評估常見自然語言處理模型，NKLM在從護理記錄中識別譫妄方面表現最佳。NKLM靈敏度達到0.9265，特異性為0.9967，以及最佳的宏觀F1分數0.9611。AUC分析確立了其優越性，NKLM達到了0.9872的AUC，接近原始BERT觀察到的最高AUC 0.9931。利用LIME進行的視覺解釋，使得護理文件中與譫妄相關的關鍵術語得以精確識別，進一步證實了模型的精準度。此外，時間相關評估凸顯了如夜班和文化節假日等影響性時間特徵，強調了它們在提升NKLM預測能力方面的重要作用。
結論：在臺灣醫療系統中應用自然語言處理於譫妄檢測被證明是有效的。NKLM模型在靈敏度和特異性方面的表現特別突出，其方法論在識別護理記錄中的譫妄關鍵詞方面表現卓越，時間相關特徵顯著增強了檢測的精確性。
Background: Delirium is a prevalent neuropsychiatric disorder observed in hospitalized pa-tients. Despite clinical assessments' ability to identify delirium using established tools, their de-ployment is often hindered by time constraints, inconsistency, and the need for specialized per-sonnel.
Methods: We included a total of 31,557 records from 56 participants in the study. The Nursing Knowledge-Enhanced Pre-trained Language Model (NKLM) was introduced, utilizing the Bi-directional Encoder Representations from Transformers (BERT) for targeted delirium detection from nursing records. The NKLM integrates temporal features, such as shift types and tradi-tional Chinese holidays, with BERT's embeddings. The model's interpretability is enriched us-ing Local Interpretable Model-agnostic Explanations (LIME), revealing key features central to delirium prediction. Comprehensive analyses, applying Chi-square and Fisher's exact tests, further evaluated the influence of temporal variables on delirium detection.
Results: Through the comprehensive evaluation of NLP models, the NKLM model emerged as a top performer in delirium prediction from records. The NKLM exhibited an exceptional Sen-sitivity of 0.9265, a Specificity of 0.9967, and a standout Macro F1-score of 0.9611. Further, the AUC analysis cemented its superiority, with NKLM achieving an AUC of 0.9872, nearly matching the highest observed AUC of 0.9931 by BERT. Visual interpretation using LIME al-lowed for pinpointing key delirium-related terms in nursing documentation, reaffirming the model's precision. Moreover, time-related assessments spotlighted influential temporal features, such as night shifts and cultural holidays, emphasizing their role in bolstering NKLM's predic-tive prowess.
Conclusion: NLP is demonstrated to be effective at detecting delirium within Taiwan's healthcare system in this study. The NKLM model distinctly outperforms in sensitivity and specificity. Our methodology excels in identifying delirium keywords in nursing records, with time-related features enhancing detection precision. |
