| 摘要: | 當前醫療領域正在經歷智慧化轉型，其核心目標是提升治療的迅速性和精確度，這種轉變在醫學與人工智慧技術的結合上尤為顯著，尤其是大型語言模型在處理醫療數據上的應用。大型語言模型的出現有助於減輕急診科醫護人員的工作壓力，提高病患管理效率。如今急診科正面臨著處理多變醫療需求和供需不平衡的挑戰，所以及早辨識病患處置非常重要，尤其是在電腦斷層處置辨識方面。因此，開發能準確初診分類和提供及時處置的輔助系統，對於提高醫療效率和應對人力不足至關重要。觀察資料集過後，發現我們的資料集存在著嚴重資料不平衡的狀況，所以為了應對不平衡問題，我們計劃透過運用大型語言模型來進行少數類別的資料增量。這一策略將使模型能夠接觸到更多元化的資訊，從而提高其在學習和訓練過程中對稀有但重要類別的識別能力。一方面模型能夠學習到常見類別的特徵，另一方面也能夠捕捉到那些在數據集中出現頻率較低的關鍵特徵。除了資料增量之外，我們還將從臨床文本中提取關鍵的臨床特徵「疼痛指數」作為外部知識融入模型中。通過整合這些臨床特徵，模型將能夠更全面地理解患者的健康狀況，進行更加縝密的學習。
實驗結果表明，研究所提出的模型在效能方面顯著優於傳統機器學習模型，尤其在處理不平衡數據時更為突出。使用AUPRC作為評估指標時，預訓練模型的性能優勢更加明顯，本實驗提出之模型效能為AUROC=0.8808、AUPRC=0.5414，值得一提的地方是其AUPRC比傳統模型高出近20%，顯示出在處理包含少量關鍵正樣本的不平衡數據集時的卓越預測精確度。綜上實驗結果，我們所提出的模型在與現有的診斷方法相比時，展現出了優異的性能這顯著的進步使我們更接近於實現最有效和精確的電腦斷層掃描處置目標。這是由於模型深入融合的醫學專業知識和創新的模型架構，使得它能夠對臨床病歷進行更細致和精確的分析和解釋，從而提供更可靠的診斷結果。此外，透過數據增量和引入外部知識特徵，讓模型得以更全面地學習和理解病患的狀態，不僅有助於識別典型的病理特徵，還能捕捉到對臨床處置來說極為關鍵的臨床情況，為醫生提供更豐富的信息以作出更準確的臨床決策，提高診斷的準確性和可靠性。綜合來看，我們的模型結合了最新的技術創新和深厚的臨床知識，為提高醫療診斷的準確性和效率開辟了新的道路。
The medical field is currently experiencing a shift towards intelligent solutions, aiming to increase the speed and accuracy of treatments. This transformation is notably seen in the integration of medical science with artificial intelligence, particularly through the use of large language models to process medical data. The advent of large language models helps alleviate the workload of emergency department medical staff and improve patient management efficiency. Emergency departments are currently facing challenges in handling variable medical needs and supply-demand imbalances, making the early identification and management of patient treatment crucial, especially in computer tomography (CT) diagnosis. Therefore, developing systems that can accurately categorize initial diagnoses and provide timely treatment is vital for improving medical efficiency and addressing workforce shortages. After examining our dataset, we found a severe imbalance in the data. To tackle data imbalances, we intend to employ large language models for augmenting data of minority classes, enhancing the model's access to a diverse information range. Moreover, we plan to extract crucial clinical features, such as the "pain index," from clinical texts and integrate them into the model as external knowledge. This integration will enable the model to gain a more comprehensive understanding of patients' health conditions and facilitate more detailed learning.
Experimental results reveal that our proposed model significantly surpasses traditional machine learning models, particularly in managing imbalanced data. When evaluated using AUPRC, the pre-trained model demonstrates a notable performance edge, achieving an AUROC of 0.8088 and AUPRC of 0.5414. Impressively, its AUPRC is almost 20% higher compared to traditional models, showcasing superior predictive accuracy in processing imbalanced datasets with a limited number of crucial positive samples. Our model advances toward achieving highly effective and precise CT scan treatment objectives, thanks to its integration of medical expertise and innovative design. It offers more detailed and accurate clinical record analysis, leading to reliable diagnostic outcomes. The model's learning is enhanced by data augmentation and integration of external knowledge, allowing it to thoroughly understand patient conditions, identify typical pathological features, and detect critical clinical scenarios. This model symbolizes a novel approach in medical diagnosis and treatment. |
