醫病平台
醫病平台
內容提供合作、相關採訪活動,或是投稿邀約,歡迎來信:
醫病平台
醫病平台
聽健康
編者按:幾週前才以ChatGPT為主題,由三位年長的醫師發表這新科技可能引起對醫病關係以及醫學教育的影響。本週我們再度以同一主題,由醫學院畢業沒幾年的年輕醫師寫出他們對人工智慧對醫療的影響。一位醫師討論「大型語言模型」對病歷書寫可能發生的影響。在目前煩忙的醫療工作中,如果能只輸入簡短正確的病史要點,而由電腦轉寫為文法通順且內容正確的住院、出院、轉院的病歷紀錄,可以改善醫療工作的效率。其他兩位醫師由另一角度探討這問題,認為這種科技的介入最大的受益者應該是普羅大眾,病人可以更容易獲得他們可以了解的醫療專業知識,讓病人在醫病關係中獲得較為平等的地位,並且因為診斷和檢驗效率的提高,而大幅降低等待時間,提高治療品質。同時,醫師也才能有時間改善與病人與家屬的互動,達到更理想的醫病關係。這三位年輕世代的醫師呼籲醫護人員、民眾和醫學生一起好好利用新技術,創造更美好的醫療環境,並在這個快速變化的世界中共同成長。
一位醫師從學生時期開始,在病歷寫作方面就有許多來自師長的要求,例如需要遵循正確的文法和用詞來撰寫病歷,以避免造成不必要的誤解和困擾,此外,還會有強制參加的病歷寫作比賽。在執業後,「病歷記載不實」、「病歷記載不詳」可說是高懸在每位醫師頭頂的達摩克利斯之劍。在評鑑的壓力下,醫師們必須確保他們的病歷完全符合標準和要求,診斷不能寫縮寫,HFrEF要寫成heart failure with reduced ejection fraction、T2DM要寫成Type 2 diabetes mellitus。然而,這些要求都造成醫療人員書寫紀錄遠大於與病人相處時間,同時給醫師們帶來壓力和負擔,特別是當他們需要在短時間內完成大量的病歷時,或連續上班36小時身心俱疲時。
隨著大型語言模型的出現,如ChatGPT,透過簡單的prompt(指令),醫師們現在可以更快速地得到他們想要的結果,進一步提高了病歷寫作的效率和準確性。此外,大型語言模型還可以幫助醫師們改進病歷寫作技能,通過檢查他們的語法和用詞來提高病歷的質量。
在過去,一些人提出了不同的解決方案,例如以自然語言書寫的報告,如病理和影像報告,改以條列式書寫,以減少病歷寫作的時間和工作量。此外,還有人提出結構化病歷,用表單的方式來幫助醫師們記錄病人的病情和治療過程。然而,這些方法也存在其缺點,例如過於僵硬和缺乏彈性,可能會限制醫師們的創造力和思維能力,並且可能會導致更多的工作量和時間浪費。
在臨床照顧上,由於內科病人的複雜度較高,如癌症病人、加護病房病人,他們有時間跨度很長的過去病史,每當有新的事件,病歷紀錄為求方便,往往是複製前後次,新增一段「This time, he developed acute abdominal pain. He was then admitted for further management.」,這樣的病歷雖然說不會遺漏任何之前的重要資訊,卻讓閱讀、接班的醫師無法在第一時間了解病人狀況。然而在我們所處的世界,萬物都傾向朝最大值熵(亂度)前進,包含病歷,要抵抗此趨勢就需要花費巨大能量,而ChatGPT就像是催化劑降低所需的功。
以下分享幾個常見的臨床情境(病人資訊皆為虛構或已經過修改)
咒語內容:
Outline and Sort the following medical record:
54M, LUL lung adenocarcinoma s/p LS1+2 segmentectomy & LND in 2019, pT3N2M1a (visceral pleural), s/p adjuvant VC x 4 concurrent RT to mediastinum 2019/11-2020/04, s/p Iressa 2020/04, with enlarged nodule at RLL, s/p RLL wedge resection, with brain mets 2021/03, s/p RT to brain tumor/skull mets (2021/5), s/p osimertinib (2021/04/14~) with leptomeningeal mets with involvement of cauda equina, esp. bilateral S2-5 roots s/p 30 Gy/ 10 Fr. RT to sacrum level nerve root (2022/11/07~11/18), under Osimertinib 2# QD (2022/12/15~) and #3 bevacizumab QOW (12/1) with more weakness of bilateral lower extremity weakness due to disease in progression, s/p L1~L5, 3000cGy/10 fractions (11/19~11/29), s/p #4 avastin (7.5mg/kg, self paid)
結果:
54M with history of LUL lung adenocarcinoma:
- LS1+2 segmentectomy & LND in 2019
- pT3N2M1a (visceral pleural)
- adjuvant VC x 4 concurrent RT to mediastinum (2019/11-2020/04)
- Iressa (2020/04)
- Enlarged nodule at RLL, s/p RLL wedge resection
- Brain mets (2021/03)
- RT to brain tumor/skull mets (2021/5)
- Osimertinib (2021/04/14~)
- Leptomeningeal mets with involvement of cauda equina, esp. bilateral S2-5 roots
30 Gy/10 Fr. RT to sacrum level nerve root (2022/11/07~11/18)
- Under Osimertinib 2# QD (2022/12/15~) and #3 bevacizumab QOW (12/1)
- Bilateral lower extremity weakness due to disease in progression
- L1L5, 3000cGy/10 fractions (11/1911/29)
#4 avastin (7.5mg/kg, self paid)
咒語內容:
Take the following information and write a medical narrative paragraph:
65M
DM, HTN on medication
p/w fever with cough 3d, AMS yesterday afternoon
no chest pain, no n/v, no abd pain
ER: vitals: 38c bp 100/60 RR 12 HR 120,
CBC: WBC 12k
Lab : Cr 2.0 aki, hypoK k 2.5
CXR: bil. infiltration
EKG: SR
DX: CAP s/p Tazocin
admitted
結果:
A 65-year-old male with a history of diabetes mellitus and hypertension, who was on medication for both, presented to the emergency room with a 3-day history of fever and cough, and an onset of altered mental status yesterday afternoon. The patient denied any chest pain, nausea, or abdominal pain. Vital signs showed a temperature of 38°C, a blood pressure of 100/60 mmHg, a respiratory rate of 12 breaths per minute, and a heart rate of 120 beats per minute. The CBC showed a white blood cell count of 12,000/µL. The patient's creatinine was elevated at 2.0 mg/dL, indicative of acute kidney injury, and the serum potassium was low at 2.5 mEq/L. A chest X-ray showed bilateral infiltrations and an EKG showed sinus rhythm. The patient was diagnosed with community-acquired pneumonia and was started on Tazocin. The patient was admitted for further management and treatment.
咒語內容:
summarize in narrative medical record
05/24 INR6 s/p FFP 4U。Afib rate s/p rate control add b-blocker for Afib
5/25: Bilateral massive plueral effusion >5cm
5/26 Afib RVR around 17:00 S/P Amiodarone loading once , check TSH/FT4 s/p NS bolus 250,
* HyperK. hold spironolactone
* 5/28 HR 150-170 > Amiodarone(300/900) > DC Digoxin
Bedside Echo: multiple B-lines, bilateral(L>R) side pleural effusion, RLL consolidation, EF
5/28 Hold Amiodarone, add concor
5/29 s/p thoracentesis
結果:
The patient presented with a high INR of 6 on 05/24 and was treated with FFP 4U. They were also diagnosed with Afib rate and underwent rate control, with a b-blocker added for Afib. On 05/25, the patient was diagnosed with bilateral massive pleural effusion greater than 5cm. On 05/26, the patient developed Afib RVR around 17:00 and was treated with a single dose of Amiodarone. TSH/FT4 was checked and the patient received an NS bolus of 250. Due to HyperK, Spironolactone was held. On 05/28, the patient had a high HR of 150-170 and was given Amiodarone (300/900), followed by DC Digoxin. Bedside Echo showed multiple B-lines, bilateral pleural effusion (L>R), RLL consolidation, and EF
從以上例子可以感受到,大型語言模型的出現為醫師們提供了一種更快速、更準確的病歷寫作解決方案。它不僅提高了病歷的質量,還可以減少醫師們的工作量和壓力,讓他們更專注於病人的護理和治療。許多人常常誤將語言模型理解成某種大型資料庫或百科全書,而對於它給出了錯誤答案嗤之以鼻,例如不知道如何算術等。實際上,語言模型所被訓練的目標,本來就不是去搜尋標準答案,而是用上文 (即prompt) 來推測出合適的下文,即見人說人話,見鬼說鬼話。如果問:某某癌症在某期別的治療,它大概率會回吐一個最常聽到的治療,而不是現今在實證上最好的治療。因此我們對於大型語言模型的期待,目前來說,比較類似一位很擅於整理文字的紀錄員。然而從另一個角度,許多醫師的門診病歷,在時間的壓力下,也都是以條列式為主,那些所謂病歷書寫的優美文句是完全看不到,取而代之的是一連串待辦事項,提示這位病人的主要診斷,以及後續的治療計畫等。也許有一天,我們可以捨棄那些來自傳統病歷寫作的繁文縟節,用精準簡潔的方式列點寫病歷即可,而那些需要看描述語句的人,自行丟給語言模型解釋即可。
3/31-醫病平台/ 醫學院教授的難題——沈默的槍手:人工智慧聊天機器人ChatGPT
責任編輯:吳依凡
這篇文章對你有幫助嗎?