導語：人工智能與檢驗(yan)醫學(xue)輔助診療現狀一文來源于網(wang)友上傳(chuan)，不(bu)代表(biao)本站觀點，若(ruo)需要原創文章可咨詢(xun)客服老師，歡迎參考(kao)。

摘要:隨著互聯網、大數據及云計算等新興技術的快速發展,人工智能(AI)已被廣泛應用于影像和病理臨床輔助診療。檢驗醫學作為現代醫學的重要支撐,與AI的融合勢在必行。該文重點介紹了檢驗大數據如何與AI結合以構建檢驗AI,并展望了檢驗AI在疾病診斷、預測及風險評估等臨床輔助診療領域的應用前景。

關鍵詞:人工智能;檢驗(yan)醫(yi)學;檢驗(yan)人工智能系統(tong);臨床輔助診(zhen)療

人工智能(AI)是研究、開發用于模擬、延伸和擴展人類智能的一門新興的科學技術[1]。AI自誕生以來,經過不斷的演進與發展,在越來越多領域得以重視和應用。目前,AI技術已廣泛應用于醫療領域,出現了影像AI[2]、手術機器人[3]、藥物研發AI[4]等諸多應用場景。檢驗醫學作為臨床輔助診療體系中不可或缺的部分,貫穿患者入院、出院、藥物治療等過程,影響著患者的醫療服務質量[5-6]。檢驗醫學具有高度自動化、數據化及數據庫存儲等特點為檢驗AI的應用與研發創造了有利條件[7-8]。為此,本文結合AI在當下臨床輔助診療中的應用現狀,重點(dian)闡述檢(jian)驗醫學(xue)如何與AI融合構建檢(jian)驗AI,并展(zhan)望檢(jian)驗AI在輔助(zhu)臨(lin)床(chuang)診療中的應用前景。

1AI與臨床輔助診療

隨著(zhu)深度學習技術在目標檢(jian)(jian)測(ce)、圖(tu)像(xiang)分(fen)(fen)類(lei)(lei)及(ji)(ji)圖(tu)像(xiang)分(fen)(fen)割等(deng)領(ling)域(yu)取得(de)突破性(xing)(xing)的(de)(de)(de)(de)(de)進(jin)展,影(ying)像(xiang)AI和(he)(he)(he)病(bing)(bing)理(li)(li)(li)AI已成(cheng)為新(xin)的(de)(de)(de)(de)(de)醫學研究熱(re)點(dian)[9]。目前,影(ying)像(xiang)AI和(he)(he)(he)病(bing)(bing)理(li)(li)(li)AI已成(cheng)功(gong)應(ying)(ying)用(yong)于(yu)(yu)疾病(bing)(bing)的(de)(de)(de)(de)(de)篩查(cha)(cha)、預測(ce)及(ji)(ji)診(zhen)斷(duan)等(deng)領(ling)域(yu)。比如,ZHANG等(deng)[2]構(gou)建了一(yi)(yi)個包(bao)括新(xin)型冠狀病(bing)(bing)毒(du)肺炎(COVID-19)、常(chang)(chang)(chang)(chang)(chang)見肺炎及(ji)(ji)正(zheng)(zheng)常(chang)(chang)(chang)(chang)(chang)對照人群(qun)的(de)(de)(de)(de)(de)大規(gui)模(mo)CT數(shu)據(ju)集,并(bing)基于(yu)(yu)CT圖(tu)像(xiang)開發(fa)了COVID-19AI診(zhen)斷(duan)系(xi)統,用(yong)于(yu)(yu)幫助準(zhun)確(que)診(zhen)斷(duan)COVID-19。TERAMOTO等(deng)[10]使用(yong)深卷積神經(jing)網絡開發(fa)了微觀圖(tu)像(xiang)中的(de)(de)(de)(de)(de)肺癌(ai)細(xi)胞(bao)病(bing)(bing)理(li)(li)(li)自動(dong)分(fen)(fen)類(lei)(lei)模(mo)型,使約71%的(de)(de)(de)(de)(de)肺癌(ai)細(xi)胞(bao)得(de)到正(zheng)(zheng)確(que)分(fen)(fen)類(lei)(lei),且(qie)分(fen)(fen)類(lei)(lei)正(zheng)(zheng)確(que)的(de)(de)(de)(de)(de)圖(tu)像(xiang)具有典型的(de)(de)(de)(de)(de)細(xi)胞(bao)形態和(he)(he)(he)排列。然(ran)而,影(ying)像(xiang)和(he)(he)(he)病(bing)(bing)理(li)(li)(li)診(zhen)斷(duan)具有局(ju)限性(xing)(xing)。一(yi)(yi)方(fang)面(mian),病(bing)(bing)理(li)(li)(li)切片后診(zhen)斷(duan)周期(qi)較長、費用(yong)較高,影(ying)像(xiang)學檢(jian)(jian)查(cha)(cha)相對迅(xun)速但費用(yong)也相對較高。另一(yi)(yi)方(fang)面(mian),病(bing)(bing)理(li)(li)(li)和(he)(he)(he)影(ying)像(xiang)檢(jian)(jian)查(cha)(cha)結(jie)果(guo)(guo)僅僅是(shi)一(yi)(yi)些特(te)定(ding)疾病(bing)(bing)的(de)(de)(de)(de)(de)部分(fen)(fen)參考指(zhi)標,無法反映患者疾病(bing)(bing)的(de)(de)(de)(de)(de)整體(ti)情況。相較于(yu)(yu)影(ying)像(xiang)和(he)(he)(he)病(bing)(bing)理(li)(li)(li)檢(jian)(jian)查(cha)(cha),實(shi)驗(yan)(yan)(yan)(yan)室(shi)檢(jian)(jian)驗(yan)(yan)(yan)(yan)具有顯著(zhu)優(you)勢(shi)。首先(xian),實(shi)驗(yan)(yan)(yan)(yan)室(shi)檢(jian)(jian)驗(yan)(yan)(yan)(yan)方(fang)法簡(jian)單、快速且(qie)成(cheng)本(ben)低(di),醫生(sheng)(sheng)(sheng)通過(guo)分(fen)(fen)析(xi)患者的(de)(de)(de)(de)(de)檢(jian)(jian)驗(yan)(yan)(yan)(yan)數(shu)據(ju)即(ji)可(ke)(ke)實(shi)現良好的(de)(de)(de)(de)(de)診(zhen)療(liao)效果(guo)(guo),可(ke)(ke)提(ti)高臨(lin)床醫生(sheng)(sheng)(sheng)的(de)(de)(de)(de)(de)診(zhen)療(liao)效率并(bing)減輕(qing)患者的(de)(de)(de)(de)(de)醫療(liao)負(fu)擔。此外,血(xue)常(chang)(chang)(chang)(chang)(chang)規(gui)、血(xue)生(sheng)(sheng)(sheng)化(hua)、尿液或糞(fen)便等(deng)檢(jian)(jian)驗(yan)(yan)(yan)(yan)結(jie)果(guo)(guo)可(ke)(ke)直接(jie)反映疾病(bing)(bing)的(de)(de)(de)(de)(de)生(sheng)(sheng)(sheng)理(li)(li)(li)及(ji)(ji)病(bing)(bing)理(li)(li)(li)變(bian)化(hua)過(guo)程。臨(lin)床上常(chang)(chang)(chang)(chang)(chang)用(yong)的(de)(de)(de)(de)(de)檢(jian)(jian)驗(yan)(yan)(yan)(yan)數(shu)據(ju)具有足夠的(de)(de)(de)(de)(de)有效性(xing)(xing)和(he)(he)(he)穩定(ding)性(xing)(xing),已經(jing)過(guo)大規(gui)模(mo)的(de)(de)(de)(de)(de)臨(lin)床實(shi)踐和(he)(he)(he)評估(gu),并(bing)廣泛用(yong)于(yu)(yu)診(zhen)斷(duan)、排查(cha)(cha)、分(fen)(fen)類(lei)(lei)或監測(ce)各種疾病(bing)(bing),可(ke)(ke)以為臨(lin)床醫生(sheng)(sheng)(sheng)疾病(bing)(bing)診(zhen)療(liao)提(ti)供(gong)一(yi)(yi)個更(geng)為全面(mian)的(de)(de)(de)(de)(de)指(zhi)導建議[11]。目前,檢(jian)(jian)驗(yan)(yan)(yan)(yan)數(shu)據(ju)的(de)(de)(de)(de)(de)真(zhen)實(shi)診(zhen)斷(duan)效能(neng)通常(chang)(chang)(chang)(chang)(chang)被低(di)估(gu),主要是(shi)由于(yu)(yu)個體(ti)化(hua)檢(jian)(jian)驗(yan)(yan)(yan)(yan)數(shu)據(ju)復雜多變(bian),不(bu)(bu)同的(de)(de)(de)(de)(de)異(yi)常(chang)(chang)(chang)(chang)(chang)檢(jian)(jian)驗(yan)(yan)(yan)(yan)結(jie)果(guo)(guo)對應(ying)(ying)的(de)(de)(de)(de)(de)臨(lin)床意(yi)義(yi)不(bu)(bu)同,聯(lian)合(he)意(yi)義(yi)則更(geng)多、更(geng)廣。臨(lin)床醫生(sheng)(sheng)(sheng)面(mian)對眾多異(yi)常(chang)(chang)(chang)(chang)(chang)檢(jian)(jian)驗(yan)(yan)(yan)(yan)結(jie)果(guo)(guo)時,由于(yu)(yu)受限于(yu)(yu)人腦的(de)(de)(de)(de)(de)推理(li)(li)(li)和(he)(he)(he)計算能(neng)力,即(ji)使是(shi)經(jing)驗(yan)(yan)(yan)(yan)豐富(fu)的(de)(de)(de)(de)(de)醫學專家也經(jing)常(chang)(chang)(chang)(chang)(chang)忽略眾多結(jie)果(guo)(guo)之間隱藏的(de)(de)(de)(de)(de)關鍵信息(xi)和(he)(he)(he)重要趨勢(shi),不(bu)(bu)可(ke)(ke)避免地造成(cheng)疾病(bing)(bing)的(de)(de)(de)(de)(de)漏診(zhen)與誤(wu)診(zhen)。因此,使用(yong)AI方(fang)法集成(cheng)和(he)(he)(he)分(fen)(fen)析(xi)檢(jian)(jian)驗(yan)(yan)(yan)(yan)數(shu)據(ju)為提(ti)高檢(jian)(jian)驗(yan)(yan)(yan)(yan)數(shu)據(ju)的(de)(de)(de)(de)(de)診(zhen)斷(duan)價值提(ti)供(gong)了巨(ju)大的(de)(de)(de)(de)(de)應(ying)(ying)用(yong)前景[12]。

2基于檢驗數據的檢驗AI

2.1檢驗AI構建的(de)基(ji)本條件

目前,構建檢(jian)驗(yan)(yan)(yan)AI的(de)(de)(de)(de)(de)基(ji)本條件已(yi)(yi)經(jing)基(ji)本具備。首先可(ke)(ke)從醫院(yuan)檢(jian)驗(yan)(yan)(yan)信(xin)息系(xi)(xi)統(tong)(tong)(tong)(LIS系(xi)(xi)統(tong)(tong)(tong))中(zhong)獲(huo)得(de)大(da)量檢(jian)驗(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)。LIS已(yi)(yi)成(cheng)(cheng)為當今醫院(yuan)管理(li)信(xin)息(HIS)系(xi)(xi)統(tong)(tong)(tong)的(de)(de)(de)(de)(de)重要組成(cheng)(cheng)部分,該系(xi)(xi)統(tong)(tong)(tong)存(cun)儲(chu)了海(hai)量的(de)(de)(de)(de)(de)檢(jian)驗(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)[13],能(neng)夠為檢(jian)驗(yan)(yan)(yan)AI的(de)(de)(de)(de)(de)構建提(ti)(ti)(ti)供(gong)(gong)大(da)數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)支(zhi)持。其次,機(ji)器學習(xi)(ML)算(suan)(suan)法(fa)(fa)層出(chu)不窮。比(bi)如,極端梯度提(ti)(ti)(ti)升(XGBoost)作為當前比(bi)較流行(xing)(xing)的(de)(de)(de)(de)(de)集(ji)成(cheng)(cheng)學習(xi)算(suan)(suan)法(fa)(fa),在(zai)提(ti)(ti)(ti)升樹(shu)模(mo)(mo)(mo)型中(zhong),具有非常強的(de)(de)(de)(de)(de)擴展性(xing)和(he)靈活性(xing),能(neng)夠集(ji)成(cheng)(cheng)多個(ge)樹(shu)模(mo)(mo)(mo)型來建立一個(ge)更強的(de)(de)(de)(de)(de)學習(xi)模(mo)(mo)(mo)型[14]。包括XGBoost在(zai)內的(de)(de)(de)(de)(de)ML算(suan)(suan)法(fa)(fa)是(shi)一種通過(guo)自(zi)動分析(xi)數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)以(yi)獲(huo)得(de)特征規律(lv)(lv),并利用這(zhe)些規律(lv)(lv)對(dui)(dui)未知數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)進行(xing)(xing)預(yu)測的(de)(de)(de)(de)(de)算(suan)(suan)法(fa)(fa),是(shi)實現AI的(de)(de)(de)(de)(de)基(ji)本方法(fa)(fa)[15]。對(dui)(dui)于(yu)給定的(de)(de)(de)(de)(de)任(ren)務,ML算(suan)(suan)法(fa)(fa)可(ke)(ke)以(yi)從數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)中(zhong)學習(xi),篩選大(da)量變量并尋找可(ke)(ke)靠的(de)(de)(de)(de)(de)組合預(yu)測結果,所(suo)以(yi)其對(dui)(dui)數(shu)(shu)(shu)(shu)(shu)(shu)(shu)量龐(pang)大(da)或復雜的(de)(de)(de)(de)(de)數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)進行(xing)(xing)挖掘分析(xi)具有絕對(dui)(dui)優勢。由(you)此,使(shi)(shi)(shi)用ML算(suan)(suan)法(fa)(fa)對(dui)(dui)LIS中(zhong)存(cun)儲(chu)的(de)(de)(de)(de)(de)數(shu)(shu)(shu)(shu)(shu)(shu)(shu)千萬患者信(xin)息進行(xing)(xing)挖掘、訓練、建模(mo)(mo)(mo)是(shi)完全可(ke)(ke)行(xing)(xing)的(de)(de)(de)(de)(de)。這(zhe)種方法(fa)(fa)不僅解(jie)決了傳統(tong)(tong)(tong)統(tong)(tong)(tong)計(ji)(ji)(ji)模(mo)(mo)(mo)型參數(shu)(shu)(shu)(shu)(shu)(shu)(shu)限制(zhi)的(de)(de)(de)(de)(de)問題(ti),還(huan)能(neng)減少人工誤差,提(ti)(ti)(ti)高(gao)篩選效(xiao)率(lv),更重要的(de)(de)(de)(de)(de)是(shi)有助于(yu)發現數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)間的(de)(de)(de)(de)(de)潛(qian)在(zai)關聯及隱藏價值。最后,算(suan)(suan)力(li)為檢(jian)驗(yan)(yan)(yan)AI提(ti)(ti)(ti)供(gong)(gong)了基(ji)本的(de)(de)(de)(de)(de)計(ji)(ji)(ji)算(suan)(suan)能(neng)力(li)支(zhi)撐。算(suan)(suan)力(li)的(de)(de)(de)(de)(de)核心在(zai)于(yu)AI芯片。由(you)于(yu)原(yuan)有的(de)(de)(de)(de)(de)CPU和(he)傳統(tong)(tong)(tong)計(ji)(ji)(ji)算(suan)(suan)架構已(yi)(yi)無法(fa)(fa)支(zhi)撐ML算(suan)(suan)法(fa)(fa)并行(xing)(xing)計(ji)(ji)(ji)算(suan)(suan)能(neng)力(li)的(de)(de)(de)(de)(de)要求,AI芯片的(de)(de)(de)(de)(de)出(chu)現使(shi)(shi)(shi)得(de)AI高(gao)強度、大(da)數(shu)(shu)(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)的(de)(de)(de)(de)(de)處理(li)需求得(de)以(yi)滿足。目前,針對(dui)(dui)AI領域(yu)運算(suan)(suan)特征,AI芯片已(yi)(yi)成(cheng)(cheng)功應(ying)用于(yu)圖像(xiang)識別、語音識別、聲紋(wen)檢(jian)測、視頻檢(jian)索及自(zi)動駕駛(shi)等(deng)任(ren)務。在(zai)做智能(neng)處理(li)任(ren)務時,AI芯片能(neng)夠使(shi)(shi)(shi)計(ji)(ji)(ji)算(suan)(suan)機(ji)以(yi)大(da)規模(mo)(mo)(mo)、低(di)成(cheng)(cheng)本、高(gao)通量、自(zi)動化等(deng)方式運行(xing)(xing)[16]。這(zhe)些均為檢(jian)驗(yan)(yan)(yan)AI的(de)(de)(de)(de)(de)實現和(he)應(ying)用提(ti)(ti)(ti)供(gong)(gong)了強大(da)的(de)(de)(de)(de)(de)硬件平臺。

2.2檢驗AI的構建過程

檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)只有(you)(you)(you)與(yu)(yu)AI深(shen)度(du)融合形成(cheng)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI,才(cai)能(neng)(neng)(neng)(neng)(neng)充分(fen)挖掘檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)指標(biao)(biao)與(yu)(yu)疾病之間隱藏的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)聯(lian)(lian)系(xi)與(yu)(yu)規律,為(wei)(wei)臨(lin)(lin)床提供更有(you)(you)(you)意(yi)義的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)診(zhen)(zhen)(zhen)療(liao)信(xin)(xin)息(xi),體現(xian)(xian)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)醫學(xue)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)價值(zhi)(zhi)。檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)構建(jian)(jian)(圖1)主(zhu)要(yao)(yao)包括(kuo)(kuo)以(yi)(yi)下步驟(1)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)收集(ji)(ji)(ji)(ji)和(he)(he)(he)預(yu)處理(li):檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)來(lai)(lai)源廣泛(fan),由于(yu)(yu)(yu)(yu)各數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)庫(LIS、HIS系(xi)統)之間缺乏(fa)有(you)(you)(you)效協(xie)調機(ji)制,聯(lian)(lian)系(xi)和(he)(he)(he)溝(gou)通(tong)(tong)不同(tong)(tong),采集(ji)(ji)(ji)(ji)和(he)(he)(he)存儲(chu)標(biao)(biao)準(zhun)(zhun)不一(yi)(yi)(yi),影響了(le)(le)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)應(ying)用(yong)(yong)(yong)(yong)(yong)(yong)[17]。因此,檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)在(zai)(zai)用(yong)(yong)(yong)(yong)(yong)(yong)于(yu)(yu)(yu)(yu)建(jian)(jian)模(mo)(mo)(mo)(mo)(mo)前需要(yao)(yao)經過(guo)預(yu)處理(li)(包括(kuo)(kuo)結(jie)構化、標(biao)(biao)準(zhun)(zhun)化、歸一(yi)(yi)(yi)化等)轉變(bian)(bian)成(cheng)ML算(suan)(suan)法(fa)(fa)(fa)能(neng)(neng)(neng)(neng)(neng)夠(gou)識(shi)別(bie)和(he)(he)(he)學(xue)習(xi)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)格式。模(mo)(mo)(mo)(mo)(mo)式識(shi)別(bie)作為(wei)(wei)AI技(ji)術(shu)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)重(zhong)(zhong)要(yao)(yao)分(fen)支,能(neng)(neng)(neng)(neng)(neng)夠(gou)將(jiang)多個(ge)標(biao)(biao)準(zhun)(zhun)來(lai)(lai)源數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)分(fen)類并結(jie)構化,節省大量(liang)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)人(ren)力(li)物力(li),解決了(le)(le)不同(tong)(tong)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)庫的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)互聯(lian)(lian)互通(tong)(tong)[18]。(2)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)知(zhi)識(shi)圖譜篩選:臨(lin)(lin)床醫生(sheng)(sheng)在(zai)(zai)疾病診(zhen)(zhen)(zhen)斷(duan)過(guo)程(cheng)中(zhong),一(yi)(yi)(yi)些(xie)疾病診(zhen)(zhen)(zhen)斷(duan)結(jie)果(guo)(guo)(guo)可(ke)(ke)能(neng)(neng)(neng)(neng)(neng)不依賴(lai)于(yu)(yu)(yu)(yu)異(yi)常(chang)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)結(jie)果(guo)(guo)(guo)而直接通(tong)(tong)過(guo)B超、放射或病理(li)等輔助檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)查結(jie)果(guo)(guo)(guo)得(de)出。由于(yu)(yu)(yu)(yu)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI是(shi)基于(yu)(yu)(yu)(yu)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)AI模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing),如果(guo)(guo)(guo)一(yi)(yi)(yi)些(xie)疾病診(zhen)(zhen)(zhen)斷(duan)結(jie)果(guo)(guo)(guo)不能(neng)(neng)(neng)(neng)(neng)通(tong)(tong)過(guo)相(xiang)(xiang)應(ying)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)異(yi)常(chang)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)結(jie)果(guo)(guo)(guo)推(tui)理(li)得(de)出,不僅增加了(le)(le)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)處理(li)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)難度(du),而且會降(jiang)低模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)預(yu)測(ce)(ce)(ce)疾病的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)準(zhun)(zhun)確率。通(tong)(tong)過(guo)人(ren)工鑒(jian)別(bie)異(yi)常(chang)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)結(jie)果(guo)(guo)(guo)是(shi)否與(yu)(yu)疾病診(zhen)(zhen)(zhen)斷(duan)結(jie)果(guo)(guo)(guo)相(xiang)(xiang)關,往(wang)(wang)往(wang)(wang)會出現(xian)(xian)其(qi)復雜程(cheng)度(du)超越了(le)(le)臨(lin)(lin)床醫生(sheng)(sheng)知(zhi)識(shi)范疇的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)情(qing)況。因此,通(tong)(tong)過(guo)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)知(zhi)識(shi)圖譜從醫學(xue)專業(ye)角(jiao)度(du)將(jiang)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)與(yu)(yu)診(zhen)(zhen)(zhen)斷(duan)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)相(xiang)(xiang)關聯(lian)(lian),排除了(le)(le)一(yi)(yi)(yi)些(xie)缺乏(fa)知(zhi)識(shi)關聯(lian)(lian)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)信(xin)(xin)息(xi),提高(gao)了(le)(le)輸(shu)入(ru)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)質(zhi)量(liang),同(tong)(tong)時也增強了(le)(le)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)學(xue)習(xi)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)能(neng)(neng)(neng)(neng)(neng)力(li)[19]。(3)特(te)(te)(te)征(zheng)選擇與(yu)(yu)建(jian)(jian)模(mo)(mo)(mo)(mo)(mo):在(zai)(zai)開(kai)發(fa)(fa)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI時,特(te)(te)(te)征(zheng)變(bian)(bian)量(liang)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)選擇非常(chang)重(zhong)(zhong)要(yao)(yao),直接影響模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)在(zai)(zai)臨(lin)(lin)床研(yan)究(jiu)中(zhong)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)表現(xian)(xian)。從特(te)(te)(te)征(zheng)自身角(jiao)度(du)選擇利用(yong)(yong)(yong)(yong)(yong)(yong)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)原(yuan)理(li)為(wei)(wei)重(zhong)(zhong)要(yao)(yao)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)特(te)(te)(te)征(zheng)應(ying)該(gai)是(shi)攜帶(dai)信(xin)(xin)息(xi)較多,也就(jiu)是(shi)特(te)(te)(te)征(zheng)值(zhi)(zhi)差異(yi)較大的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)變(bian)(bian)量(liang)。從輸(shu)入(ru)特(te)(te)(te)征(zheng)和(he)(he)(he)輸(shu)出特(te)(te)(te)征(zheng)相(xiang)(xiang)關性(xing)角(jiao)度(du)選擇主(zhu)要(yao)(yao)是(shi)利用(yong)(yong)(yong)(yong)(yong)(yong)一(yi)(yi)(yi)些(xie)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)算(suan)(suan)法(fa)(fa)(fa)本身的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)特(te)(te)(te)點,構造不同(tong)(tong)特(te)(te)(te)征(zheng)與(yu)(yu)目標(biao)(biao)特(te)(te)(te)征(zheng)之間的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)相(xiang)(xiang)關重(zhong)(zhong)要(yao)(yao)程(cheng)度(du),若重(zhong)(zhong)要(yao)(yao)性(xing)低的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)特(te)(te)(te)征(zheng)變(bian)(bian)量(liang)(如低于(yu)(yu)(yu)(yu)某閾值(zhi)(zhi))則可(ke)(ke)以(yi)(yi)拋棄。建(jian)(jian)模(mo)(mo)(mo)(mo)(mo)方法(fa)(fa)(fa)主(zhu)要(yao)(yao)包括(kuo)(kuo)監督(du)學(xue)習(xi)和(he)(he)(he)無(wu)監督(du)學(xue)習(xi)[20]。監督(du)學(xue)習(xi)是(shi)使用(yong)(yong)(yong)(yong)(yong)(yong)有(you)(you)(you)診(zhen)(zhen)(zhen)斷(duan)結(jie)果(guo)(guo)(guo)標(biao)(biao)注的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)訓(xun)練(lian)(lian)開(kai)發(fa)(fa)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing),之后,所得(de)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)用(yong)(yong)(yong)(yong)(yong)(yong)于(yu)(yu)(yu)(yu)預(yu)測(ce)(ce)(ce)或診(zhen)(zhen)(zhen)斷(duan)新數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)集(ji)(ji)(ji)(ji)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)結(jie)果(guo)(guo)(guo)變(bian)(bian)量(liang)。無(wu)監督(du)學(xue)習(xi)則僅使用(yong)(yong)(yong)(yong)(yong)(yong)無(wu)診(zhen)(zhen)(zhen)斷(duan)結(jie)果(guo)(guo)(guo)標(biao)(biao)注的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)輸(shu)入(ru)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)來(lai)(lai)建(jian)(jian)模(mo)(mo)(mo)(mo)(mo)。(4)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)評估(gu)(gu):模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)性(xing)能(neng)(neng)(neng)(neng)(neng)可(ke)(ke)以(yi)(yi)使用(yong)(yong)(yong)(yong)(yong)(yong)同(tong)(tong)一(yi)(yi)(yi)來(lai)(lai)源(內(nei)部(bu)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng))或獨立(li)來(lai)(lai)源(外部(bu)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng))數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)進(jin)行(xing)評估(gu)(gu)。交(jiao)(jiao)叉驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng)方法(fa)(fa)(fa)是(shi)一(yi)(yi)(yi)種準(zhun)(zhun)確的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)內(nei)部(bu)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng)方法(fa)(fa)(fa)[21]。交(jiao)(jiao)叉驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)最(zui)(zui)簡單形式是(shi)將(jiang)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)集(ji)(ji)(ji)(ji)劃(hua)分(fen)為(wei)(wei)訓(xun)練(lian)(lian)集(ji)(ji)(ji)(ji)(用(yong)(yong)(yong)(yong)(yong)(yong)于(yu)(yu)(yu)(yu)開(kai)發(fa)(fa)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing))和(he)(he)(he)測(ce)(ce)(ce)試(shi)(shi)(shi)集(ji)(ji)(ji)(ji)(用(yong)(yong)(yong)(yong)(yong)(yong)于(yu)(yu)(yu)(yu)評估(gu)(gu)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)性(xing)能(neng)(neng)(neng)(neng)(neng))。訓(xun)練(lian)(lian)集(ji)(ji)(ji)(ji)和(he)(he)(he)測(ce)(ce)(ce)試(shi)(shi)(shi)集(ji)(ji)(ji)(ji)之間沒有(you)(you)(you)最(zui)(zui)佳劃(hua)分(fen),例如研(yan)究(jiu)人(ren)員(yuan)習(xi)慣將(jiang)80%的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)研(yan)究(jiu)人(ren)群分(fen)配給訓(xun)練(lian)(lian)集(ji)(ji)(ji)(ji),其(qi)余20%分(fen)配給測(ce)(ce)(ce)試(shi)(shi)(shi)集(ji)(ji)(ji)(ji)。然(ran)而,這(zhe)(zhe)種方法(fa)(fa)(fa)可(ke)(ke)能(neng)(neng)(neng)(neng)(neng)會使結(jie)果(guo)(guo)(guo)產生(sheng)(sheng)偏差,使得(de)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)不具有(you)(you)(you)普(pu)適(shi)性(xing)。n倍交(jiao)(jiao)叉驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng)是(shi)一(yi)(yi)(yi)種替代方法(fa)(fa)(fa),在(zai)(zai)這(zhe)(zhe)種方法(fa)(fa)(fa)中(zhong),數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)集(ji)(ji)(ji)(ji)被劃(hua)分(fen)成(cheng)n個(ge)大小(xiao)相(xiang)(xiang)等的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)分(fen)區,執行(xing)多輪交(jiao)(jiao)叉驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng)(即每輪使用(yong)(yong)(yong)(yong)(yong)(yong)不同(tong)(tong)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)分(fen)區作為(wei)(wei)測(ce)(ce)(ce)試(shi)(shi)(shi)集(ji)(ji)(ji)(ji)),平均(jun)結(jie)果(guo)(guo)(guo)用(yong)(yong)(yong)(yong)(yong)(yong)于(yu)(yu)(yu)(yu)估(gu)(gu)計模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)性(xing)能(neng)(neng)(neng)(neng)(neng)。模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)外部(bu)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng)即用(yong)(yong)(yong)(yong)(yong)(yong)非建(jian)(jian)模(mo)(mo)(mo)(mo)(mo)數(shu)(shu)(shu)(shu)(shu)據(ju)(ju)(ju)(ju)(ju)集(ji)(ji)(ji)(ji)測(ce)(ce)(ce)試(shi)(shi)(shi)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)性(xing)能(neng)(neng)(neng)(neng)(neng),提高(gao)其(qi)推(tui)廣性(xing)以(yi)(yi)支持(chi)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)普(pu)適(shi)性(xing)。(5)臨(lin)(lin)床驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)證(zheng)(zheng):為(wei)(wei)了(le)(le)證(zheng)(zheng)實檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI具有(you)(you)(you)很(hen)高(gao)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)臨(lin)(lin)床應(ying)用(yong)(yong)(yong)(yong)(yong)(yong)價值(zhi)(zhi),即可(ke)(ke)以(yi)(yi)有(you)(you)(you)效輔助臨(lin)(lin)床醫生(sheng)(sheng)進(jin)行(xing)臨(lin)(lin)床決策,需要(yao)(yao)將(jiang)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI診(zhen)(zhen)(zhen)療(liao)性(xing)能(neng)(neng)(neng)(neng)(neng)與(yu)(yu)醫學(xue)專家進(jin)行(xing)比(bi)較。(6)模(mo)(mo)(mo)(mo)(mo)型(xing)(xing)(xing)應(ying)用(yong)(yong)(yong)(yong)(yong)(yong):為(wei)(wei)了(le)(le)將(jiang)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI運用(yong)(yong)(yong)(yong)(yong)(yong)到臨(lin)(lin)床實踐中(zhong),需要(yao)(yao)基于(yu)(yu)(yu)(yu)互聯(lian)(lian)網技(ji)術(shu)開(kai)發(fa)(fa)相(xiang)(xiang)應(ying)的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)應(ying)用(yong)(yong)(yong)(yong)(yong)(yong)程(cheng)序以(yi)(yi)方便臨(lin)(lin)床醫生(sheng)(sheng)使用(yong)(yong)(yong)(yong)(yong)(yong),使得(de)檢(jian)(jian)(jian)(jian)(jian)(jian)(jian)驗(yan)(yan)(yan)(yan)(yan)(yan)(yan)(yan)AI可(ke)(ke)以(yi)(yi)在(zai)(zai)真實場(chang)景運行(xing)以(yi)(yi)獲得(de)疾病的(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)(de)預(yu)測(ce)(ce)(ce)值(zhi)(zhi)等。

3檢驗AI在輔助診療中的應用前景

3.1疾病診斷

快速(su)而(er)準確(que)(que)的(de)(de)(de)(de)診(zhen)(zhen)斷(duan)對(dui)于指(zhi)(zhi)導(dao)患(huan)者(zhe)的(de)(de)(de)(de)診(zhen)(zhen)療(liao)活動至(zhi)關重要,檢(jian)(jian)驗(yan)(yan)AI基于檢(jian)(jian)驗(yan)(yan)數據(ju)可輔助臨床醫生進行疾(ji)(ji)(ji)病(bing)診(zhen)(zhen)斷(duan)。比如,GUNCAR等(deng)[22]使用了隨機森(sen)林ML算(suan)法并基于實驗(yan)(yan)室(shi)血(xue)液(ye)(ye)檢(jian)(jian)測(ce)(ce)結果,建(jian)立了兩種(zhong)模型(xing)用于血(xue)液(ye)(ye)疾(ji)(ji)(ji)病(bing)診(zhen)(zhen)斷(duan),一種(zhong)預(yu)測(ce)(ce)模型(xing)(SBA-HEM168)使用所有(you)可用的(de)(de)(de)(de)血(xue)液(ye)(ye)測(ce)(ce)試(shi)(shi)參數;另(ling)一種(zhong)預(yu)測(ce)(ce)模型(xing)(SBA-HEM061)僅使用一組簡化的(de)(de)(de)(de)血(xue)液(ye)(ye)測(ce)(ce)試(shi)(shi)參數。兩種(zhong)模型(xing)在診(zhen)(zhen)斷(duan)血(xue)液(ye)(ye)疾(ji)(ji)(ji)病(bing)上都(dou)具(ju)(ju)有(you)良好的(de)(de)(de)(de)診(zhen)(zhen)斷(duan)性能(neng),在預(yu)測(ce)(ce)患(huan)者(zhe)5種(zhong)最可能(neng)的(de)(de)(de)(de)疾(ji)(ji)(ji)病(bing)時,預(yu)測(ce)(ce)準確(que)(que)率(lv)分別為0.88和(he)(he)(he)0.86;僅考慮患(huan)者(zhe)最可能(neng)的(de)(de)(de)(de)疾(ji)(ji)(ji)病(bing)時,預(yu)測(ce)(ce)準確(que)(que)率(lv)為0.59和(he)(he)(he)0.57[22]。此外(wai),臨床試(shi)(shi)驗(yan)(yan)表明兩種(zhong)預(yu)測(ce)(ce)模型(xing)的(de)(de)(de)(de)準確(que)(que)性與血(xue)液(ye)(ye)學專家相當[23]。本課題組采用了730113例合格參與者(zhe)的(de)(de)(de)(de)89116949個(ge)(ge)檢(jian)(jian)驗(yan)(yan)數據(ju)和(he)(he)(he)10423581個(ge)(ge)實驗(yan)(yan)診(zhen)(zhen)斷(duan)數據(ju),構建(jian)了基于知識和(he)(he)(he)數據(ju)雙驅動的(de)(de)(de)(de)多功(gong)能(neng)智慧(hui)檢(jian)(jian)驗(yan)(yan)平臺[23](即檢(jian)(jian)驗(yan)(yan)AI),能(neng)夠(gou)自動識別并綜合分析2071種(zhong)檢(jian)(jian)驗(yan)(yan)指(zhi)(zhi)標,完成(cheng)對(dui)包(bao)括呼吸系(xi)統(tong)及血(xue)液(ye)(ye)系(xi)統(tong)等(deng)在內的(de)(de)(de)(de)10種(zhong)器官系(xi)統(tong)疾(ji)(ji)(ji)病(bing)和(he)(he)(he)441種(zhong)具(ju)(ju)體疾(ji)(ji)(ji)病(bing)診(zhen)(zhen)斷(duan)的(de)(de)(de)(de)多元(yuan)關系(xi)推(tui)(tui)理,且所有(you)推(tui)(tui)理過程具(ju)(ju)備良好的(de)(de)(de)(de)可解釋性。

3.2疾病預測

檢驗(yan)AI可(ke)深度(du)挖掘常(chang)規(gui)檢驗(yan)指(zhi)(zhi)標隱藏(zang)的(de)信息,能更好地預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)疾病。目前(qian),腫瘤的(de)預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)僅(jin)限于幾種特定的(de)標志物。PATRíCIO等[24]開發了(le)基于常(chang)規(gui)檢驗(yan)指(zhi)(zhi)標的(de)乳腺癌(ai)(ai)預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)模(mo)(mo)型(xing)(xing),研究人(ren)員運用了(le)邏輯(ji)回歸、隨機森林和(he)支持(chi)向(xiang)量機3種不同的(de)ML算法以不同數量的(de)變量(包括(kuo)葡萄(tao)糖(tang)、抵(di)抗(kang)素(su)(su)、年(nian)齡(ling)、體(ti)質量指(zhi)(zhi)數、胰島素(su)(su)抵(di)抗(kang)指(zhi)(zhi)數、瘦素(su)(su)、胰島素(su)(su)、脂聯(lian)素(su)(su)、單核細胞(bao)趨(qu)化(hua)(hua)蛋白-1)作為預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)因(yin)子(zi)構(gou)建了(le)乳腺癌(ai)(ai)預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)模(mo)(mo)型(xing)(xing)。該研究發現,支持(chi)向(xiang)量機預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)模(mo)(mo)型(xing)(xing)僅(jin)使(shi)(shi)用葡萄(tao)糖(tang)、抵(di)抗(kang)素(su)(su)、年(nian)齡(ling)和(he)體(ti)質量指(zhi)(zhi)數作為決定因(yin)素(su)(su)即(ji)可(ke)預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)女性(xing)(xing)乳腺癌(ai)(ai)的(de)存在(zai)。同時,該預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)模(mo)(mo)型(xing)(xing)在(zai)測(ce)(ce)(ce)(ce)(ce)試集上實現了(le)高(gao)靈(ling)敏度(du)(82%、88%)和(he)高(gao)特異(yi)度(du)(84%、90%)的(de)預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)[24]。該研究充分證實基于檢驗(yan)AI,一些常(chang)規(gui)檢測(ce)(ce)(ce)(ce)(ce)參數完全可(ke)以替代一些腫瘤標志物成為新(xin)的(de)癌(ai)(ai)癥早期篩查工(gong)具[24]。SOMNAY等[25]使(shi)(shi)用術前(qian)血(xue)清鈣、磷(lin)、甲(jia)狀(zhuang)旁(pang)(pang)腺激素(su)(su)、維(wei)生素(su)(su)D和(he)肌酐水(shui)平作為原發性(xing)(xing)甲(jia)狀(zhuang)旁(pang)(pang)腺功能亢(kang)進(jin)的(de)潛(qian)在(zai)預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)因(yin)子(zi)構(gou)建了(le)ML模(mo)(mo)型(xing)(xing),在(zai)所測(ce)(ce)(ce)(ce)(ce)試的(de)ML模(mo)(mo)型(xing)(xing)中(zhong),貝葉斯網絡(luo)模(mo)(mo)型(xing)(xing)被證明是(shi)最準確的(de),對95%的(de)原發性(xing)(xing)甲(jia)狀(zhuang)旁(pang)(pang)腺功能亢(kang)進(jin)患者進(jin)行了(le)正確分類(lei)[曲(qu)線下面(mian)積(ji)(AUC)=0.99],值得(de)注(zhu)意的(de)是(shi),從模(mo)(mo)型(xing)(xing)中(zhong)省略預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)因(yin)子(zi)甲(jia)狀(zhuang)旁(pang)(pang)腺激素(su)(su)并沒(mei)有明顯降低其準確性(xing)(xing)。該研究表明即(ji)使(shi)(shi)在(zai)患者出(chu)現輕微疾病病理變化(hua)(hua)的(de)情況下,ML模(mo)(mo)型(xing)(xing)仍可(ke)準確預(yu)(yu)(yu)(yu)測(ce)(ce)(ce)(ce)(ce)出(chu)原發性(xing)(xing)甲(jia)狀(zhuang)旁(pang)(pang)腺功能亢(kang)進(jin)癥[25]。

3.3疾病(bing)風險(xian)評估

檢(jian)驗AI可(ke)以根據(ju)檢(jian)驗數(shu)(shu)據(ju)對(dui)患者(zhe)疾(ji)(ji)病進行風(feng)險(xian)評(ping)(ping)估(gu),為(wei)臨(lin)床(chuang)醫生(sheng)改進治療(liao)策略(lve)做出預(yu)警。對(dui)于(yu)COVID-19,快速準(zhun)確(que)地(di)識別(bie)和(he)(he)評(ping)(ping)估(gu)疾(ji)(ji)病的(de)(de)(de)嚴(yan)重(zhong)(zhong)程度至關重(zhong)(zhong)要。MA等[26]收(shou)集了(le)COVID-19患者(zhe)入院(yuan)臨(lin)床(chuang)數(shu)(shu)據(ju),包括癥狀、合并(bing)癥、人口(kou)統計(ji)學、生(sheng)命體(ti)征、CT掃(sao)描(miao)結果及(ji)檢(jian)驗數(shu)(shu)據(ju),應(ying)用多(duo)元(yuan)邏輯回歸方法確(que)定具有統計(ji)學意義的(de)(de)(de)臨(lin)床(chuang)特征,研究人員最終使用了(le)乳酸脫氫酶、C反(fan)應(ying)蛋(dan)白(bai)和(he)(he)年齡作為(wei)死亡預(yu)測因素構建(jian)(jian)了(le)多(duo)元(yuan)邏輯回歸模(mo)型(xing),用于(yu)識別(bie)CO-VID-19入院(yuan)患者(zhe)的(de)(de)(de)嚴(yan)重(zhong)(zhong)程度并(bing)評(ping)(ping)估(gu)其(qi)死亡風(feng)險(xian)。該模(mo)型(xing)預(yu)測292例(li)患者(zhe)病死率的(de)(de)(de)效(xiao)能(AUC=0.9521)優于(yu)肺(fei)炎CURB-65評(ping)(ping)分(fen)模(mo)型(xing)(AUC=0.8501)和(he)(he)XGBoost模(mo)型(xing)(AUC=0.4530)。由數(shu)(shu)十個臨(lin)床(chuang)癥狀和(he)(he)檢(jian)驗數(shu)(shu)據(ju)(如靜脈(mo)(mo)曲(qu)張(zhang)(zhang)(zhang)程度、纖維蛋(dan)白(bai)原水平和(he)(he)紅斑等)作為(wei)患者(zhe)出血(xue)風(feng)險(xian)的(de)(de)(de)潛在(zai)決定因素所構建(jian)(jian)的(de)(de)(de)ML模(mo)型(xing)能夠預(yu)測肝功能受損兒童(tong)(tong)食管(guan)靜脈(mo)(mo)曲(qu)張(zhang)(zhang)(zhang)出血(xue)的(de)(de)(de)風(feng)險(xian),該模(mo)型(xing)能夠準(zhun)確(que)地(di)預(yu)測約(yue)85%兒童(tong)(tong)患者(zhe)食管(guan)靜脈(mo)(mo)曲(qu)張(zhang)(zhang)(zhang)破裂(lie)出血(xue)情況,使高危兒童(tong)(tong)患者(zhe)能夠優先進行緊急肝移(yi)植(zhi),該研究結果已(yi)通過在(zai)線應(ying)用程序(xu)(ht-tp://hrs2c2.com),在(zai)獨立患者(zhe)隊列中(zhong)得到了(le)驗證[27]。

3.4疾病治療與預后檢驗AI可根(gen)據(ju)檢驗數據(ju)評(ping)

估患(huan)(huan)(huan)者(zhe)的(de)(de)(de)(de)疾病(bing)(bing)治療(liao)及預(yu)后情況。比如,鑒(jian)于(yu)華法(fa)林的(de)(de)(de)(de)復(fu)雜特性(xing)(xing)(xing),LIU等(deng)[28]使用(yong)了(le)兩(liang)個(ge)經典的(de)(de)(de)(de)檢驗數據即丙氨(an)酸氨(an)基轉移酶及血(xue)清肌酐,結(jie)合患(huan)(huan)(huan)者(zhe)的(de)(de)(de)(de)華法(fa)林使用(yong)劑量、性(xing)(xing)(xing)別、年齡(ling)和(he)體(ti)質量等(deng)數據構建了(le)ML分(fen)(fen)類模(mo)(mo)型,用(yong)于(yu)預(yu)測(ce)(ce)(ce)華法(fa)林的(de)(de)(de)(de)治療(liao)是否充分(fen)(fen)。研(yan)究人員使用(yong)了(le)377例住(zhu)院患(huan)(huan)(huan)者(zhe)的(de)(de)(de)(de)臨(lin)床(chuang)數據,比較(jiao)了(le)7種不(bu)同ML分(fen)(fen)類模(mo)(mo)型的(de)(de)(de)(de)性(xing)(xing)(xing)能(neng),最終C4.5決(jue)策樹和(he)隨機森林得(de)分(fen)(fen)最高(gao),并發現相較(jiao)于(yu)臨(lin)床(chuang)醫生(sheng)的(de)(de)(de)(de)主(zhu)觀決(jue)定,ML模(mo)(mo)型更能(neng)準確(que)預(yu)測(ce)(ce)(ce)華法(fa)林治療(liao)的(de)(de)(de)(de)充分(fen)(fen)性(xing)(xing)(xing)[28]。盡管急(ji)性(xing)(xing)(xing)腎(shen)(shen)(shen)(shen)損(sun)傷通常(chang)是可(ke)逆的(de)(de)(de)(de),但(dan)一(yi)(yi)些(xie)患(huan)(huan)(huan)者(zhe)發生(sheng)腎(shen)(shen)(shen)(shen)功(gong)能(neng)恢復(fu)不(bu)全、慢(man)性(xing)(xing)(xing)腎(shen)(shen)(shen)(shen)病(bing)(bing),甚(shen)至腎(shen)(shen)(shen)(shen)功(gong)能(neng)衰竭的(de)(de)(de)(de)風(feng)險(xian)增加[29-30]。急(ji)性(xing)(xing)(xing)腎(shen)(shen)(shen)(shen)損(sun)傷患(huan)(huan)(huan)者(zhe)住(zhu)院治療(liao)后,對慢(man)性(xing)(xing)(xing)腎(shen)(shen)(shen)(shen)病(bing)(bing)的(de)(de)(de)(de)預(yu)測(ce)(ce)(ce)可(ke)使患(huan)(huan)(huan)者(zhe)獲(huo)得(de)早期(qi)干預(yu),并可(ke)能(neng)改善(shan)疾病(bing)(bing)預(yu)后,且可(ke)避免不(bu)必要的(de)(de)(de)(de)醫療(liao)資源使用(yong)。JAMES等(deng)[31]使用(yong)了(le)常(chang)規檢驗數據(包括基線血(xue)清肌氨(an)酸值、蛋(dan)白尿、急(ji)性(xing)(xing)(xing)腎(shen)(shen)(shen)(shen)損(sun)傷嚴重程度指標及出院血(xue)清肌氨(an)酸值)開發了(le)一(yi)(yi)個(ge)多(duo)變量風(feng)險(xian)分(fen)(fen)層模(mo)(mo)型,該模(mo)(mo)型能(neng)夠(gou)識(shi)別急(ji)性(xing)(xing)(xing)腎(shen)(shen)(shen)(shen)損(sun)傷住(zhu)院后可(ke)能(neng)患(huan)(huan)(huan)有慢(man)性(xing)(xing)(xing)腎(shen)(shen)(shen)(shen)病(bing)(bing)的(de)(de)(de)(de)高(gao)風(feng)險(xian)患(huan)(huan)(huan)者(zhe)。該模(mo)(mo)型有助(zhu)于(yu)患(huan)(huan)(huan)者(zhe)向門診醫療(liao)過(guo)渡期(qi)間,臨(lin)床(chuang)醫生(sheng)做出合理的(de)(de)(de)(de)預(yu)后評估、隨訪指導和(he)護理建議(yi)。

3.5亞健康的預測

生(sheng)理生(sheng)化指標量化檢測(ce)(ce)(ce)是(shi)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)檢測(ce)(ce)(ce)方(fang)法中較為客觀的(de)(de)(de)檢測(ce)(ce)(ce)方(fang)法,可最大(da)可能(neng)避免人(ren)(ren)為因(yin)素(su)影響(xiang),利(li)于(yu)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)的(de)(de)(de)準確(que)預(yu)測(ce)(ce)(ce)。目前(qian),針對亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)人(ren)(ren)群(qun)的(de)(de)(de)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)實驗(yan)室(shi)檢測(ce)(ce)(ce)參(can)考值(zhi)很難界(jie)定。處(chu)于(yu)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)狀(zhuang)態的(de)(de)(de)人(ren)(ren)雖然(ran)有諸多自覺不適癥(zheng)狀(zhuang),實驗(yan)室(shi)檢驗(yan)可能(neng)僅(jin)僅(jin)有某(mou)些指標的(de)(de)(de)變(bian)化。如何在正常值(zhi)范圍內劃分(fen)(fen)出(chu)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)人(ren)(ren)群(qun)和(he)(he)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)人(ren)(ren)群(qun)一直是(shi)研究(jiu)的(de)(de)(de)難點。檢驗(yan)AI的(de)(de)(de)出(chu)現(xian)為亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)的(de)(de)(de)預(yu)測(ce)(ce)(ce)及診斷指明了方(fang)向。以(yi)預(yu)定時間內收集的(de)(de)(de)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)患者與疾病(bing)人(ren)(ren)群(qun)的(de)(de)(de)檢驗(yan)數據(ju)為預(yu)測(ce)(ce)(ce)因(yin)素(su),構建(jian)(jian)區分(fen)(fen)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)人(ren)(ren)群(qun)和(he)(he)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)人(ren)(ren)群(qun)的(de)(de)(de)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)分(fen)(fen)類器,可實現(xian)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)的(de)(de)(de)快速甄(zhen)別。該分(fen)(fen)類器類似于(yu)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)狀(zhuang)態軌跡(ji),可提(ti)示(shi)機體(ti)是(shi)否(fou)處(chu)于(yu)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)狀(zhuang)態,以(yi)及亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)人(ren)(ren)群(qun)是(shi)否(fou)有疾病(bing)發生(sheng)的(de)(de)(de)可能(neng)[32]。比如,王嬙等[33]將(jiang)采(cai)集到的(de)(de)(de)出(chu)入境人(ren)(ren)員(yuan)臨床檢驗(yan)指標等數據(ju)錄入計算機,基于(yu)極(ji)限學習(xi)機和(he)(he)支持向量機方(fang)法建(jian)(jian)立(li)預(yu)測(ce)(ce)(ce)模型并自動分(fen)(fen)析錄入的(de)(de)(de)信(xin)息,從而預(yu)測(ce)(ce)(ce)該個體(ti)發生(sheng)亞(ya)(ya)(ya)(ya)健(jian)(jian)(jian)(jian)(jian)(jian)(jian)康(kang)(kang)的(de)(de)(de)趨勢及可能(neng)性。

4檢驗AI的局限性和挑戰

檢(jian)(jian)(jian)(jian)驗(yan)(yan)AI在臨床(chuang)(chuang)實際應用中(zhong),還存在一些局(ju)限(xian)(xian)性(xing)和挑戰。首先(xian),由于(yu)(yu)疾(ji)病(bing)(bing)診(zhen)斷的復雜性(xing),僅依(yi)靠檢(jian)(jian)(jian)(jian)驗(yan)(yan)AI提供的疾(ji)病(bing)(bing)預測(ce)(ce)或(huo)診(zhen)斷存在局(ju)限(xian)(xian)性(xing)[34],臨床(chuang)(chuang)醫生仍需結合其(qi)他輔助檢(jian)(jian)(jian)(jian)查(如影像、病(bing)(bing)理(li)及(ji)超聲等)結果(guo)(guo)對患者疾(ji)病(bing)(bing)進行(xing)綜(zong)合判斷。其(qi)次,數據(ju)(ju)質量問題是檢(jian)(jian)(jian)(jian)驗(yan)(yan)AI面(mian)臨的主(zhu)要挑戰。檢(jian)(jian)(jian)(jian)驗(yan)(yan)AI的疾(ji)病(bing)(bing)預測(ce)(ce)是基于(yu)(yu)檢(jian)(jian)(jian)(jian)驗(yan)(yan)大(da)數據(ju)(ju),由于(yu)(yu)檢(jian)(jian)(jian)(jian)驗(yan)(yan)大(da)數據(ju)(ju)的采(cai)集標準不(bu)一,導(dao)致(zhi)檢(jian)(jian)(jian)(jian)驗(yan)(yan)數據(ju)(ju)缺乏質量保證(zheng),影響了檢(jian)(jian)(jian)(jian)驗(yan)(yan)AI模型(xing)預測(ce)(ce)疾(ji)病(bing)(bing)的準確性(xing),進而(er)(er)限(xian)(xian)制(zhi)了檢(jian)(jian)(jian)(jian)驗(yan)(yan)AI的臨床(chuang)(chuang)實際應用[35]。最后,受(shou)限(xian)(xian)于(yu)(yu)不(bu)同(tong)醫院檢(jian)(jian)(jian)(jian)驗(yan)(yan)科對檢(jian)(jian)(jian)(jian)驗(yan)(yan)數據(ju)(ju)的定義(yi)、檢(jian)(jian)(jian)(jian)測(ce)(ce)標準及(ji)檢(jian)(jian)(jian)(jian)測(ce)(ce)系統不(bu)同(tong),導(dao)致(zhi)不(bu)同(tong)醫院之間檢(jian)(jian)(jian)(jian)驗(yan)(yan)數據(ju)(ju)難以(yi)互通互用,而(er)(er)依(yi)據(ju)(ju)單中(zhong)心檢(jian)(jian)(jian)(jian)驗(yan)(yan)數據(ju)(ju)構(gou)建的檢(jian)(jian)(jian)(jian)驗(yan)(yan)AI缺乏普適(shi)性(xing),其(qi)預測(ce)(ce)結果(guo)(guo)在不(bu)同(tong)醫院之間存在偏差。

5結語

檢(jian)驗(yan)(yan)醫學(xue)是臨床(chuang)輔助診療(liao)的重要支撐。隨(sui)著AI與醫學(xue)領域的深度(du)融合,檢(jian)驗(yan)(yan)醫學(xue)與AI的結合作為一種新(xin)的醫療(liao)模式(shi)正在(zai)快(kuai)速(su)發展。檢(jian)驗(yan)(yan)AI通過對檢(jian)驗(yan)(yan)大數(shu)據的分析,可為臨床(chuang)醫生提供(gong)及時準確的輔助診斷信(xin)息(xi),這將徹底改變(bian)傳統(tong)(tong)檢(jian)驗(yan)(yan)醫學(xue)的傳統(tong)(tong)運(yun)用模式(shi),推動檢(jian)驗(yan)(yan)醫學(xue)的智能(neng)化(hua)發展。

作者(zhe):井杰 綜述 黃曉春 單位:中國人民(min)解放(fang)軍海軍軍醫大學第一附屬醫院檢(jian)驗(yan)科(ke)