(1.19)綜上所述，結(jié)合醫(yī)療診斷的實(shí)際需求，本文在后續(xù)章節(jié)所進(jìn)行的醫(yī)療數(shù)據(jù)分類(lèi)實(shí)驗(yàn)中將以AUC、Recall、F1、G-mean以及MCC為主要的性能度量指標(biāo)。1.1.3參數(shù)設(shè)置首先，對(duì)代價(jià)敏感深度學(xué)習(xí)方法中的參數(shù)在區(qū)間上進(jìn)行取值，根據(jù)在不同大小的數(shù)據(jù)集上的實(shí)驗(yàn)結(jié)果及其變化設(shè)置參數(shù)的取值范圍。圖1.5給出了隨著取值的變化，CS-SmDAE在All-cause、T-carer、RA、Diabetic這四個(gè)數(shù)據(jù)集上的實(shí)驗(yàn)結(jié)果。圖1.5參數(shù)敏感性分析從圖1.5中可以發(fā)現(xiàn)隨著超參數(shù)在區(qū)間上逐漸增大，各個(gè)數(shù)據(jù)集上的各種評(píng)價(jià)指標(biāo)都發(fā)生了較為明顯的改變。總體而言，當(dāng)時(shí)，CS-SmDAE在各個(gè)數(shù)據(jù)集上基本都能達(dá)到一個(gè)較好的水平。鑒于此，本文在后續(xù)的對(duì)比實(shí)驗(yàn)中將超參數(shù)設(shè)置為0.2。1.1.4對(duì)比實(shí)驗(yàn)及分析為了驗(yàn)證本文設(shè)計(jì)的CS-SmDAE模型的有效性，首先分別在RA（IR=5.14），Diabetic（IR=10.36）和Hcc（IR=16.11）這三個(gè)不平衡率跨度較為明顯的醫(yī)療數(shù)據(jù)集上，將CS-SmDAE與原始的SmDAE進(jìn)行了對(duì)比實(shí)驗(yàn)。實(shí)驗(yàn)設(shè)置中，除了softmax分類(lèi)器和交叉熵?fù)p失函數(shù)之外，SmDAE網(wǎng)絡(luò)中其它部分均與CS-SmDAE模型保持一致。兩個(gè)模型在各數(shù)據(jù)集上的對(duì)比實(shí)驗(yàn)結(jié)果如圖1.6所示。（a）RA數(shù)據(jù)集對(duì)比結(jié)果（b）Diabetic數(shù)據(jù)集對(duì)比結(jié)果（c）Hcc數(shù)據(jù)集對(duì)比結(jié)果圖1.6不同IR的數(shù)據(jù)集下的實(shí)驗(yàn)對(duì)比結(jié)果從圖1.6中可以得出：與SmDAE算法相比CS-SmDAE在不平衡醫(yī)療數(shù)據(jù)集上的分類(lèi)效果更好，更值得一提的是，隨著實(shí)驗(yàn)數(shù)據(jù)集的不平衡率的增大，CS-SmDAE提升效果越明顯。接下來(lái)，為了評(píng)估本章模型的綜合性能，我們將CS-SmDAE與常見(jiàn)的處理不平衡數(shù)據(jù)分類(lèi)問(wèn)題的算法進(jìn)行對(duì)比實(shí)驗(yàn)，用于性能對(duì)比實(shí)驗(yàn)的對(duì)比算法如下：（1）SMOTE-DNNADDINEN.CITE<EndNote><Cite><Author>Roumani</Author><Year>2018</Year><RecNum>309</RecNum><DisplayText>[63]</DisplayText><record><rec-number>309</rec-number><foreign-keys><keyapp="EN"db-id="x0r9955slp0fvnev95spxd0qwrsa9xz2rt2z">309</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>Roumani,YazanF.</author><author>Roumani,Yaman</author><author>Nwankpa,JosephK.</author><author>Tanniru,Mohan</author></authors></contributors><titles><title>Classifyingreadmissionstoacardiacintensivecareunit</title><secondary-title>AnnalsofOperationsResearch</secondary-title></titles><periodical><full-title>AnnalsofOperationsResearch</full-title></periodical><volume>263</volume><dates><year>2018</year></dates><urls></urls></record></Cite></EndNote>[\o"Roumani,2018#309"63]。SMOTE-DNN從數(shù)據(jù)的角度出發(fā)，采用過(guò)采樣方法來(lái)平衡各類(lèi)數(shù)目，然后結(jié)合傳統(tǒng)的深度神經(jīng)網(wǎng)絡(luò)模型進(jìn)行訓(xùn)練，以改善分類(lèi)性能。（2）CSMLPADDINEN.CITE<EndNote><Cite><Author>L</Author><Year>2013</Year><RecNum>310</RecNum><DisplayText>[64]</DisplayText><record><rec-number>310</rec-number><foreign-keys><keyapp="EN"db-id="x0r9955slp0fvnev95spxd0qwrsa9xz2rt2z">310</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>CastroCristianoL</author><author>BragaAnt?nioP</author></authors></contributors><titles><title>Novelcost-sensitiveapproachtoimprovethemultilayerperceptronperformanceonimbalanceddata</title><secondary-title>IEEEtransactionsonneuralnetworksandlearningsystems</secondary-title></titles><periodical><full-title>IEEEtransactionsonneuralnetworksandlearningsystems</full-title></periodical><volume>24</volume><number>6</number><dates><year>2013</year></dates><urls></urls><remote-database-provider>Cnki</remote-database-provider></record></Cite></EndNote>[\o"L,2013#310"64]。CSMLP基于聯(lián)合目標(biāo)函數(shù)，該函數(shù)使用單個(gè)成本參數(shù)來(lái)區(qū)分分類(lèi)錯(cuò)誤的重要性，以提高M(jìn)LP對(duì)于兩類(lèi)的辨別能力。（3）CSDNNADDINEN.CITE<EndNote><Cite><Author>Wang</Author><Year>2018</Year><RecNum>311</RecNum><DisplayText>[65]</DisplayText><record><rec-number>311</rec-number><foreign-keys><keyapp="EN"db-id="x0r9955slp0fvnev95spxd0qwrsa9xz2rt2z">311</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>HaishuaiWang</author><author>ZhichengCui</author><author>YixinChen</author><author>MichaelAvidan</author><author>ArbiBenAbdallah</author><author>AlexanderKronzer</author></authors></contributors><titles><title>PredictingHospitalReadmissionviaCost-SensitiveDeepLearning</title><secondary-title>IEEE/ACMTransactionsonComputationalBiologyandBioinformatics(TCBB)</secondary-title></titles><periodical><full-title>IEEE/ACMTransactionsonComputationalBiologyandBioinformatics(TCBB)</full-title></periodical><volume>15</volume><number>6</number><dates><year>2018</year></dates><isbn>1545-5963</isbn><urls></urls><remote-database-provider>Cnki</remote-database-provider></record></Cite></EndNote>[\o"Wang,2018#311"65]。CSDNN使用一個(gè)修正的損失函數(shù)，在深度學(xué)習(xí)框架中嵌入成本信息。首先對(duì)每種方法的參數(shù)進(jìn)行調(diào)整，并按照文獻(xiàn)ADDINEN.CITE<EndNote><Cite><Author>Jie</Author><Year>2020</Year><RecNum>312</RecNum><DisplayText>[66]</DisplayText><record><rec-number>312</rec-number><foreign-keys><keyapp="EN"db-id="x0r9955slp0fvnev95spxd0qwrsa9xz2rt2z">312</key></foreign-keys><ref-typename="JournalArticle">17</ref-type><contributors><authors><author>ZhouJie</author><author>JiangZhibin</author><author>WangShitong</author></authors></contributors><titles><title>Laplacianleastlearningmachinewithdynamicupdatingforimbalancedclassification</title><secondary-title>AppliedSoftComputing</secondary-title></titles><periodical><full-title>AppliedSoftComputing</full-title></periodical><volume>88</volume><number>C</number><dates><year>2020</year></dates><isbn>1568-4946</isbn><urls></urls><remote-database-provider>Cnki</remote-database-provider></record></Cite></EndNote>[\o"Jie,2020#312"66]的建議將每個(gè)數(shù)據(jù)集隨機(jī)分為訓(xùn)練（60%）和測(cè)試（40%）部分。然后在1.1.1節(jié)中所介紹的5個(gè)不平衡的醫(yī)療數(shù)據(jù)集上分別用AUC、F1、G-mean和MCC這四個(gè)指標(biāo)進(jìn)行了評(píng)估。在各個(gè)模型上利用交叉驗(yàn)證并求取平均結(jié)果的方法來(lái)降低隨機(jī)影響，其中各數(shù)據(jù)集上最好的結(jié)果用黑體突出顯示。表1.5AUC指標(biāo)對(duì)比結(jié)果數(shù)據(jù)集SMOTE-DNNCSMLPCSDNNCS-SmDAEAll-cause0.7980.6470.7650.817Hcc0.6480.5960.6450.668T-carer0.7030.5060.6820.733RA0.8740.9440.8960.912Diabetic0.8020.5480.7750.832表1.6F1指標(biāo)對(duì)比結(jié)果數(shù)據(jù)集SMOTE-DNNCSMLPCSDNNCS-SmDAEAll-cause0.7460.6250.7550.788Hcc0.5690.5030.5450.574T-carer0.3930.7150.6750.652RA0.8820.8180.8510.881Diabetic0.8260.5190.8280.842表1.7G-mean指標(biāo)對(duì)比結(jié)果數(shù)據(jù)集SMOTE-DNNCSMLPCSDNNCS-SmDAEAll-cause0.7720.5350.7190.797Hcc0.4290.4640.4320.516T-carer0.7030.5090.6330.752RA0.8740.8430.8640.881Diabetic0.8020.3540.7710.835表1.8MCC指標(biāo)對(duì)比結(jié)果數(shù)據(jù)集SMOTE-DNNCSMLPCSDNNCS-SmDAEAll-cause0.3930.6390.6370.722Hcc0.3180.4250.4330.479T-carer0.3360.4730.5160.546RA0.7290.8