iindaba

IModeli yoLwimi olukhulu (LLM) inokubhala amanqaku acengayo asekelwe kumagama akhawulezayo, iphumelele iimviwo zobuchule bobuchule, kwaye ibhale ulwazi olunobubele novelwano ngesigulana. Nangona kunjalo, ukongeza kwimingcipheko eyaziwayo ye-fiction, fragility, kunye neenyaniso ezingachanekanga kwi-LLM, eminye imiba engasonjululwanga ngokuthe ngcembe ibe yinto ekugxilwe kuyo, njengemizekelo ye-AI equlethe "imilinganiselo yabantu" enokuba nocalucalulo ekudalweni nasekusetyenzisweni kwayo, kwaye nokuba i-LLM ayisayi kuqulunqa umxholo kwaye ishenxise iziphumo ezinobungozi ngokucacileyo, "amaxabiso e-LLM avela ebantwini" asenokuthi athobe.

Imizekelo emininzi ibonisa indlela idatha esetyenziswayo ekuqeqesheni imodeli ye-AI ifaka iikhowudi zomntu ngamnye kunye nezentlalo, ezinokuqiniswa ngaphakathi kwimodeli. Le mizekelo ibandakanya uluhlu lwezicelo, kubandakanywa ukutolika ngokuzenzekelayo kwe-X-rays yesifuba, ukuhlelwa kwezifo zesikhumba, kunye nokwenziwa kwesigqibo se-algorithmic malunga nokwabiwa kwezibonelelo zonyango. Njengoko kuchaziwe kwinqaku lakutshanje kwijenali yethu, idatha yoqeqesho enomkhethe inokukhulisa kwaye iveze iinqobo ezisemgangathweni kunye nokuthambekela okukhoyo eluntwini. Ngokuchasene noko, uphando luye lwabonisa ukuba i-AI ingasetyenziselwa ukunciphisa i-bias. Ngokomzekelo, abaphandi basebenzise iimodeli zokufunda ezinzulu kwiifilimu ze-X-ray zamadolo kwaye bafumanisa izinto eziye zaphoswa zizibonakaliso zobunzima obuqhelekileyo (zihlelwe ngama-radiologists) ngaphakathi kwedolo elihlangeneyo, ngaloo ndlela kuncitshiswa ukungafani kweentlungu ezingabonakaliyo phakathi kwezigulane ezimnyama nezimhlophe.

Nangona abantu abaninzi ngakumbi beqonda i-bias kwiimodeli ze-AI, ngokukodwa ngokubhekiselele kwidatha yoqeqesho, ezinye iindawo ezininzi zokungena kwimilinganiselo yabantu azinikwa ingqwalasela eyaneleyo ekuphuhliseni nasekuhanjisweni kweemodeli ze-AI. I-AI yezoNyango isandul 'ukufezekisa iziphumo ezinomdla, kodwa kwinqanaba elikhulu, ayizange iqwalasele ngokucacileyo imilinganiselo yabantu kunye nokusebenzisana kwabo kunye novavanyo lomngcipheko kunye nokuqiqa okunokwenzeka, kwaye akuzange kwenziwe imodeli.

Ukuqinisekisa ezi ngqikelelo zingabonakaliyo, khawufane ucinge ukuba uyingcali ye-endocrinologist ekufuneka ichaze ihomoni yokukhula komntu kwinkwenkwe eneminyaka esi-8 engaphantsi kwepesenti yesi-3 yobudala bayo. Inqanaba le-hormone yokukhula kwenkwenkwe yenkwenkwe lingaphantsi kwe-2 ng / mL (ixabiso lokubhekisela,> 10 ng / mL, ixabiso lereferensi kumazwe amaninzi angaphandle kwe-United States> 7 ng / mL), kunye ne-hormone yokukhula komntu i-coding gene ifumene ukuguqulwa okungaqhelekanga kokungasebenzi. Sikholelwa ukuba ukusetyenziswa kwonyango lwe-hormone yokukhula komntu kubonakala kwaye akunakuphikiswa kwesi silungiselelo seklinikhi.

Ukusetyenziswa kwonyango lwe-hormone yokukhula kwabantu kule meko ilandelayo kunokubangela ukungqubuzana: ukuphakama kwenkwenkwe eneminyaka eyi-14 kuye kwahlala kwipesenti ye-10 yeentanga zakhe, kunye nencopho ye-hormone yokukhula komntu emva kokuvuselela i-8 ng / mL. Akukho zinguqu zokusebenza ezaziwayo ezinokuchaphazela ubude, okanye ezinye izizathu ezaziwayo zesiqu esifutshane, kwaye ubudala bakhe bethambo buneminyaka eyi-15 ubudala (okt akukho kulibaziseka kophuhliso). Inxalenye kuphela yempikiswano kungenxa yomahluko kwimilinganiselo emiselwe ziingcali ezisekwe kwizifundo ezininzi malunga namanqanaba ehomoni yokukhula kwabantu esetyenziselwa ukufumanisa ukusilela kwehomoni yokukhula ekwanti. Ubuncinci impikiswano enkulu isuka kumngcipheko wenzuzo yokusebenzisa unyango lwe-hormone yokukhula kwabantu ngokwembono yezigulana, abazali abaguli, abaqeqeshi bezempilo, iinkampani zamayeza kunye nabahlawuli. Iingcali ze-endocrinologists zabantwana zinokulinganisa iziphumo ezingathandekiyo ezingathandekiyo zenaliti yemihla ngemihla ye-hormone yokukhula iminyaka eyi-2 kunye namathuba okungabikho okanye ukukhula okuncinci kuphela kubungakanani bomzimba womntu omdala xa kuthelekiswa nokhoyo. Amakhwenkwe angakholelwa ukuba nangona ubude babo bunokwenyuka kuphela nge-2 cm, kuyafaneleka ukujova i-hormone yokukhula, kodwa umhlawuli kunye nenkampani yamachiza inokubamba iimbono ezahlukeneyo.

Sithatha i-creatinine esekelwe kwi-eGFR njengomzekelo, esetyenziswa ngokubanzi kwisalathisi sokusebenza kweentso ekuxilongeni kunye nokubeka isifo sezintso ezingapheliyo, ukubeka ukutshintshwa kwezintso okanye iimeko zokunikela, kunye nokumisela iindlela zokunciphisa kunye nokuchasana kwamachiza amaninzi achazwe ngugqirha. I-EGFR yindlela elula yokubuyisela i-equation esetyenziselwa ukuqikelela izinga lokuhluza i-glomerular filtration (mGFR), kumgangatho wereferensi, kodwa indlela yokuvavanya inzima kakhulu. Le equation yokuhlehla ayinakuthathwa njengemodeli ye-AI, kodwa ibonisa imigaqo emininzi malunga neenqobo zobuntu kunye nokuqiqa okunokwenzeka.

Indawo yokuqala yokungena kumaxabiso omntu ukungena kwi-eGFR kuxa ukhetha idata yokulinganisa iiequations. Umgca wokuqala osetyenzisiweyo ukuyila ifomula ye-eGFR ubukhulu becala yenziwe ngabantu abamnyama nabamhlophe, kwaye ukusebenza kwayo kwezinye iintlanga ezininzi akucaci. Amanqaku alandelayo okungena kwiimpawu zobuntu kule fomyula abandakanya: ukukhetha ukuchaneka kwe-mGFR njengeyona njongo iphambili yokuvavanya ukusebenza kwezintso, yintoni inqanaba elamkelekileyo lokuchaneka, indlela yokulinganisa ukuchaneka, kunye nokusebenzisa i-eGFR njengomqobo wokuqalisa ukwenza izigqibo zeklinikhi (njengokumisela iimeko zokutshintshwa kwezintso okanye ukumisela imithi). Ekugqibeleni, xa ukhetha umxholo wemodeli yokufaka, ixabiso lomntu liya kungena kwakhona kule fomula.

Umzekelo, ngaphambi kowama-2021, izikhokelo zicebisa ukulungelelanisa amanqanaba e-creatinine kwifomula ye-eGFR esekwe kwiminyaka yesigulana, isini, kunye nobuhlanga (abahlelwa kuphela njengabantu abamnyama okanye abangengabo abamnyama). Uhlengahlengiso olusekwe kugqatso lujolise ekuphuculeni ukuchaneka kwefomula ye-mGFR, kodwa ngo-2020, izibhedlele ezinkulu zaqala ukuthandabuza ukusetyenziswa kwe-eGFR esekelwe kubuhlanga, bechaza izizathu ezinjengokulibazisa ukufaneleka kwesigulane sokutshintshwa kunye nohlanga lwe-concretizing njengengcamango yebhayoloji. Uphando lubonise ukuba ukuyila iimodeli ze-eGFR ngokobuhlanga kunokuba neempembelelo ezinzulu kunye nezahlukeneyo zokuchaneka kunye neziphumo zonyango; Ke ngoko, ukujolisa ngokukhethekayo ekuchanekeni okanye ukugxila kwinxalenye yeziphumo kubonakalisa izigwebo zexabiso kwaye kunokufihla ukuthathwa kwezigqibo elubala. Okokugqibela, iqela elisebenzayo lesizwe licebise ngefomula entsha eyathi ifakwe kwakhona ngaphandle kokuthathela ingqalelo uhlanga ukulungelelanisa ukusebenza kunye nemiba yobulungisa. Lo mzekelo ubonisa ukuba nefomula yeklinikhi elula inamanqaku amaninzi okungena kwiindlela zobuntu.

Xa kuthelekiswa neefomula zeklinikhi ezinenani elincinci lezalathi eziqikelelwayo, i-LLM inokubandakanya iibhiliyoni ukuya kumakhulu eebhiliyoni zeeparamitha (ubunzima bemodeli) okanye ngaphezulu, okwenza kube nzima ukuyiqonda. Isizathu sokuba sisithi “kunzima ukuqonda” kukuba kuninzi lwee-LLMs, eyona ndlela ichanekileyo yokufumana iimpendulo ngokubuza ayinakwenziwa imephu. Inani leeparamitha ze-GPT-4 alikabhengezwa okwangoku; I-GPT-3 yangaphambili yayine-175 yeebhiliyoni zeeparamitha. Iiparamitha ezingaphezulu azithethi ukuba namandla anamandla, njengemizekelo emincinci ebandakanya imijikelo engaphezulu yokubala (efana ne-LLaMA [iModeli yoLwimi olukhulu lweMeta AI] uthotho lwemodeli) okanye iimodeli ezilungelelaniswe ngokucokisekileyo ngokusekelwe kwingxelo yomntu ziya kwenza ngcono kuneemodeli ezinkulu. Ngokomzekelo, ngokumalunga nabavavanyi babantu, imodeli ye-InstrumentGPT (imodeli ene-1.3 yeebhiliyoni zeeparamitha) idlula i-GPT-3 ekuphuculeni iziphumo zemveliso yemodeli.

Iinkcukacha zoqeqesho ezikhethekileyo ze-GPT-4 azikachazwa okwangoku, kodwa iinkcukacha zemifuziselo yesizukulwana sangaphambili kuquka i-GPT-3, i-InstrumentGPT, kunye nezinye ii-LLM ezininzi ezivulelekileyo ziye zabhengezwa. Kule mihla, iimodeli ezininzi ze-AI ziza namakhadi emodeli; Idatha yokuvavanya kunye nokhuseleko lwe-GPT-4 ipapashwe kwikhadi lenkqubo efanayo enikezelwa yinkampani yokudala imodeli ye-OpenAI. Ukwenziwa kwe-LLM kunokwahlulwa ngokwezigaba zibe zimbini: inqanaba lokuqala loqeqesho lwangaphambili kunye nenqanaba lokulungisa kakuhle elijolise ekwandiseni iziphumo zemveliso yemodeli. Kwinqanaba loqeqesho lwangaphambili, imodeli ibonelelwa ngecorpus enkulu ebandakanya umbhalo wokuqala we-Intanethi ukuyiqeqeshela ukuqikelela igama elilandelayo. Le nkqubo ibonakala ilula "yokugqiba ngokuzenzekelayo" ivelisa imodeli yesiseko esinamandla, kodwa inokukhokelela ekuziphatheni okuyingozi. Iimpawu zobuntu ziya kungena kwisigaba soqeqesho lwangaphambili, kubandakanywa nokukhetha idatha yoqeqesho lwangaphambili lwe-GPT-4 kunye nokwenza isigqibo sokususa umxholo ongafanelekanga onjengomxholo wephonografi kwidatha yoqeqesho lwangaphambili. Ngaphandle kwezi nzame, imodeli esisiseko isenokungabi luncedo okanye ingabinako ukuqulatha iziphumo ezinobungozi. Kwinqanaba elilandelayo lokulungiswa kakuhle, iindlela ezininzi zokuziphatha eziluncedo nezingenabungozi ziya kuvela.

Kwinqanaba lohlengahlengiso, indlela yokuziphatha yeemodeli zolwimi isoloko iguqulwa ngokunzulu ngokulungiswa okucokisekileyo kunye nokufunda okomeleza ngokusekelwe kwingxelo yomntu. Kwinqanaba lokulungiswa kakuhle, abasebenzi bekontraka abaqeshiweyo baya kubhala imizekelo yokuphendula kumagama akhawulezayo kwaye baqeqeshe ngokuthe ngqo imodeli. Kwinqanaba lokufunda lokuqiniswa ngokusekelwe kwingxelo yomntu, abaphononongi abangabantu baya kuhlela iziphumo zemveliso yemodeli njengemizekelo yomxholo wegalelo. Emva koko sebenzisa ezi ziphumo zothelekiso zingentla ukuze ufunde "imodeli yomvuzo" kwaye uphucule ngakumbi imodeli ngokufunda okomeleza. Ukubandakanyeka kwabantu kwinqanaba elisezantsi elimangalisayo kunokuphucula le modeli mikhulu. Umzekelo, imodeli ye-InstrumentGPT isebenzise iqela labasebenzi bekontraka abamalunga nama-40 abaqeshwe kwiiwebhusayithi ze-crowdsourcing kwaye baphumelele uvavanyo lokuhlola olujolise ekukhetheni iqela labachasi abanovelwano kwizinto ezikhethwa ngamaqela abantu ahlukeneyo.

Njengoko le mizekelo mibini igqithisileyo, eyifomula yekliniki elula [i-eGFR] kunye ne-LLM enamandla [GPT-4], ibonisa, ukwenza izigqibo zabantu kunye nemilinganiselo yabantu idlala indima ebaluleke kakhulu ekubumbeni iziphumo zeziphumo zemodeli. Ngaba ezi modeli ze-AI zingabamba amaxabiso azo ahlukeneyo ezigulane kunye noogqirha? Indlela yokukhokela esidlangalaleni ukusetyenziswa kwe-AI kumayeza? Njengoko kukhankanyiwe ngezantsi, ukuphononongwa kwakhona kohlalutyo lwesigqibo sezonyango kunokubonelela ngesisombululo esisisiseko kule miba.

Uhlalutyo lwesigqibo sezonyango aluqhelekanga kwiikliniki ezininzi, kodwa lunokwahlula phakathi kokuqiqa okunokwenzeka (ngenxa yeziphumo ezingaqinisekanga ezinxulumene nokwenza izigqibo, njengokuba ngaba ukulawula i-hormone yokukhula komntu kwimeko yeklinikhi ephikisanayo eboniswe kuMzobo 1) kunye nezinto zokuqwalaselwa (kumaxabiso axhomekeke kwezi ziphumo, ixabiso labo lilinganiswe njenge "ixabiso le-cm ye-utility", ukubonelela ngexabiso le-cm yendlela yokuphakama kwe-2. kwizigqibo ezinzima zonyango. Ekuhlalutyweni kwesigqibo, oogqirha kufuneka baqale banqume zonke izigqibo ezinokwenzeka kunye nezinto ezinokwenzeka ezinxulumene nesiphumo ngasinye, kwaye ke zibandakanya isigulane (okanye elinye iqela) into eluncedo ehambelana nesiphumo ngasinye ukukhetha eyona ndlela ifanelekileyo. Ngoko ke, ukuqinisekiswa kohlalutyo lwesigqibo kuxhomekeke ekubeni ngaba ukusetwa kwesiphumo kubanzi, kunye nokuba umlinganiselo wokusetyenziswa kunye noqikelelo lwamathuba achanekile. Ngokufanelekileyo, le ndlela inceda ekuqinisekiseni ukuba izigqibo zisekelwe kubungqina kwaye zihambelana nokukhethwa kwesigulane, ngaloo ndlela inciphisa i-gap phakathi kwedatha yenjongo kunye nemilinganiselo yomntu. Le ndlela yaziswa kwicandelo lezonyango kumashumi eminyaka eyadlulayo kwaye yasetyenziswa ekuthatheni izigqibo kwesigulana ngasinye kunye novavanyo lwempilo yabemi, njengokubonelela ngeengcebiso zovavanyo lomhlaza wesikhumba kubemi ngokubanzi.

Kuhlalutyo lwesigqibo sonyango, iindlela ezahlukeneyo ziye zaphuhliswa ukufumana uncedo. Uninzi lweendlela zemveli ngokuthe ngqo zifumana ixabiso kwisigulane ngasinye. Indlela elula kakhulu kukusebenzisa umlinganiselo wokulinganisa, apho izigulane zivavanya inqanaba labo lokukhetha kwisiphumo esithile kwi-digital scale (njenge-linear scale ukusuka kwi-1 ukuya kwi-10), kunye neziphumo ezibi kakhulu zempilo (ezifana nempilo epheleleyo kunye nokufa) ezibekwe kuzo zombini iziphelo. Indlela yokutshintsha ixesha yenye indlela esetyenziswa ngokuqhelekileyo. Kule ndlela, izigulane kufuneka zenze isigqibo malunga nokuba lingakanani ixesha eliphilileyo elizimisele ukulichitha ngokutshintshisa ixesha lempilo enkenenkene. Indlela yokungcakaza eqhelekileyo yenye indlela esetyenziswa ngokuqhelekileyo yokumisela into eluncedo. Kule ndlela, izigulane zibuzwa ukuba zeziphi ezimbini ezikhethiweyo ezikhethayo: nokuba ziphila inani elithile leminyaka kwimpilo eqhelekileyo kunye nethuba elithile (p) (t), kwaye zithwala umngcipheko wokufa nge-1-p amathuba; Qiniseka ukuba uphila iminyaka emithandathu phantsi kweemeko zempilo. Buza izigulane izihlandlo ezininzi kwii-p-values ​​ezahlukeneyo de zibonise ukungakhethi naluphi na ukhetho, ukuze uncedo lunokubalwa ngokusekelwe kwiimpendulo zesigulane.
Ukongeza kwiindlela ezisetyenziselwa ukukhangela ukhetho lwesigulana ngasinye, iindlela ziye zaphuhliswa ukufumana into eluncedo kubantu bezigulane. Ingakumbi iingxoxo zamaqela ekugxilwe kuzo (ukuzisa izigulane kunye ukuze zixoxe ngamava athile) zinokunceda ukuqonda iimbono zazo. Ukuze kuhlanganiswe ngokukuko ukusetyenziswa kweqela, iindlela ezahlukeneyo zokuxoxa zeqela ezicwangcisiweyo ziye zacetywa.
Enyanisweni, ukuqaliswa ngokuthe ngqo kwe-utility ekuxilongweni kweklinikhi kunye nenkqubo yonyango kudla ixesha elide. Njengesisombululo, iikhweshine zophando zihlala zisasazwa kubantu abakhethiweyo ngokungenamkhethe ukufumana amanqaku aluncedo kumgangatho wabemi. Eminye imizekelo ibandakanya i-questionnaire ye-EuroQol 5-dimensional, i-6-dimensional utility weight short form, i-Health Utility Index, kunye neCancer Specific European Cancer Research and Treatment Organization Quality of Life Questionnaire Core 30 isixhobo.