izindaba

Imodeli Yolimi Olukhulu (LLM) ingabhala izindatshana ezihehayo ezisuselwe emagameni asheshayo, iphumelele izivivinyo zobuchwepheshe, futhi ibhale ulwazi lwesiguli olunobungane noluzwelayo. Kodwa-ke, ngaphezu kwezingozi ezaziwayo zezinganekwane, ubuthakathaka, namaqiniso anganembile ku-LLM, ezinye izinkinga ezingaxazululiwe ziya ngokuya zigxila, njengamamodeli e-AI aqukethe "amagugu omuntu" angase abandlulule ekudaleni nasekusetshenzisweni kwawo, futhi ngisho noma i-LLM ingasaqambi okuqukethwe futhi isuse imiphumela eyingozi ngokusobala, "amanani e-LLM" angase ahlehle.

Izibonelo ezingenakubalwa zibonisa ukuthi idatha esetshenziswa ukuqeqesha amamodeli e-AI ifaka amakhodi amanani angawodwana nawezenhlalo, angase aqine ngaphakathi kwemodeli. Lezi zibonelo zibandakanya uhla lwezinhlelo zokusebenza, okuhlanganisa ukutolika okuzenzakalelayo kwe-X-ray yesifuba, ukuhlukaniswa kwezifo zesikhumba, kanye nokwenziwa kwezinqumo nge-algorithmic mayelana nokwabiwa kwensiza yezokwelapha. Njengoba kushiwo esihlokweni sakamuva kujenali yethu, idatha yokuqeqeshwa echemile ingase ikhulise futhi iveze izindinganiso nokuchema okukhona emphakathini. Ngokuphambene nalokho, ucwaningo luye lwabonisa nokuthi i-AI ingasetshenziswa ukunciphisa ukuchema. Isibonelo, abacwaningi basebenzise amamodeli okufunda okujulile kumafilimu e-X-ray emadolweni futhi bathola izici eziphuthelwe izinkomba zokuqina ezijwayelekile (ezihlelwe yi-radiologists) ngaphakathi kwamadolo, ngaleyo ndlela kuncishiswe ukungezwani kobuhlungu obungachazeki phakathi kweziguli ezimnyama nezimhlophe.

Nakuba abantu abaningi ngokwengeziwe beqaphela ukuchema kumamodeli we-AI, ikakhulukazi mayelana nedatha yokuqeqeshwa, amanye amaphuzu amaningi okungena ezindinganiso zomuntu awanikezwa ukunakwa okwanele ekuthuthukisweni nasekusetshenzisweni kwamamodeli we-AI. I-Medical AI isanda kuthola imiphumela emangazayo, kodwa ngokwezinga elikhulu, ayizange icabangele ngokusobala izindinganiso zomuntu kanye nokusebenzisana kwabo nokuhlola ubungozi kanye nokucabanga okunokwenzeka, futhi akuzange kumodelwe.

Ukuze uqinise le mibono engabonakali, cabanga ukuthi ungudokotela we-endocrinologist okudingeka ukuthi unikeze i-hormone yokukhula yomuntu ehlanganiswe nomfana oneminyaka engu-8 ongaphansi kwephesenti lesithathu leminyaka yakhe. Izinga le-hormone yokukhula komuntu eligqugquzelwe lomfana lingaphansi kuka-2 ng/mL (inani eliyireferensi,>10 ng/mL, inani lereferensi lamazwe amaningi angaphandle kwe-United States lingu>7 ng/mL), futhi isakhi sekhodi se-hormone yokukhula komuntu sithole ukuguqulwa okungajwayelekile kokungasebenzi. Sikholelwa ukuthi ukusetshenziswa kokwelashwa ngama-hormone okukhula komuntu kusobala futhi akunakuphikiswa kulesi silungiselelo somtholampilo.

Ukusetshenziswa kokwelashwa kwe-hormone yokukhula komuntu kulezi zimo ezilandelayo kungabangela ingxabano: ukuphakama komfana oneminyaka engu-14 bekulokhu ku-percentile ye-10 yontanga yakhe, futhi inani eliphakeme le-hormone yokukhula komuntu ngemva kokukhuthazwa yi-8 ng / mL. Azikho izinguquko zokusebenza ezaziwayo ezingathinta ubude, noma ezinye izimbangela ezaziwayo zesiqu esifushane, futhi ubudala bakhe bethambo buyiminyaka eyi-15 (okungukuthi akukho ukubambezeleka kokuthuthuka). Ingxenye kuphela yempikiswano ingenxa yomehluko embundwini amanani anqunywa ochwepheshe ngokusekelwe ezifundweni eziningi eziphathelene namazinga e-hormone okukhula komuntu asetshenziselwa ukuhlonza ukuntula kwe-hormone yokukhula ehlukanisiwe. Okungenani impikiswano eningi isuka ebhalansini yenzuzo yengozi yokusebenzisa ukwelashwa kwe-hormone okukhula ngokombono weziguli, abazali abanesineke, ochwepheshe bezokunakekelwa kwempilo, izinkampani zemithi, nabakhokhi. Izazi ze-endocrinologists zezingane zingase zikale imiphumela emibi engavamile yemijovo yansuku zonke ye-hormone yokukhula iminyaka engu-2 namathuba okuthi akukho noma ukukhula okuncane kuphela kusayizi womzimba womuntu omdala uma kuqhathaniswa nokwamanje. Abafana bangase bakholelwe ukuthi ngisho noma ukuphakama kwabo kungase kukhuphuke ngo-2 cm kuphela, kufanelekile ukujova i-hormone yokukhula, kodwa umkhokhi kanye nenkampani yemithi ingase ibambe imibono ehlukene.

Sithatha i-creatinine esekelwe eGFR njengesibonelo, okuyinkomba yokusebenza kwezinso esetshenziswa kakhulu ekuxilongeni nasekuhloleni isifo sezinso esingalapheki, ukubeka izimo zokufakelwa kwezinso noma zokunikela, nokunquma indlela yokunciphisa kanye nokungqubuzana kwemithi eminingi kadokotela. I-EGFR isibalo esilula sokuhlehla esisetshenziselwa ukulinganisa isilinganiso sokuhlunga kwe-glomerular esilinganisiwe (mGFR), okuyindinganiso eyireferensi, kodwa indlela yokuhlola inzima uma kuqhathaniswa. Lesi sibalo sokuhlehla asikwazi ukubhekwa njengemodeli ye-AI, kodwa sibonisa izimiso eziningi mayelana namagugu omuntu nokucabanga okunokwenzeka.

Iphuzu lokuqala lokungena lamanani omuntu ukuze angene ku-eGFR yilapho kukhethwa idatha yokulinganisa izibalo. Ulayini wangempela osetshenziselwa ukuklama ifomula ye-eGFR iningi labo linabahlanganyeli abamnyama nabamhlophe, futhi ukusebenza kwayo kwezinye izizwe eziningi akucacile. Amaphuzu okungena alandelayo ezindinganiso zomuntu kule fomula ahlanganisa: ukukhetha ukunemba kwe-mGFR njengenhloso eyinhloko yokuhlola ukusebenza kwezinso, iliphi izinga elamukelekayo lokunemba, indlela yokulinganisa ukunemba, nokusebenzisa i-eGFR njengomkhawulo wokucupha izinqumo zomtholampilo (njengokunquma izimo zokufakelwa kwezinso noma ukunquma imithi). Okokugcina, lapho kukhethwa okuqukethwe kwemodeli yokufaka, amanani omuntu nawo azofaka le fomula.

Isibonelo, ngaphambi kuka-2021, imihlahlandlela iphakamisa ukulungisa amaleveli e-creatinine kufomula ye-eGFR ngokusekelwe eminyakeni yobudala yesiguli, ubulili, nohlanga (okuhlukaniswa kuphela njengabantu abamnyama noma abangebona abamnyama). Ukulungiswa okusekelwe emncintiswaneni kuhloswe ngawo ukuthuthukisa ukunemba kwefomula ye-mGFR, kodwa ngo-2020, izibhedlela ezinkulu zaqala ukungabaza ukusetshenziswa kwe-eGFR esekelwe kuhlanga, ziveza izizathu ezinjengokubambezeleka kokufaneleka kwesiguli ukufakelwa kabusha kanye nomjaho wokuhlanganisa njengomqondo webhayoloji. Ucwaningo luye lwabonisa ukuthi ukuklama amamodeli e-eGFR ngokohlanga kungaba nemithelela ejulile nehlukahlukayo ekunembeni nasemiphumeleni yomtholampilo; Ngakho-ke, ukukhetha ngokukhetha ukunemba noma ukugxila engxenyeni yemiphumela kubonisa ukwahlulela kwenani futhi kungase kufihle ukuthathwa kwezinqumo okusobala. Ekugcineni, ithimba labasebenzi likazwelonke laphakamisa ifomula entsha eyafakwa kabusha ngaphandle kokucabangela ubuhlanga ukuze kulinganiswe izindaba zokusebenza kanye nokungenzeleli. Lesi sibonelo sibonisa ukuthi ngisho nefomula yomtholampilo elula inamaphuzu amaningi okungena kumanani abantu.

Uma kuqhathaniswa namafomula omtholampilo anenani elincane kuphela lezinkomba eziqagelayo, i-LLM ingase ihlanganise izigidigidi kuya kumakhulu ezigidigidi zamapharamitha (imodeli yesisindo) noma ngaphezulu, okwenza kube nzima ukuyiqonda. Isizathu esenza sithi “kunzima ukukuqonda” ukuthi kuma-LLM amaningi, indlela eyiyo yokuthola izimpendulo ngokubuza ayinakudwetshwa. Inombolo yamapharamitha we-GPT-4 ayikamenyezelwa okwamanje; I-GPT-3 eyandulelayo yayinamapharamitha ayizigidi eziyizinkulungwane ezingu-175. Amapharamitha engeziwe awasho ngempela amandla aqinile, njengoba amamodeli amancane afaka imijikelezo yokubala (efana nochungechunge lwemodeli ye-LLaMA [Large Language Model Meta AI]) noma amamodeli ashunwe kahle asekelwe kumpendulo yomuntu azosebenza kangcono kunamamodeli amakhulu. Isibonelo, ngokuya kubahloli babantu, imodeli ye-InstrumentGPT (imodeli enamapharamitha ayizigidi eziyizinkulungwane ezingu-1.3) idlula i-GPT-3 ekuthuthukiseni imiphumela yokuphuma kwemodeli.

Imininingwane ethile yokuqeqeshwa kwe-GPT-4 ayikadalulwa okwamanje, kodwa imininingwane yamamodeli esizukulwane sangaphambilini afaka i-GPT-3, InstrumentGPT, namanye ama-LLM amaningi omthombo ovulekile adaluliwe. Namuhla, amamodeli amaningi e-AI eza namakhadi angamamodeli; Idatha yokuhlola nezokuphepha ye-GPT-4 ishicilelwe ekhadini lesistimu elifanayo elinikezwe inkampani yokudala imodeli i-OpenAI. Ukwakhiwa kwe-LLM kungahlukaniswa ngokwezigaba ezimbili: isigaba sokuqala sokuqeqeshwa kanye nesigaba sokulungisa kahle okuhloswe ngaso ukuthuthukisa imiphumela yokuphuma kwemodeli. Esigabeni sangaphambi kokuqeqeshwa, imodeli ihlinzekwa ngekhorasi enkulu okuhlanganisa nombhalo we-inthanethi wangempela ukuyiqeqeshela ukubikezela igama elilandelayo. Le nqubo ebonakala ilula "yokuqedela ngokuzenzakalela" ikhiqiza imodeli yesisekelo enamandla, kodwa futhi ingaholela ekuziphatheni okulimazayo. Izimiso zobuntu zizongena esigabeni sangaphambi kokuqeqeshwa, okuhlanganisa ukukhetha idatha yokuqeqeshwa kwangaphambili ye-GPT-4 nokunquma ukususa okuqukethwe okungalungile njengokuqukethwe kwezithombe zobulili ezingcolile kudatha yokuqeqeshwa kwangaphambili. Ngaphandle kwale mizamo, imodeli eyisisekelo isengase ingabi wusizo noma ikwazi ukuqukatha imiphumela eyingozi yokukhiphayo. Esigabeni esilandelayo sokulungisa kahle, kuzovela izindlela eziningi zokuziphatha eziwusizo nezingenabungozi.

Esigabeni sokucushwa kahle, ukuziphatha kwamamodeli olimi kuvame ukuguqulwa ngokujulile ngokuhlela kahle okugadiwe nokufunda okuqiniswayo okusekelwe empendulweni yomuntu. Esigabeni sokulungisa kahle okugadwayo, izisebenzi zenkontileka eqashiwe zizobhala izibonelo zokuphendula zamagama asheshayo futhi ziqeqeshe imodeli ngokuqondile. Esigabeni sokufunda sokuqinisa ngokusekelwe empendulweni yomuntu, abahloli abangabantu bazohlunga imiphumela yemodeli njengezibonelo zokuqukethwe okokufaka. Bese usebenzisa imiphumela yokuqhathanisa engenhla ukuze ufunde "imodeli yomvuzo" futhi uthuthukise imodeli ngokufunda ngokuqinisa. Ukuzibandakanya kwabantu okumangalisayo kwezinga eliphansi kungashuna kahle lawa mamodeli amakhulu. Isibonelo, imodeli ye-InstrumentGPT yasebenzisa ithimba labasebenzi bosonkontileka abacishe babe ngu-40 ababuthwe kusukela kumawebhusayithi e-crowdsourcing futhi baphumelela ukuhlolwa kokuhlolwa okuhloswe ukukhetha iqembu lezichasiselo ezizwelayo kokuncanyelwayo kwamaqembu ahlukene abantu.

Njengoba lezi zibonelo ezimbili ezeqisayo, okuyifomula yomtholampilo elula [eGFR] kanye ne-LLM enamandla [GPT-4], zibonisa, ukuthathwa kwezinqumo komuntu kanye nezindinganiso zobuntu zidlala indima ebalulekile ekulolongeni imiphumela yemodeli. Ingabe lezi zinhlobo ze-AI zingathwebula amanani azo ahlukahlukene esiguli kanye nodokotela? Ungakuqondisa kanjani esidlangalaleni ukusetshenziswa kwe-AI kwezokwelapha? Njengoba kushiwo ngezansi, ukuhlolwa kabusha kokuhlaziywa kwesinqumo sezokwelapha kungase kunikeze isixazululo esimisiwe kulezi zinkinga.

Ukuhlaziywa kwesinqumo sezokwelapha akuvamile kodokotela abaningi, kodwa kungahlukanisa phakathi kokucabanga okunokwenzeka (ngemiphumela engaqinisekile ehlobene nokwenziwa kwezinqumo, njengokuthi kufanele kunikezwe i-hormone yokukhula komuntu esimweni somtholampilo esiyimpikiswano esiboniswe kuMfanekiso 1) kanye nezici zokucatshangelwa (kumanani azimele anamathiselwe kule miphumela, inani lawo elilinganiswa njengokuthi "ukuphakama kwe-cm) ukuhlinzeka ngesisombululo sobude be-cm) ngezinqumo eziyinkimbinkimbi zezokwelapha. Ekuhlaziyeni izinqumo, odokotela kufanele baqale banqume zonke izinqumo ezingase zibe khona kanye namathuba ahlobene nomphumela ngamunye, bese behlanganisa usizo lwesiguli (noma elinye iqembu) elihlotshaniswa nomphumela ngamunye ukuze bakhethe inketho efanelekile kakhulu. Ngakho-ke, ukufaneleka kokuhlaziywa kwesinqumo kuncike ekutheni isilungiselelo somphumela siphelele, kanye nokuthi ukulinganiswa kokusetshenziswa kanye nesilinganiso samathuba kunembile yini. Ngokufanelekile, le ndlela isiza ukuqinisekisa ukuthi izinqumo zisekelwe ebufakazini futhi zihambisana nokuthandwayo kwesiguli, ngaleyo ndlela kunciphise igebe phakathi kwedatha yenhloso kanye namanani omuntu siqu. Le ndlela yethulwa emkhakheni wezokwelapha emashumini ambalwa eminyaka edlule futhi yasetshenziswa ekuthathweni kwesinqumo sesiguli ngasinye kanye nokuhlolwa kwezempilo kwenani labantu, njengokuhlinzeka ngezincomo zokuhlolelwa umdlavuza we-colorectal kubantu abaningi.

Ekuhlaziyweni kwezinqumo zezokwelapha, izindlela ezehlukene zenzelwe ukuthola usizo. Izindlela eziningi zendabuko zithola ukubaluleka esigulini ngasinye. Indlela elula iwukusebenzisa isikali sokulinganisa, lapho iziguli zihlola izinga lazo lokuthanda umphumela othile esikalini sedijithali (njengesilinganiso somugqa esisukela ku-1 kuya ku-10), nemiphumela yezempilo eyedlulele kakhulu (njengempilo ephelele nokufa) etholakala kuzo zombili iziphetho. Indlela yokushintshana kwesikhathi ingenye indlela evame ukusetshenziswa. Ngale ndlela, iziguli zidinga ukwenza isinqumo sokuthi singakanani isikhathi esinempilo ezizimisele ukusichitha ngokushintshanisa isikhathi sempilo ebuthaka. Indlela yokugembula evamile ingenye indlela evame ukusetshenziswa yokunquma insiza. Kule ndlela, iziguli zibuzwa ukuthi yiziphi izinketho ezimbili ezizithandayo: noma ziphile inombolo ethile yeminyaka empilweni evamile ngamathuba athile (p) (t), futhi zithwale ingozi yokufa nge-1-p okungenzeka; Qinisekisa ukuthi uphila iminyaka engu-t ngaphansi kwezimo zempilo ezihlukene. Buza iziguli izikhathi eziningi ngamavelu ahlukene we-p kuze kube yilapho zingabonisi ukukhetha kunoma iyiphi inketho, ukuze usizo lungabalwe ngokusekelwe ezimpendulo zesiguli.
Ngaphezu kwezindlela ezisetshenziselwa ukuthola okuthandwa yisiguli ngasinye, kuye kwasungulwa izindlela zokuthola usizo kubantu besiguli. Ikakhulukazi izingxoxo zamaqembu (ukuhlanganisa iziguli ukuze zixoxe ngokuhlangenwe nakho okukhethekile) kungasiza ukuqonda imibono yazo. Ukuze kuhlanganiswe ngempumelelo insiza yeqembu, kuhlongozwe izindlela ezehlukene zezingxoxo zeqembu.
Empeleni, ukwethulwa okuqondile kwensiza ekuxilongweni komtholampilo kanye nenqubo yokwelashwa kudla isikhathi esiningi. Njengesixazululo, imibuzo yenhlolovo ivamise ukusatshalaliswa kubantu abakhethwe ngokungahleliwe ukuze kutholwe amaphuzu wokusetshenziswa ezingeni labantu. Ezinye izibonelo zihlanganisa uhlu lwemibuzo lwe-EuroQol 5-dimensional, ifomu elifushane lesisindo sensiza engu-6-dimensional, I-Health Utility Index, kanye nethuluzi le-Cancer Specific European Cancer Research and Treatment Organization Quality of Life Questionnaire Core 30 ithuluzi.