Kwenzekeni
Abacwaningi bahlole imodeli yolimi enkulu ekukhishweni kokubhaliswa komtholampilo besebenzisa idatha yerekhodi yezokwelapha engacutshungulwanga evela kuzilungiselelo ezimbili ze-American College of Cardiology National Cardiovascular Data Registry. Ucwaningo luhlele imibuzo yokubhalisa yaba izigaba eziyisithupha ngokusekelwe ekungaqondakalini nasekucabangeni komtholampilo okudingekayo ukuyiphendula.
Iphepha, elibuyekezwe ku-arXiv ngo-Aug. 25, libika umshayeli wendiza kanye nocwaningo lokuqinisekisa olubandakanya imibuzo yokubhalisa emitholampilo evela ku-American College of Cardiology's National Cardiovascular Data Registry. Esivivinyweni sesikhungo sezokwelapha sezemfundo, imodeli ikhombe imithombo yedatha yekhandidethi yombuzo ngamunye wokubhalisa. Ama-abstractors anolwazi abe esesebenzisa leyo miphumela ukuchaza amasethi emibhalo eqondene nombuzo othile. Ocwaningweni lokuqinisekisa esikhungweni sesibili, kusetshenziswa ukubhaliswa kwesibili kwe-ACC NCDR, imodeli iphendule imibuzo kulawo masethi. Lo mklamo ugxile ekuhloleni ekukhipheni nasekuhumusheni ulwazi oluvela ezintweni ezikhona ezirekhodiwe zezokwelapha esikhundleni sethebula ledatha elenziwe lula, elikhethwe kusengaphambili.
Ngaphambi kokubuyekeza okuphumayo okuyimodeli, abacaphuni ababili basungula iqiniso eliyisisekelo ngokuzimela futhi banikeza umbuzo ngamunye esigabeni esisodwa kweziyisithupha: Ifulegi Lemithi/Isehlakalo, Ukuba Khona Komtholampilo Okunambambili, Ukuphatha, Ilabhorethri Engakanani/I-Physiologic, Ukutolikwa Komtholampilo, kanye Nesikhathi Somcimbi. Izigaba zihlelwe ngokungaqondakali nokucabanga komtholampilo okudingekayo ukuze kuxazululwe umbuzo ngamunye. Iphepha libika izimpendulo ze-abstractors eziyi-9,430 ezihlanganiswe zaba yizimpendulo zokuvumelana eziyi-4,715, okuhlanganisa nezimpendulo zomshayeli wendiza ezingama-501 kanye nezimpendulo zokuqinisekisa eziyi-4,214. Esivivinyweni sokuhlola, isilinganiso senani lemithombo yedatha yekhandidethi sisuka ku-14.6 ngezibalo zabantu saya kokungu-89.2 kumlando nezici zobungozi, nokuhluka okukhulu okubonakala ekuchezukeni okujwayelekile okubikiwe.
Ekuqinisekiseni, isivumelwano se-inter-rater yabantu sasicishe sibe ngama-98%, ngokusho kocwaningo. Izimpendulo ze-LLM zifana ncamashi nokuvumelana kuma-87% wamacala, zahlukaniswa njengengxenye yokufanisa ngo-2%, futhi azifananga ku-9%. Leli phepha libika ukunemba kwezinga lemibuzo elisho ukuthini elingu-91.5%, nokuchezuka okujwayelekile okungu-13.4%, emibuzweni eyi-157 ngayinye ebinezimpendulo okungenani ezingama-20. Ukunemba kuhlukanisiwe ngesigaba sokungaqondakali, kwehla kusuka ku-96% emibuzweni ye-Medication/Event Flag kuya ku-62% yemibuzo Yesikhathi Somcimbi. Umthombo uhlonza ucwaningo njengokuphrinta kusengaphambili futhi awusho imodeli ehlolwe ku-abstract.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Imiphumela iphakamisa ukuthi ukunemba okumaphakathi kungafihla ubuthakathaka obubalulekile ku-AI yokunakekelwa kwezempilo. Imodeli yenze kahle kakhulu emithini eqondile uma kuqhathaniswa nemibuzo yefulegi lomcimbi futhi okubi kakhulu lapho kufanele ihumushe umongo womtholampilo noma inqume ukuthi umcimbi wenzeke nini.
Okutholakele okumaphakathi akukhona nje ukuthi i-LLM yenze amaphutha. Iwukuthi ukusebenza kwehle ngendlela ehlelekile njengoba imibuzo iba yimpicabadala futhi idinga ukucatshangwa okwengeziwe komtholampilo. Ngakho-ke isibalo esisodwa sokunemba esingama-91.5% singanikeza isithombe esingaphelele sengozi yokusebenza. Ukugeleza komsebenzi wokubhalisa okuqukethe izinkambu eziningi eziqondile kungase kubonakale kuthembekile kuyilapho kungasebenzi kahle kusethi emincane yemibuzo edinga ukwakhiwa kabusha komongo, ukutolika ubufakazi bomtholampilo, noma ukubeka umcimbi ngesikhathi.
Amarejista omtholampilo asekela ucwaningo, ukukalwa kwekhwalithi, ukulinganisa, nezinye izinhlobo zokubikwa kokunakekelwa kwezempilo. Amaphutha ekukhishweni angathinta ikhwalithi yalawo marekhodi angezansi ngisho noma isistimu ingenzi ukuxilonga noma izincomo zokwelashwa. Igebe elibikiwe phakathi kwesivumelwano sabantu esicishe sibe ngu-98% kanye nezinga eliphansi lokufana ncamashi lemodeli libonisa ukuthi ukubuyekezwa komuntu kuhlala kubalulekile kulo msebenzi, ikakhulukazi lapho impendulo incike emibhalweni eminingi noma ekuchazeni ukulandelana kwezenzakalo.
Lolu cwaningo luphinde lunikeze indlela engokoqobo yokuhlola amamodeli olimi lokunakekelwa kwezempilo: ukuhlukanisa imibuzo ngokuya ngohlobo lokungaqondakali oluqukethwe, kunokuphatha yonke imisebenzi yokukhipha njengelinganayo. Leyo ndlela ingasiza izinhlangano ukuthi zibone ukuthi yimiphi imikhakha elungele usizo futhi edinga ukubuyekezwa okuseduze. Umthombo awubonisi ukuthi imodeli ibangele ukulimala kwesiguli, ukusebenza okuthuthukisiwe kokubhalisa, ukunciphisa izindleko, noma yasetshenziswa ekunakekelweni okujwayelekile. Leyo miphumela ihlala ingakaziwa.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ucwaningo alutholi ukuthi ingabe okutholakele kuvamile kwamanye amamodeli, izibhedlela, amakhono ezokwelapha, amarejistri, noma izinqumo zomtholampilo. Umsebenzi owengeziwe kufanele uhlole izilungiselelo ezibanzi, uqhathanise izinguqulo zamamodeli, uhlole imiphumela yamaphutha amancane, futhi uhlole ukuthi ukubuyekezwa komuntu kungawabamba ngokuthembekile amaphutha anzima kakhulu.
Ukhiye ongaziwa ukuthi imiphumela isebenza kabanzi kangakanani. Umthombo uchaza umshayeli wendiza esikhungweni sezokwelapha sezemfundo esisodwa kanye nokuqinisekiswa esikhungweni sesibili esebenzisa enye incwadi yokubhalisa ye-ACC NCDR, kodwa awunikezi ubufakazi ku-abstract mayelana nezibhedlela zomphakathi, ezinye izici ezikhethekile, amasistimu amarekhodi ahlukene, noma ezinye imiklamo yokubhalisa. Imodeli ehloliwe nayo ayihlonzwa ku-abstract. Ukuqhathanisa kuwo wonke amamodeli nezinguqulo zamamodeli kuzodingeka ngaphambi kokwenza iziphetho mayelana nokusebenza kwe-LLM ngokuvamile.
Ukuhlola okuzayo kufanele kubike ngaphezu kokunemba kwesilinganiso esinembile sokufana. Imibuzo ebalulekile ihlanganisa ukuthi ingabe izimpendulo eziyingxenye ziyasebenza emtholampilo, yiziphi izinhlobo zamaphutha ezivame kakhulu, ukuthi amaphutha afaka kaningi kangakanani izinsuku noma amarekhodi aphikisanayo, nokuthi ababuyekezi bangakwazi yini ukubona amaphutha emodeli njalo. Imiphumela yezinga lesigaba socwaningo yenza Isikhathi Somcimbi sibe yindawo ebaluleke kakhulu yokulandelelwa, kodwa umthombo awucacisi imibuzo ngayinye, izibonelo zamaphutha, noma ubukhali bezimpendulo ezingalungile.
Akwaziwa futhi ukuthi ukukhethwa komthombo wemodeli kumshayeli wendiza kuthinte ukuhamba komsebenzi kokuqinisekiswa kwakamuva noma ukuthi ukusebenza okufanayo kuzokwenzeka yini lapho amasethi amadokhumenti engacutshungulwa abacaphuni abanolwazi. Olunye ucwaningo lungahlola amarekhodi angacutshunguliwe ngokugcwele, amazinga ahlukene osizo lwabantu, kanye nokuqapha okulindelwe emisebenzini yokubhalisa yangempela. Kuze kube yilapho lobo bufakazi butholakala, okutholakele kusekela ukwengamela komuntu okuhlosiwe kwemisebenzi engaqondakali yokukhipha kunesiphetho sokuthi ama-LLM asekulungele ukufaka esikhundleni se-abstractors yomtholampilo. Ngakho-ke umthombo ushiya obala ukuthi indlela efanayo ingenza kanjani uma ukukhethwa kwamadokhumenti kungaqondiswa abaqaphi abanolwazi, lapho amarekhodi equkethe amazinga ahlukene esakhiwo, noma lapho ababuyekezi bengena enqubeni ezindaweni ezihlukahlukene.