Autè yo: Ittai Dayan Holger R. Roth nan Aoxiao Zhong Ahmed Harouni Amilcare jèn Anas Z. Abidin Andrey Liu Anthony Beardsworth nan Costa Bradford J. Wood Chèn-Sung Tsai Chih-Hung Wang nan Chun-Nan Hsu C. K. Lee Peiying Ruan Daguang Xu Dufan Wu Eddie Huang Felipe Campos Kitamura Griffin Lacey Gustavo César de Antônio Corradi Gustavo Nino Hao-Hsin Shin Hirofumi Obinata Rèn nan Jason C. Crane nan Jesse Tetreault Jiahui Guan John W. Garrett Joshua D. Kaggie Jung Gil Park Keith Dreyer Krishna Juluru Kristopher Kersten Marcio Aloisio Bezerra Cavalcanti Rockenbach Marius George Linguraru Masoom A. Haider Meena AbdelMaseeh Nicola Rieke Pablo F. Damasceno Pedro Mario Cruz e Silva Pochuan Wang Sheng Xu Shuichi Kawano Sira Sriswasdi Soo Young Park Thomas M. Grist Varun Buch Watsamon Jantarabenjakul Weichung Wang Won Young Tak Xiang Li Xihong Lin Young Joon Kwon Abood Quraini Andrew Feng Andrew N. Priest Baris Turkbey Benjamin Glicksberg Bernardo Bizzo Byung Seok Kim Carlos Tor-Díez Chia-Cheng Lee Chia-Jung Hsu Chin Lin Chiu-Ling Lai Christopher P. Hess Colin Compas Deepeksha Bhatia Eric K. Oermann Evan Leibovitz Hisashi Sasaki Hitoshi Mori Isaac Yang Jae Ho Sohn Krishna Nand Keshava Murthy Li-Chen Fu Matheus Ribeiro Furtado de Mendonça Mike Fralick Min Kyu Kang Mohammad Adil Natalie Gangai Peerapon Vateekul Pierre Elnajjar Sarah Hickman Sharmila Majumdar Shelley L. McLeod Sheridan Reed Stefan Gräf Stephanie Harmon Tatsuya Kodama Thanyawee Puthanakit Tony Mazzulli Vitor Lima de Lavor Yothin Rakvongthai Yu Rim Lee Yuhong Wen Fiona J. Gilbert Mona G. Flores Quanzheng Li Autè yo: Pwodwi pou Telefòn Holger R. Roth nan Pwodwi pou Telefòn Pwodwi pou Telefòn Amilcare jèn Anas Z. Abidin nan Andrey Liu Anthony Beardsworth nan Costa Bradford J. Wood nan Chèn-Sung Tsai Chih-Hung Wang nan Chun-Nan Hsu C. K. Li nan Pwodwi pou Telefòn Pwodwi pou Pwodwi pou Eddy Huang Felipe Campos nan Kitamura Griphin Lacey nan Gustavo César nan Antônio Corradi Gouvènman Nino Pwofesè Shin Hirofumi Obinata nan Rèn nan Jason C. Crane nan Jesse Tetreault nan Jiahui Guan nan John W. Garrett nan Joshua D. Kaggie nan Park nan Jung Gil Pwofesè Dreyer Kreyòl Ayisyen Kris la nan Marcio Aloisio Bezerra Cavalcanti nan Rockenbach Marius George Linguraru nan Masoom A. Haider nan Pwodwi pou Telefòn Nicole Rieke nan Pablo F. nan Damasceno Pwoteksyon nan Pedro Mario Cruz e Silva Pòchuan Wang Pwodwi pou Pwodwi pou Telefòn Shuichi Kawano Pwodwi pou Jwèt nan Soo Young Park Thomas M. Grist nan Liv nan Varun Pwodwi pou Telefòn Wouj Wang Jwenn yon jèn Li nan Pwodwi nou an Jèn Joon Kwon Pwodwi pou Telefòn Andrey Feng nan Andrey N. Priest nan Pwodwi pou Telefòn Benjamin Glicksberg nan Bernardo Bizzo nan Pwoteksyon nan Kim Carlos Tor-Díez nan Li nan Cheng Lee Pwodwi pou Telefòn Chin Lin nan Pwoteksyon Christophe P. Hess nan Kolòn Compas Pwofesyonèl Bhatia Eric K. Oermann nan Evan Leibovitz nan Pwoteksyon nan Sasaki Pwoteksyon Isit la nan Pwensipalman jodi a Kreyòl Ayisyen Li-Chen Fu nan Matheus Ribeiro Furtado nan Mendonça Mike Fralick nan Min Kyu Kang nan Pou egzanp Natali Gangai nan Pwodwi pou Telefòn Pyè Elnajjar Sarah Hickman nan Pwodwi pou Telefòn Shelley L. McLeod nan Règleman Pwofesè nan Pwofesè Harmon Tatsuya Kodama nan Pwodwi pou Telefòn Tòni Mazzulli Vitor Lima nan travay Pwodwi pou Telefòn Yu Rim Lee nan Pwofesyonèl Wen Fiona J. Gilbert nan Mona G. Flores nan Pwodwi pou Abstraksyon Federated Learning setting (FL) se yon metòd ki itilize pou fòmasyon done modèl enstitiyèl ak done ki sòti nan plizyè sous pandan y ap kenbe anonimite done, ak sa ki elimine anpil obstak yo nan pataje done. Isit la nou itilize done ki sòti nan 20 enstiti atravè lemond yo fòme yon modèl FL, rele EXAM (Electronic Medical Record (EMR) chest X-ray AI modèl), ki prezante bezwen oksijèn nan tan kap vini an nan pasyan sentòm ak COVID-19 lè l sèvi avèk entèlijans nan siy vital, done laboratwa ak ray X-ray chest. EXAM te jwenn yon zòn mwayèn anba koub la (AUC) >0.92 pou prezante rezilta nan 24 ak 72 èdtan soti nan tan prezantasyon an premye nan chanm a, epi li Pwodwi Kominote syantifik, akademik, medikal ak syans done yo te vin ansanm nan fòs la COVID-19 pandemik yo rapidman evalye paradigm nouvo nan entèlijans atifisyèl (AI) ki se vit ak san danje, ak potansyèlman ankouraje pataje done ak fòmasyon modèl ak tès san yo pa objaktif prive ak pwoteksyon done obstak nan kolaborasyon konvansyonèl. , Pwofesè swen swen sante, rechèchè ak endistri yo te chanje konsantrasyon yo nan reponn a nesesè ak bezwen klinik kritik ki te kreye pa kri a, ak rezilta remakab , , , , , , Rekrutman nan etid klinik te akselere ak fasilite pa òganizasyon regilatè nasyonal yo ak yon espwa koperasyon entènasyonal , , Analitik done ak AI disiplin yo te toujou ankouraje apwòch louvri ak kolaboratif, enkli konpozisyon tankou lojisyèl sous louvri, replikab rechèch, depo done ak fè disponib anonim dataset piblikman. , Pandèmik la te enstale nesesite a fè kolaborasyon done rapidman ki pèmèt kominote klinik ak syantifik nan reponn a rapidman evolye ak lajman difisil mondyal defi. Done pataje gen etik, règlemantè ak legal konplèksite ki yo enspire, e ka yon ti kras konplike, pa dènye antre nan gwo konpayi teknoloji nan mond lan nan done sante , , . 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Yon egzanp konplè nan kalite kolaborasyon sa yo se travay anvan nou an sou yon modèl SARS-COV-2 sipò desizyon klinik (CDS) ki baze sou AI. Modèl sa a CDS te devlope nan Mass General Brigham (MGB) ak te valide atravè done plizyè sistèm sante. Input yo nan modèl la CDS yo te imaj X-ray (CXR) tete, siy vital, done demografik ak valè laboratwa ki te montre nan piblisasyon anvan yo dwe prezante rezilta yo nan pasyan ak COVID-19 , , , CXR te chwazi kòm entwodiksyon imaj paske li lajman disponib ak souvan endike pa direksyon tankou moun ki bay pa ACR Konpayi Fleischner Pwoteksyon WHO Nasyonal Toracic Societies , Minis Sante nasyonal COVID manuels ak radioloji sosyete atravè lemond pwodiksyon an nan modèl la CDS te yon pousantaj, rele CORISK , ki koresponn ak kondisyon sipò oksijèn ak ki ta ka ede nan triage pasyan pa klini yo nan frontline , , Pwofesè swen swen sante yo te ye prefere modèl ki te valide sou done yo pwòp yo Jodi a, pifò modèl AI, ki gen ladan modèl la CDS ki te anvan anviwònman an, yo te fòme ak valide sou done 'genyen' ki souvan manke diversite , , potansyèlman rezilta nan overfitting ak pi ba generalizability. Sa a ka diminye pa fòmasyon ak done divès kalite soti nan plizyè sit san yo pa santralize done lè l sèvi avèk metòd tankou transfè aprantisaj , FL se yon metòd ki itilize yo fòmasyon modèl AI sou sous done divès kalite, san yo pa done yo dwe transpòte oswa ekspoze deyò kote orijinal yo. Malgre ke aplike nan anpil endistri, FL te dènyèman ofri pou rechèch nan swen sante ant enstitisyon . 18 19 20 21 22 23 24 25 26 27 28 29 30 27 31 32 33 34 35 36 Federated Learning sipòte lanse rapid la nan eksperyans santralman òkestrasyon ak amelyore traceability nan done ak evalyasyon nan chanjman algorithmic ak enpak Yon apwòch nan FL, ki rele client-server, voye yon modèl "non-tren" nan lòt sèvè ( "nodes") ki fè travay fòmasyon pati, nan lòd yo voye rezilta yo tounen yo dwe fonde nan santral la ( "federated") sèvè. Sa a se pote soti kòm yon pwosesis iteratif jiskaske fòmasyon an se konplè. . 37 36 Gouvènman nan done pou FL se kenbe lokalman, diminye pwoblèm prive, ak sèlman pwa modèl oswa gradients kominikasyon ant sit kliyan ak sèvè a federasyon , FL te deja montre pwomèt nan aplikasyon yo medikal imaj , , , , ki gen ladan nan analiz la COVID-19 , , Yon egzanp enpòtan se yon modèl prediksyon mortalité nan pasyan enfekte ak SARS-COV-2 ki sèvi ak karakteristik klinik, sepandan limite nan kantite modalit yo ak skalè . 38 39 40 41 42 43 8 44 45 46 Objè nou an te devlope yon modèl robust, generalizable ki ta ka ede nan triage pasyan yo. Nou teyori ke modèl la CDS ka federalize avèk siksè, paske li te itilize nan enpòtan done ki se relatif komen nan pratik klinik ak ki pa depann anpil sou evalyasyon operatè-depann nan kondisyon pasyan (tankou impressions klinik oswa simptom rapòte). Anplis de sa, rezilta laboratwa, siy vit, yon etid imajinasyon ak yon demografik anjeneral retire (ki se, laj), yo te itilize. Se poutèt sa, nou retrained modèl la CDS ak done divès kalite lè l sèvi avèk yon apwòch FL kliyan-server yo devlope yon nouvo modèl FL mondyal, ki te rele EXAM, lè l sèvi avèk karakteristik CXR ak EM Hypothesis nou an te ke EXAM pral pèfòmans pi bon pase modèl lokal yo ak pral jeneralize pi bon nan tout sistèm swen sante. Rezilta Architecture modèl egzamen an Modèl EXAM se ki baze sou modèl la CDS ki te di pi wo a Totalman, 20 karakteristik (19 soti nan EMR ak yon CXR) te itilize kòm entwodiksyon nan modèl la. Rezilta (ki se, 'tou a tè') etikèt yo te bay ki baze sou tretman oksijèn nan pasyan an apre 24 ak 72 èdtan soti nan admisyon an inisyal nan depatman an enkyetid (ED). Yon lis detaye nan karakteristik yo mande ak rezilta yo ka wè nan Tab la . 27 1 Etikèt rezilta yo nan pasyan yo te mete nan 0, 0,25, 0,50 ak 0,75 depann sou terapi oksijèn ki pi enpòtan ki te resevwa pasyan an nan fenèt pratik la. Kategori yo nan tretman oksijèn te, respektivman, lè nan chanm (RA), oksijèn ki ba-flux (LFO), segondè-flux oksijèn (HFO)/ventilasyon noninvasif (NIV) oswa vantilasyon mekanik (MV). Si pasyan a te mouri nan fenèt pratik la, etikèt la rezilta te mete nan 1. Sa a te rezilta nan chak ka yo bay de etikèt nan ranje a 0-1 ki koresponn ak chak nan fenèt pratik la (ki se, 24 ak 72 èdtan). Pou karakteristik EMR, sèlman valè yo premye ki te capture nan ED a te itilize ak preprocessing done ki gen ladan deidentifikasyon, imputation valè manke ak normalization nan zè-mwayen ak varyans inite. Pou imaj CXR, sèlman premye ki te jwenn nan ED a te itilize. Modèl la se konsa fusion enfòmasyon ki soti nan tou de EMR ak CXR karakteristik, lè l sèvi avèk yon 34-layer konvolisyonèl rezo neural (ResNet34) yo ekstrè karakteristik soti nan yon CXR ak yon rezo Deep & Cross yo concatenate karakteristik yo ansanm ak karakteristik yo EMR (pou plis detay, wè ). pwodiksyon modèl se yon pousantaj risk, rele pousantaj EXAM, ki se yon valè kontinyèl nan ranje a 0-1 pou chak nan pratik yo 24 ak 72 èdtan ki koresponn ak etikèt yo dekri pi wo a. Metòd Federasyon nan modèl la Modèl la EXAM te fòme lè l sèvi avèk yon kohorte de 16.148 ka, fè li pa sèlman nan mitan premye modèl FL pou COVID-19 men tou yon pwojè devlopman trè gwo ak multicontinents nan klinikman enpòtan AI (Fig. ). Done ant sit yo pa te harmonize anvan ekstraksyon an, ak, nan figi a reyèl konpòtman an nan òdinatè klinik, yon harmonize meticulous nan done enpòte pa te pote pa otè yo (Fig. ) nan 1A ak B 1C, D nan , Mondyal kat ki montre 20 diferan sit kliyan kontribye nan etid la EXAM. , Nimewo ka kontribye pa chak enstitisyon oswa sit (client 1 reprezante sit la kontribye pi gwo kantite ka). , Distribisyon entansite X-ray chest nan chak sit kliyan. , Èd nan pasyan nan chak sit kliyan, ki montre laj minimòm ak maksimòm (asterisks), laj mwayèn (triangles) ak devwa estanda (bar orizontal). Nimewo echantiyon nan chak sit kliyan se montre nan Tables Supplementary . a b c d 1 Nou te konpare modèl lokalman fòmasyon ak modèl global FL sou done tès chak kliyan. fòmasyon modèl la nan FL rezilta nan yon amelyorasyon enpòtan nan pèfòmans ( « 1 × 10–3, Wilcoxon signe-ranke tès) nan 16% (kòm defini pa AUC mwayèn lè kouri modèl la sou set tès lokal yo respektivman: soti nan 0.795 nan 0.920, oswa 12.5 pwen pwosesis) (Fig. ). Li te tou rezilta nan 38% amelyorasyon nan generalizability (kòm defini pa AUC an mwayen lè kouri modèl la sou tout seri tès: soti nan 0.667 nan 0.920, oswa 25.3 pwen pousantaj) nan pi bon modèl mondyal pou pratik nan tretman oksijèn 24 èdtan konpare ak modèl fòmasyon sèlman sou done pwòp nan yon sit (Fig. Pou rezilta pratik nan 72 èdtan tretman oksijèn, pi bon fòmasyon modèl mondyal la te rezilta nan yon amelyorasyon nan pèfòmans an mwayen de 18% konpare ak modèl yo fòme lokalman, pandan y ap generalizability nan modèl la mondyal la amelyore an mwayen de 34% (Extended Data Fig. ). Stabilite nan rezilta nou yo te valide pa repete twa kouri nan fòmasyon lokal ak FL sou diferan randomize divizyon done. P 2a nan 2b nan 1 , pèfòmans sou tès chak kliyan mete nan pratik nan 24 èdtan tretman oksijèn pou modèl fòmasyon sou done lokal sèlman (Local) konpare ak pi bon modèl mondyal ki disponib sou sèvè a (FL). Av, pèfòmans tès mwayèn nan tout sit. , Generalizability (mwayen pèfòmans sou done tès nan lòt sit, ki reprezante pa AUC mwayen) kòm yon fonksyon nan gwosè dataset la nan yon kliyan (pa gen okenn ka). Liy la orizontal jòn endike pèfòmans nan generalizability nan modèl la pi bon mondyal. pèfòmans la pou 18 nan 20 kliyan se montre, paske kliyan 12 te gen rezilta sèlman pou 72 èdtan oksijèn (Extended Data Fig. ) ak kliyan 14 te gen ka sèlman ak tretman RA, se konsa ke metrik evalyasyon (avèk AUC) pa te aplike nan nenpòt nan ka sa yo ( Done pou kliyan 14 yo tou te eksepsyonèlman soti nan kalkil la nan generalizability an mwayen nan modèl lokal. a b 1 Metòd Modèl lokal ki te fòme lè l sèvi avèk kohorte ki pa balans (pou egzanp, sitou vakans mild nan COVID-19) te benefisye nan enfòmatizasyon an FL, ak yon amelyorasyon enpòtan nan pratik performans AUC mwayen pou kategori ak sèlman kèk ka. Sa a te evidan nan sit kliyan 16 (ak yon dataset ki pa balans), ak pifò pasyan ki gen gravite maladi mild ak ak sèlman kèk ka grav. modèl la FL te reyalize yon pousantaj pi wo verite-pozitif pou de ka pozitif (gravite) ak yon pousantaj pi ba falsite-pozitif konpare ak modèl la lokal, tou de yo montre nan ROC (Receiver Operating Characteristic) plots ak matris konfuzyon (Fig. Epitou plis done Fig. Pifò enpòtan, generalizability nan modèl la FL te ogmante anpil sou modèl la lokalman fòme. 3a nan 2 , ROC nan sit kliyan 16, ak done ki pa balans ak sitou ka miltip. , ROC nan modèl lokal la nan sit kliyan 12 (petèt dataset), ROC medyeval nan modèl fòmasyon sou dataset pi gwo ki koresponn ak senk sit kliyan nan zòn nan Boston (1, 4, 5, 6, 8) ak ROC nan modèl la pi bon mondyal nan prezante tretman oksijèn 72 èdtan pou diferan pousantaj nan pousantaj EXAM (lwa, mitan, dwat). ROC medyeval se kalkil ki baze sou senk modèl yo fòmasyon lokalman pandan y ap zòn la gri denote deviasyon estanda ROC. ROCs pou twa valè diferan koupe ( Pos ak neg denote kantite ka pozitif ak negatif, respektivman, kòm definye pa varyete sa a nan pousantaj egzamen an. a b t Nan ka a nan sit kliyan ak datasèt relativman ti, modèl FL a pi bon te parèt pa sèlman modèl la lokal, men tou moun ki fòmasyon sou datasèt pi gwo soti nan senk sit kliyan nan Boston zòn nan USA a (Fig. ) nan 3b nan Modèl la mondyal te fè byen nan prezante bezwen oksijèn nan 24/72 èdtan nan pasyan tou de COVID pozitif ak negatif (Extended Data Fig. ) nan 3 Validasyon nan sit inik Following initial training, EXAM was subsequently tested at three independent validation sites: Cooley Dickinson Hospital (CDH), Martha’s Vineyard Hospital (MVH) and Nantucket Cottage Hospital (NCH), all in Massachusetts, USA. The model was not retrained at these sites and it was used only for validation purposes. The cohort size and model inference results are summarized in Table , ak curves ROC ak matris konfuzyon pou dataset la pi gwo (nan CDH) yo montre nan Figi. . Pwen operasyon te mete nan diferans ant vantilasyon ki pa mekanik ak vantilasyon mekanik (MV) tretman (oswa mouri). Modèl la fòmasyon FL mondyal la, EXAM, te jwenn yon AUC medyeval de 0,944 ak 0,924 pou 24 ak 72 èdtan pratik travay, respektivman (Tab la ), ki te depase pèfòmans an mwayen ant sit yo itilize nan fòmasyon EXAM. Pou prezante tretman MV (oswa mouri) nan 24 èdtan, EXAM te jwenn yon sensibilite nan 0.950 ak spesifikite nan 0.882 nan CDH, ak yon sensibilite nan 1.000 spesifikite nan 0.934 nan MVH. NCH pa gen okenn ka ak MV / mouri nan 24 èdtan. Kòm pou 72-èdtan MV prezantasyon, EXAM te jwenn yon sensibilite nan 0.929 ak spesifikite nan 0.880 nan CDH, sensibilite nan 1.000 ak spesifikite nan 0.976 nan MVH ak sensibilite nan 1.000 ak spesifikite nan 0.929 nan NCH. 2 4 2 , , Performance (ROC) (top) ak confusion matriks (bottom) nan modèl la EXAM FL sou dataset la CDH pou prezante bezwen oksijèn nan 24 èdtan ( ) ak 72 èdtan ( ). ROCs for three different cutoff values ( ) nan pousantaj egzamen risk yo montre. a b a b t Pou MV nan CDH nan 72 èdtan, EXAM te gen yon pousantaj ba false-negatif nan 7.1%. Reprezentatif ka default yo prezante nan Extended Data Fig. , ki montre de ka false-negatif soti nan CDH kote yon ka te gen anpil manke karakteristik done EMR ak lòt la te gen yon CXR ak yon artefakt mouvman ak kèk manke karakteristik EMR. 4 Use of differential privacy Yon motivasyon prensipal pou enstitisyon swen swen sante sèvi ak FL se pou kenbe sekirite a ak prive nan done yo, osi byen ke aderans nan mas yo konformite done. Pou FL, gen yon risk potansyèl nan modèl 'inversion' oswa menm rekonstriksyon nan imaj fòmasyon soti nan gradients modèl yo menm . To counter these risks, security-enhancing measures were used to mitigate risk in the event of data ‘interception’ during site-server communication . Nou te eksperyans ak teknik pou evite interception nan FL done, epi ajoute yon karakteristik sekirite ki nou kwè ta ka ankouraje plis enstitisyon yo sèvi ak FL. Nou se konsa valide konklizyon anvan yo ki montre ke pataje pwa, ak lòt teknik diferan prive, ka siksè aplike nan FL . Through investigation of a partial weight-sharing scheme , , , nou te montre ke modèl ka jwenn yon pèfòmans konparab menm lè sèlman 25% nan ajou pwa yo pataje (Extended Data Fig. ) nan 47 48 49 50 50 51 52 5 Diskisyon Nan etid sa a, gen yon gwo, reyèl etid FL nan swen swen sante nan kantite sit ak kantite pwen done itilize. Nou kwè ke li bay yon pwouve-of-konsep fòs nan itilizasyon FL pou devlopman rapid ak kolaboratif nan modèl AI nesesè nan swen sante. Etid nou an gen ladan plizyè sit sou kat kontinan ak anba supervizyon nan òganizasyon diferan, ak se konsa kenbe pwomèt la pou yo bay nan diferan mache reglemante nan yon fason akselere. Modèl FL mondyal la, EXAM, te pwouve yo pi fò ak reyalize pi bon rezilta sou sit endividyèl pase nenpòt modèl fòme sou sèlman done lokal yo. Nou kwè ke amelyorasyon konsistan te rive akòz yon gwo, men tou plis divès, seri done, itilizasyon Pou yon sit kliyan ak yon dataset relativman ti, de metòd tipik yo ka itilize pou mete yon modèl itil: youn se fòmasyon lokal ak done pwòp li yo, lòt la se aplike yon modèl fòmasyon sou yon dataset pi gwo. Pou sit ak dataset ti, li ta dwe pratikman pa posib yo bati yon modèl aprantisaj fonksyonèl lè l sèvi avèk sèlman done lokal yo. Konstatasyon an, ke de metòd sa yo te depase sou tout twa travay pratik pa modèl la FL mondyal la, indike ke benefis la pou sit kliyan ak dataset ti ki rezilta nan patisipasyon nan kolaborasyon FL se substantif. Sa a se posib yon refleksyon nan kapasite FL a pou retire plis divèsite pase fòmasyon lokal la, ak diminye bias yo prezan nan modèl sa yo fòmasyon sou yon popilasyon homogène. Soti nan . 46 Rezilta yo validasyon konfime ke modèl la mondyal se solid, sipòte hypothesis nou an ke modèl fòmasyon FL yo generalizable atravè sistèm swen swen sante. Yo bay yon ka enpòtan pou itilize nan algorithms pratik nan swen pasyan COVID-19, ak itilize nan FL nan kreyasyon modèl ak tès. Pa patisipe nan etid sa a, sit kliyan yo te resevwa aksè nan EXAM, yo dwe valide plis anvan pwomèt nenpòt approbation règlemantè oswa futures introduksyon nan swen klinik. Plan yo se nan wout la pou valide EXAM prospectively nan 'produksyon' anviwònman nan MGB ki sèvi ak COVID-19 resous objektif , as well as at different sites that were not a part of the EXAM training. 53 Plis pase 200 modèl pratik pou sipòte desizyon nan pasyan ak COVID-19 yo te pibliye . Unlike the majority of publications focused on diagnosis of COVID-19 or prediction of mortality, we predicted oxygen requirements that have implications for patient management. We also used cases with unknown SARS-COV-2 status, and so the model could provide input to the physician ahead of receiving a result for PCR with reverse transcription (RT–PCR), making it useful for a real-life clinical setting. The model’s imaging input is used in common practice, in contrast with models that use chest computed tomography, a nonconsensual diagnostic modality. The model’s design was constrained to objective predictors, unlike many published studies that leveraged subjective clinical impressions. The data collected reflect varied incidence rates, and thus the ‘population momentum’ we encountered is more diverse. This implies that the algorithm can be useful in populations with different incidence rates. 19 Patient cohort identification and data harmonization are not novel issues in research and data science , but are further complicated, when using FL, given the lack of visibility on other sites’ datasets. Improvements to clinical information systems are needed to streamline data preparation, leading to better leverage of a network of sites participating in FL. This, in conjunction with hyperparameter engineering, can allow algorithms to ‘learn’ more effectively from larger data batches and adapt model parameters to a particular site for further personalization—for example, through further fine-tuning on that site Yon sistèm ki pral pèmèt konplètman, prèske a tan reyèl modèl inferans ak pwosesis rezilta yo pral tou benefisye epi yo pral 'touche bouch la' soti nan fòmasyon nan deplwaman modèl. 54 39 Pandan ke done yo pa te santralize, yo pa fasil aksè. Dapre sa, nenpòt analiz nan tan kap vini an nan rezilta yo, plis pase sa ki te derive ak kolekte, se limite. Similar to other machine learning models, EXAM is limited by the quality of the training data. Institutions interested in deploying this algorithm for clinical care need to understand potential biases in the training. For example, the labels used as ground truth in the training of the EXAM model were derived from 24- and 72-h oxygen consumption in the patient; it is assumed that oxygen delivered to the patient equates the oxygen need. However, in the early phase of the COVID-19 pandemic, many patients were provided high-flow oxygen prophylactically regardless of their oxygen need. Such clinical practice could skew the predictions made by this model. Kòm aksè a nan done nou yo te limit, nou pa te gen ase enfòmasyon ki disponib pou kreye estatistik detaye sou ka mal, post-hoc, nan pifò sit. Sepandan, nou te etidye ka mal soti nan pi gwo sit tès endividyèl la, CDH, epi yo te kapab kreye ipotèsis ke nou ka tès nan tan kap vini an. Pou sit pèfòmans segondè, li sanble ke pifò ka mal ranplase nan youn nan de kategori: (1) bon jan kalite ba nan done enprime - pou egzanp, manke done oswa artefakt mouvman nan CXR; oswa (2) done soti nan distribisyon - pou egzanp yon pasyan trè jèn. In future, we also intend to investigate the potential for a ‘population drift’ due to different phases of disease progression. We believe that, owing to the diversity across the 20 sites, this risk may have been mitigated. A feature that would enhance these kinds of large-scale collaboration is the ability to predict the contribution of each client site towards improving the global FL model. This will help in client site selection, and in prioritization of data acquisition and annotation efforts. The latter is especially important given the high costs and difficult logistics of these large-consortia endeavors, and it will enable these endeavors to capture diversity rather than the sheer quantity of data samples. Pwovens apwòch ka entegre otomatik hyperparameter rechèch Search nan Neural Architecture ak lòt aprantisaj machin otomatik apwòch yo jwenn paramèt fòmasyon optimum pou chak sit kliyan pi efikasman. 55 56 57 Konnen pwoblèm nan normalization batch (BN) nan FL motivated nou yo fikse modèl baz nou an pou ekstraksyon karakteristik imaj to reduce the divergence between unbalanced client sites. Future work might explore different types of normalization techniques to allow the training of AI models in FL more effectively when client data are nonindependent and identically distributed. 58 49 Pwosesis ki sot pase a sou atak pwoteksyon an nan anviwònman an FL te rele pwoblèm sou lekòl done pandan fòmasyon modèl Mwayènman, algorithms pwoteksyon yo toujou pa eksplore ak limite pa faktori plizyè. Pandan ke diferansal algorithms pwoteksyon , , montre bon pwoteksyon, yo ka ranfòse pèfòmans nan modèl la. Algoritm enkripsyon, tankou enkripsyon homomorf , kenbe pèfòmans, men ka substantivman ogmante gwosè mesaj ak tan fòmasyon. Yon fason ki kvantifye pou mesye prive pèmèt chwa pi bon pou deside paramèt minimòm prive nesesè pandan y ap kenbe pèfòmans klinikman akseptab , , . 59 36 48 49 60 36 48 49 Apre validasyon plis, nou konsidere deplwaman nan modèl la EXAM nan anviwònman an ED kòm yon fason pou evalye risk nan tou de pati-pati ak nivo popilasyon, epi yo bay klinikè ak yon pwen referans adisyonèl lè fè travay la souvan difisil nan triage pasyan yo. Nou konsidere tou lè l sèvi avèk modèl la kòm yon metrik pi sensitif nan nivo popilasyon yo ede balans resous ant rejyon, hospitalis ak departman yo. Espere nou se ke efò FL menm jan an ka rompe silos done yo ak pèmèt devlopman pi vit nan modèl AI ki trè bezwen nan tan kap vini an. Methods Etik apwouve All procedures were conducted in accordance with the principles for human experimentation as defined in the Declaration of Helsinki and International Conference on Harmonization Good Clinical Practice guidelines, and were approved by the relevant institutional review boards at the following validation sites: CDH, MVH, NCH and at the following training sites: MGB, Mass General Hospital (MGH), Brigham and Women’s Hospital, Newton-Wellesley Hospital, North Shore Medical Center and Faulkner Hospital (all eight of these hospitals were covered under MGB’s ethics board reference, no. 2020P002673, and informed consent was waived by the instititional review board (IRB). Similarly, participation of the remaining sites was approved by their respective relevant institutional review processes: Children’s National Hospital in Washington, DC (no. 00014310, IRB certified exempt); NIHR Cambridge Biomedical Research Centre (no. 20/SW/0140, informed consent waived); The Self-Defense Forces Central Hospital in Tokyo (no. 02-014, informed consent waived); National Taiwan University MeDA Lab and MAHC and Taiwan National Health Insurance Administration (no. 202108026 W, informed consent waived); Tri-Service General Hospital in Taiwan (no. B202105136, informed consent waived); Kyungpook National University Hospital in South Korea (no. KNUH 2020-05-022, informed consent waived); Faculty of Medicine, Chulalongkorn University in Thailand (nos. 490/63, 291/63, informed consent waived); Diagnosticos da America SA in Brazil (no. 26118819.3.0000.5505, informed consent waived); University of California, San Francisco (no. 20-30447, informed consent waived); VA San Diego (no. H200086, IRB certified exempt); University of Toronto (no. 20-0162-C, informed consent waived); National Institutes of Health in Bethesda, Maryland (no. 12-CC-0075, informed consent waived); University of Wisconsin-Madison School of Medicine and Public Health (no. 2016-0418, informed consent waived); Memorial Sloan Kettering Cancer Center in New York (no. 20-194, informed consent waived); and Mount Sinai Health System in New York (no. IRB-20-03271, informed consent waived). MI-CLAIM guidelines for reporting of clinical AI models were followed (Supplementary Note ) 2 etidye mete The study included data from 20 institutions (Fig. ): MGB, MGH, Brigham ak Women's Hospital, Newton-Wellesley Hospital, North Shore Medical Center ak Faulkner Hospital; Children's National Hospital nan Washington, DC; NIHR Cambridge Biomedical Research Centre; The Self-Defense Forces Central Hospital nan Tokyo; National Taiwan University MeDA Lab ak MAHC ak Taiwan National Health Insurance Administration; Tri-Service General Hospital nan Taiwan; Kyungpook National University Hospital nan Kore di Sid; Fakilte Medsin, Chulalongkorn University nan Thailand; Diagnosticos da America SA nan Brezil; University of California, San Francisco; VA San Diego; University of Toronto; National Institutes of Health nan Bethesda, Maryland; University of Wisconsin-Madison School of Medicine and Public Health; Memorial Sloan Kettering Cancer Center nan New York; ak Mount Sinai , , Done soti nan twa sit endividyèl yo te itilize pou validasyon endividyèl: CDH, MVH ak NCH, tout nan Massachusetts, USA. Sa yo twa syans te gen karakteristik popilasyon pasyan diferan de sit fòmasyon yo. Done yo itilize pou validasyon algorithm konsiste de pasyan admis nan ED nan sit sa yo ant mwa mas 2020 ak fevriye 2021, ak ki satisfè kritè yo enklizyon menm jan ak done yo itilize nan fòmasyon modèl la FL. 1a nan 61 62 63 Koleksyon done 20 sit kliyan prepare yon total de 16,148 ka (tou pozitif ak negatif) pou aplikasyon pou fòmasyon, validasyon ak tès modèl la (Fig. ). Done medikal te jwenn nan relasyon ak pasyan ki te satisfè kritè yo enklizyon nan etid la. Sit kliyan te eseye entegre tout ka COVID-pozitif depi kòmansman an nan pandèmyen an nan mwa desanm 2019 ak jiska lè yo te kòmanse fòmasyon lokal pou etid la EXAM. Tout fòmasyon lokal te kòmanse pa 30 septanm 2020. Sit la tou enkli lòt pasyan nan menm peryòd la ak rezilta tès RT-PCR negatif. Pandan ke pi fò nan sit yo te gen plis pasyan SARS-COV-2-negatif pase -pozitif, nou limite kantite pasyan negatif enkli nan, pi wo a, 95% nan kantite ka nan chak sit kliyan. 1b Yon 'cas' enkli yon CXR ak done enpòtan yo te pran soti nan dosye medikal la nan pasyan an. Yon desann nan gwosè cohort nan dataset la pou chak sit kliyan se montre nan Figi. Distribisyon an ak modèl nan entansite imaj CXR (valè piksèl) varye anpil ant sit akòz yon plizyè faktè pasyan- ak sit-spesifik, tankou diferan manifakti aparèy ak pwotokòl imaj, tankou montre nan Figi. . laj pasyan ak distribisyon karakteristik EMR varye anpil ant sit, kòm te espere akòz diferans demografik ant hospitalis distribiye atravè lemond (Extended Data Fig. ) nan 1b nan 1C, D nan 6 Kritè entegre nan pasyan Kritè yo enklizyon pasyan yo te: (1) pasyan an te prezante nan syans la ED oswa ekivalan; (2) pasyan an te gen yon tès RT-PCR te fè nan nenpòt ki lè ant prezantasyon an nan syans la ED ak retire soti nan syans la; (3) pasyan an te gen yon CXR nan syans la ED; ak (4) dosye a pasyan an te gen omwen senk nan valè yo EMR detaye nan Tables la. , all obtained in the ED, and the relevant outcomes captured during hospitalization. Of note, The CXR, laboratory results and vitals used were the first available for capture during the visit to the ED. The model did not incorporate any CXR, laboratory results or vitals acquired after leaving the ED. 1 Model input Totalman, 21 karakteristik EMR yo te itilize kòm entwodiksyon nan modèl la. Etikèt rezilta (ki se, fondasyon an) yo te bay ki baze sou kondisyon pasyan an apre 24 ak 72 èdtan nan admisyon an inisyal nan ED. Yon lis detaye nan karakteristik EMR yo mande ak rezilta yo ka wè nan Tables la . 1 Distribisyon nan tretman oksijèn lè l sèvi avèk aparèy diferan nan sit kliyan diferan yo montre nan Extended Data Fig. , which details the device usage at admission to the ED and after 24- and 72-h periods. The difference in dataset distribution between the largest and smallest client sites can be seen in Extended Data Fig. . 7 8 Nimewo de ka COVID-19 pozitif, kòm konfime pa yon sèl RT-PCR tès te jwenn nan nenpòt ki tan ant prezantasyon an nan ED ak retire soti nan hospitalizasyon an, se lis nan Tables Supplementary . Chak sit kliyan te mande yo randomize dataset li yo nan twa pati: 70% pou fòmasyon, 10% pou valizyon ak 20% pou tès. Pou tou de 24 ak 72 èdtan modèl pratik rezilta, randomize divizyon pou chak nan twa repete lokal ak FL fòmasyon ak evalyasyon eksperyans te kreye otomatikman. 1 EXAM model development Gen yon varyasyon laj nan kou klinik nan pasyan ki prezante nan hospitalizasyon ak sentòm yo nan COVID-19, ak kèk ki gen yon deteriorasyon rapid nan fonksyon respiratwa ki mande pou diferan entèvyou pou anpeche oswa diminye hipoksemi. , Yon desizyon enpòtan te pran pandan evalyasyon an nan yon pasyan nan pwen inisyal swen, oswa nan ED, se si pasyan an ta ka mande plis envasif oswa resous-limited kontre-mesaj oswa entèvyou (tankou MV oswa antitè monoklòn), ak se konsa ta dwe resevwa yon terapi rare men efikas, yon terapi ak yon ranje risk-benefit rapò akòz efè adisyonèl oswa yon nivo pi wo nan swen, tankou admisyon nan inisyal swen Nan kontrè, yon pasyan ki se nan yon risk ki pi ba nan bezwen tretman oksijèn envasif ka mete nan yon anviwònman sante mwens-intensif tankou yon chanm regilye, oswa menm libere soti nan ED pou kontinye self-monitoring nan kay la . EXAM was developed to help triage such patients. 62 63 64 65 Of note, the model is not approved by any regulatory agency at this time and it should be used only for research purposes. egzamen an EXAM was trained using FL; it outputs a risk score (termed EXAM score) similar to CORISK (Done ekstansyon nan Fig. ) ak ka itilize nan menm fason an pou triche pasyan yo. Li koresponn ak kondisyon sipò oksijèn nan yon pasyan nan twa fenèt - 24 ak 72 èdtan - apre prezantasyon an premye nan ED. Extended Data Fig. illustre ki jan CORISK ak pousantaj EXAM ka itilize pou sètifikasyon pasyan. 27 9a nan 9b nan Ranje X-ray imaj yo preprocessed yo chwazi imaj la pozisyon anvan ak elimine imaj gade latéral, ak Lè sa a, scaled nan yon rezolisyon de 224 × 224. Kòm yo montre nan Extended Data Fig. , modèl la fonde enfòmasyon ki soti nan tou de karakteristik EMR ak CXR (ki baze sou yon modifye ResNet34 ak atansyon espesifik pretrained on the CheXpert dataset) ak rezo a Deep & Cross Pou konvergans sa yo diferan kalite done, yon vektè karakteristik 512-dimansyon te ekstrè soti nan chak imaj CXR lè l sèvi avèk yon ResNet34 prétrainé, ak atansyon espesyalis, Lè sa a, konkate ak karakteristik yo EMR kòm envantè a pou rezo a Deep & Cross. pwodiksyon final la te yon valè kontinyèl nan ranje a 0-1 pou tou de 24 ak 72 èdtan prediksyon, ki koresponn ak etikèt yo dekri pi wo a, jan yo montre nan Figi Done Extended. . We used cross-entropy as the loss function and ‘Adam’ as the optimizer. The model was implemented in Tensorflow Sèvi ak NVIDIA Clara Train SDK . The average AUC for the classification tasks (≥LFO, ≥HFO/NIV or ≥MV) was calculated and used as the final evaluation metric, with normalization to zero-mean and unit variance. CXR images were preprocessed to select the correct series and exclude lateral view images, then scaled to a resolution of 224 × 224 (ref. ) nan 9a 66 67 68 9b nan 69 70 27 Karakteristik imputation ak normalization A MissForest algorithm te itilize pou imite karakteristik EMR, ki baze sou dataset la fòmasyon lokal. Si yon karakteristik EMR te manke konplètman nan yon dataset sit kliyan, valè mwayèn nan karakteristik sa a, ki te kalkule sèlman sou done ki soti nan sit kliyan MGB, te itilize. Lè sa a, karakteristik EMR te reskalye nan zèlen-mwayen ak varyans inite ki baze sou estatistik yo kalkil sou done ki soti nan sit kliyan MGB. 71 Details of EMR–CXR data fusion using the Deep & Cross network To model the interactions of features from EMR and CXR data at the case level, a deep-feature scheme was used based on a Deep & Cross network architecture . Karakteristik binè ak kategori pou antrepriz yo EMR, osi byen ke karakteristik imaj 512-dimansyon nan CXR a, te transfere nan vètikal dense fonksyonèl nan valè reyèl pa embedding ak stacking layers. Vètikal yo transformed dense te sèvi kòm antrepriz nan framework a fonksyonèl, ki espesifikman te itilize yon rezo krize pou egzekite fusion ant antrepriz ki soti nan sous diferan. Rezo krize te fè eksplicit karakteristik krize nan layers li yo, pa kondwi pwodwi enteryè ant karakteristik enprime orijinal la ak pwodiksyon ki soti nan layer anvan, ak sa ki ogmante nivo a nan interaksyon ant karakteristik yo. An menm tan an, de rezo klasik dyp ak plizyè ki konekte konplètman- 68 FL details Pwobableman fòm la ki pi etabli nan FL se implemantasyon an nan algorithm la medya federasyon tankou pwopoze pa McMahan et al. , oswa varyasyon yo. Algorithm sa a ka realize lè l sèvi avèk yon Customer-Server konfigirasyon kote chak sit patisipan ap travay kòm yon kliyan. Yon moun ka panse FL kòm yon metòd ki vle minimize yon fonksyon pèdi mondyal pa diminye yon seri de fonksyon pèdi lokal, ki yo estime nan chak sit. Pa minimize pèdi lokal nan chak sit kliyan pandan y ap tou senkronize pèdi nan sit kliyan aprann sou yon sèvè agrégation santralize, yon ka minimize pèdi mondyal san yo pa bezwen jwenn aksè nan tout dataset la nan yon kote santralize. Chak sit kliyan aprann lokalman, epi pataje ajou pèdi modèl ak yon sèvè santral ki agrege kontribisyon lè l sèvi avèk enkripsyon couchèt sekirite ak pwotokòl kominikasyon. ). 72 9c Yon pseudoalgoritm nan FL se montre nan Notasyon Supplementè Nan eksperyans nou yo, nou mete kantite ronde federasyon nan = 200, ak yon etap fòmasyon lokal pou chak ronde nan chak kliyan. Nimewo nan kliyan, , te jiska 20 depann sou koneksyon rezo a nan kliyan yo oswa done ki disponib pou yon peryòd rezilta espesifik (24 oswa 72 èdtan). Nimewo iterasyon fòmasyon lokal, , depann sou gwosè dataset nan chak kliyan and is used to weigh each client’s contributions when aggregating the model weights in federated averaging. During the FL training task, each client site selects its best local model by tracking the model’s performance on its local validation set. At the same time, the server determines the best global model based on the average validation scores sent from each client site to the server after each FL round. After FL training finishes, the best local models and the best global model are automatically shared with all client sites and evaluated on their local test data. 1 T t K Nwa k Nan fòmasyon sou done lokal sèlman (baseline a), nou mete nimewo a epòk nan 200. Optimize a Adam te itilize pou tou de fòmasyon lokal ak FL ak yon pousantaj aprantisaj inisyal de 5 × 10-5 ak yon pousantaj aprantisaj etap ak yon faktori 0.5 apre chak 40 epòk, ki se enpòtan pou konvergans la nan mwayèn federated Random affine transformations, ki gen ladan rotasyon, tradiksyon, koupe, scaling ak son ak chanjman nan entansite alantou, yo te aplike nan imaj yo pou agrandi done pandan fòmasyon an. 73 Akòz sensibilite a nan layers nan BN lè fè fas ak kliyan diferan nan yon anviwònman ki pa enpòtan ak idènman distribiye, nou te jwenn pi bon pèfòmans modèl rive lè kenbe ResNet34 prétrainé ak atansyon espesifik paramèt fixe pandan fòmasyon FL (ki se lè l sèvi avèk yon vitès aprantisaj nan zèb pou sa ki layers). Yon rezo Deep & Cross ki konbine karakteristik imaj ak karakteristik EMR pa gen layers BN ak se konsa pa afekte pa pwoblèm nan instability BN. 58 47 Nan etid sa a, nou te etidye yon sistèm pwoteksyon prive ki pataje sèlman ajou modèl pati ant sèvè ak sit kliyan. Ajou pousantaj yo te rankontre pandan chak iterasyon pa gwosè a nan kontribisyon, ak sèlman yon sèten pousantaj nan ajou pousantaj pi gwo te pataje ak sèvè a. Pou egzanp, ajou pousantaj (ki rele tou gradients) te pataje sèlman si valè absoli yo te pi wo pase yon sèten pousantaj percentil, (t) (Extended Data Fig. ), ki te kalkile soti nan tout gradients ki pa zèb, Δ , epi yo ka diferan pou chak kliyan Nan chak ronde nan FL . Variations of this scheme could include additional clipping of large gradients or differential privacy schemes ki ajoute zwazo aleksyonèl nan gradients, oswa menm nan done a brik, anvan manje nan rezo a . k 5 Pwodwi pou Telefòn k t 49 51 Analiz estatistik Nou te fè yon Wilcoxon signed-ranking tès pou konfime enpòtan nan amelyorasyon nan pèfòmans observe ant modèl la fòmasyon lokalman ak modèl la FL pou 24 ak 72 èdtan (Fig. and Extended Data Fig. ) Hypothèse a null te respekte ak yon sèl-sidè « 1 × 10–3 nan tou de ka. 2 1 P Pearson korelasyon te itilize pou evalye generalizability (robustity nan valè a AUC mwayèn nan done tès nan lòt sit kliyan) nan modèl lokalman fòmasyon konpare ak gwosè dataset lokal yo. Se sèlman yon korelasyon modere te observe ( = 0,43, nan = 0.035, degre libète (df) = 17 pou modèl la 24 èdtan ak = 0.62, = 0.003, df = 16 for the 72-h model). This indicates that dataset size alone is not the only factor determining a model’s robustness to unseen data. r P r P Pou konpare curves ROC soti nan modèl la FL mondyal ak modèl lokal fòmasyon nan diferan sit (Extended Data Fig. ), nou bootstrapped 1,000 echantiyon soti nan done yo ak konvèti AUC yo rezilta. Lè sa a, nou kalkil diferans ant de seri ak estandaize lè l sèvi avèk fòmil la = (AUC1 – AUC2) Ki kote is the standardized difference, is the standard deviation of the bootstrap differences and AUC1 and AUC2 are the corresponding bootstrapped AUC series. By comparing with normal distribution, we obtained the valè ki montre nan Tables Supplementary Rezilta yo montre ke hypothesis la null te respekte ak trè ba values, indicating the statistical significance of the superiority of FL outcomes. The computation of valè yo te fèt nan R ak bibliyotèk la pROC . 3 D s D s D P 2 P P 74 Since the model predicts a discrete outcome, a continuous score from 0 to 1, a straightforward calibration evaluation such as a qqplot is not possible. Hence, for a quantified estimate of calibration we quantified discrimination (Extended Data Fig. ). Nou te fè yon sèl-way analiz nan varyasyon (ANOVA) tès yo konpare pousantaj modèl lokal ak FL nan mitan kat kategori fondasyonal verite (RA, LFO, HFO, MV). -statistike, ki te kalkil kòm varyasyon an ant echantiyon an vle di divize pa varyasyon an nan echantiyon yo ak reprezante degre a nan dispersion ant gwoup diferan, te itilize pou kwantifye modèl yo. Rezilta nou yo montre ke -valè nan senk sit lokal diferan se 245.7, 253.4, 342.3, 389.8 ak 634.8, pandan y ap valè nan modèl la FL se 843.5. -valè vle di ke gwoup yo se plis separe, pousantaj yo soti nan modèl FL nou an klè montre yon dispersion pi gwo ant kat kategori fondamantal verite. Anplis de sa, value of the ANOVA test on the FL model is <2 × 10–16, indicating that the FL prediction scores are statistically significantly different among the different prediction classes. 10 F F F P Reporting Summary plis enfòmasyon sou konsepsyon rechèch disponib nan linked to this article. Nature Research Reporting Summary Disponibilite nan done The dataset from the 20 institutes that participated in this study remains under their custody. These data were used for training at each of the local sites and were not shared with any of the other participating institutions or with the federated server, and they are not publicly available. Data from the independent validation sites are maintained by CAMCA, and access can be requested by contacting Q.L. Based on determination by CAMCA, a data-sharing review and amendment of IRB for research purposes can be conducted by MGB research administration and in accordance with MGB IRB and policy. Kòd Disponibilite Tout kòd ak lojisyèl ki itilize nan etid sa a yo piblikman disponib nan NGC. Pou aksè, log kòm yon envite oswa kreye yon pwofil Lè sa a, antre youn nan URL yo anba a. Modèl la fòmasyon, direksyon pou preparasyon done, kòd pou fòmasyon, validasyon tès modèl la, dosye readme, direksyon enstalasyon ak dosye lisans yo piblikman disponib nan NVIDIA NGC : Yon lojisyèl fòmasyon federated ki disponib kòm yon pati nan Clara Train SDK: Alternativman, sèvi ak komando sa a download modèl la "wget --content-disposition -O clara_train_covid19_exam_ehr_xray_1.zip”. 61 https://ngc.nvidia.com/catalog/models/nvidia:med:clara_train_covid19_exam_ehr_xray https://ngc.nvidia.com/catalog/containers/nvidia:clara-train-sdk https://api.ngc.nvidia.com/v2/models/nvidia/med/clara_train_covid19_exam_ehr_xray/versions/1/zip Referans Budd, J. et al. Digital technologies in the public-health response to COVID-19. , 1183–1192 (2020). Nat. Med. 26 Moorthy, V., Henao Restrepo, A. M., Preziosi, M.-P. & Swaminathan, S. Data sharing for novel coronavirus (COVID-19). , 150 (2020). Bull. World Health Organ. 98 Chen, Q., Allot, A. & Lu, Z. Keep up with the latest coronavirus research. , 193 (2020). Nature 579 Fabbri, F., Bhatia, A., Mayer, A., Schlotter, B. & Kaiser, J. BCG IT spend pulse: how COVID-19 is shifting tech priorities. (2020). https://www.bcg.com/publications/2020/how-covid-19-is-shifting-big-it-spend Candelon, F., Reichert, T., Duranton, S., di Carlo, R. C. & De Bondt, M. The rise of the AI-powered company in the postcrisis world. (2020). https://www.bcg.com/en-gb/publications/2020/business-applications-artificial-intelligence-post-covid Chao, H. et al. Integrative analysis for COVID-19 patient outcome prediction. , 101844 (2021). Med. Image Anal. 67 Zhu, X. et al. Joint prediction and time estimation of COVID-19 developing severe symptoms using chest CT scan. , 101824 (2021). Med. Image Anal. 67 Yang, D. et al. Federated semi-supervised learning for Covid region segmentation in chest ct using multi-national data from China, Italy, Japan. , 101992 (2021). Med. Image Anal. 70 Minaee, S., Kafieh, R., Sonka, M., Yazdani, S. & Jamalipour Soufi, G. Deep-COVID: predicting COVID-19 from chest X-ray images using deep transfer learning. , 101794 (2020). Med. Image Anal. 65 COVID-19 Studies from the World Health Organization Database. (2020). https://clinicaltrials.gov/ct2/who_table ACTIV. (2020). https://www.nih.gov/research-training/medical-research-initiatives/activ Coronavirus Treatment Acceleration Program (CTAP). US Food and Drug Administration (2020). https://www.fda.gov/drugs/coronavirus-covid-19-drugs/coronavirus-treatment-acceleration-program-ctap Gleeson, P., Davison, A. P., Silver, R. A. & Ascoli, G. A. A commitment to open source in neuroscience. , 964–965 (2017). Neuron 96 Piwowar, H. et al. The state of OA: a large-scale analysis of the prevalence and impact of open access articles. , e4375 (2018). PeerJ. 6 European Society of Radiology (ESR). What the radiologist should know about artificial intelligence – an ESR white paper. , 44 (2019). Insights Imaging 10 Pesapane, F., Codari, M. & Sardanelli, F. Artificial intelligence in medical imaging: threat or opportunity? Radiologists again at the forefront of innovation in medicine. , 35 (2018). Eur. Radiol. Exp. 2 Price, W. N. 2nd & Cohen, I. G. Privacy in the age of medical big data. , 37–43 (2019). Nat. Med. 25 Liang, W. et al. Development and validation of a clinical risk score to predict the occurrence of critical illness in hospitalized patients with COVID-19. , 1081–1089 (2020). JAMA Intern. Med. 180 Wynants, L. et al. Prediction models for diagnosis and prognosis of covid-19 infection: systematic review and critical appraisal. , m1328 (2020). Brit. Med. J. 369 Zhang, L. et al. D-dimer levels on admission to predict in-hospital mortality in patients with Covid-19. , 1324–1329 (2020). J. Thromb. Haemost. 18 Sands, K. E. et al. Patient characteristics and admitting vital signs associated with coronavirus disease 2019 (COVID-19)-related mortality among patients admitted with noncritical illness. (2020). https://doi.org/10.1017/ice.2020.461 American College of Radiology. CR recommendations for the use of chest radiography and computed tomography (CT) for suspected COVID-19 infection. (2020). https://www.acr.org/Advocacy-and-Economics/ACR-Position-Statements/Recommendations-for-Chest-Radiography-and-CT-for-Suspected-COVID19-Infection Rubin, G. D. et al. The role of chest imaging in patient management during the COVID-19 pandemic: a multinational consensus statement from the Fleischner Society. , 172–180 (2020). Radiology 296 World Health Organization. Use of chest imaging in COVID-19. (2020). https://www.who.int/publications/i/item/use-of-chest-imaging-in-covid-19 Jamil, S. et al. Diagnosis and management of COVID-19 disease. , 10 (2020). Am. J. Respir. Crit. Care Med. 201 Redmond, C. E., Nicolaou, S., Berger, F. H., Sheikh, A. M. & Patlas, M. N. Emergency radiology during the COVID-19 pandemic: The Canadian Association of Radiologists Recommendations for Practice. , 425–430 (2020). Can. Assoc. Radiologists J. 71 Buch, V. et al. Development and validation of a deep learning model for prediction of severe outcomes in suspected COVID-19 Infection. Preprint at (2021). https://arxiv.org/abs/2103.11269 Lyons, C. & Callaghan, M. The use of high-flow nasal oxygen in COVID-19. , 843–847 (2020). Anaesthesia 75 Whittle, J. S., Pavlov, I., Sacchetti, A. D., Atwood, C. & Rosenberg, M. S. Respiratory support for adult patients with COVID-19. , 95–101 (2020). J. Am. Coll. Emerg. Physicians Open 1 Ai, J., Li, Y., Zhou, X. & Zhang, W. COVID-19: treating and managing severe cases. , 370–371 (2020). Cell Res. 30 Esteva, A. et al. A guide to deep learning in healthcare. , 24–29 (2019). Nat. Med. 25 Cahan, E. M., Hernandez-Boussard, T., Thadaney-Israni, S. & Rubin, D. L. Putting the data before the algorithm in big data addressing personalized healthcare. , 78 (2019). NPJ Digit. Med. 2 Thrall, J. H. et al. Artificial intelligence and machine learning in radiology: opportunities, challenges, pitfalls, and criteria for success. , 504–508 (2018). J. Am. Coll. Radiol. 15 Shilo, S., Rossman, H. & Segal, E. Axes of a revolution: challenges and promises of big data in healthcare. , 29–38 (2020). Nat. Med. 26 Gao, Y. & Cui, Y. Deep transfer learning for reducing health care disparities arising from biomedical data inequality. , 5131 (2020). Nat. Commun. 11 Rieke, N. et al. The future of digital health with federated learning. , 119 (2020). NPJ Dig. Med. 3 Yang, Q., Liu, Y., Chen, T. & Tong, Y. Federated machine learning: concept and applications. , 12 (2019). ACM Trans. Intell. Syst. Technol. 10 Ma, C. et al. On safeguarding privacy and security in the framework of federated learning. , 242–248 (2020). IEEE Netw. 34 Brisimi, T. S. et al. Federated learning of predictive models from federated Electronic Health Records. , 59–67 (2018). Int. J. Med. Inform. 112 Roth, H. R. et al. Federated learning for breast density classification: a real-world implementation. In , (eds. Albarqouni, S. et al.) Vol. 12,444, 181–191 (Springer International Publishing, 2020). Proc. Second MICCAI Workshop, DART 2020 and First MICCAI Workshop, DCL 2020 Domain Adaptation and Representation Transfer, and Distributed and Collaborative Learning Sheller, M. J. et al. Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. , 12598 (2020). Sci. Rep. 10 Remedios, S. W., Butman, J. A., Landman, B. A. & Pham, D. L. in (eds Remedios, S. W. et al.) (Springer, 2020). Federated Gradient Averaging for Multi-Site Training with Momentum-Based Optimizers Xu, Y. et al. A collaborative online AI engine for CT-based COVID-19 diagnosis. Preprint at (2020). https://www.medrxiv.org/content/10.1101/2020.05.10.20096073v2 Raisaro, J. L. et al. SCOR: A secure international informatics infrastructure to investigate COVID-19. , 1721–1726 (2020). J. Am. Med. Inform. Assoc. 27 Vaid, A. et al. Federated learning of electronic health records to improve mortality prediction in hospitalized patients with COVID-19: machine learning approach. , e24207 (2021). JMIR Med. Inform. 9 Nino, G. et al. Pediatric lung imaging features of COVID-19: a systematic review and meta-analysis. , 252–263 (2021). Pediatr. Pulmonol. 56 Fredrikson, M., Jha, S. & Ristenpart, T. Model inversion attacks that exploit confidence information and basic countermeasures. In 1322–1333, (2015). Proc. 22nd ACM SIGSAC Conference on Computer and Communications Security https://doi.org/10.1145/2810103.2813677 Zhu, L., Liu, Z. & Han, S. in (eds Wallach, H. et al.) 14774–14784 (Curran Associates, Inc., 2019). Advances in Neural Information Processing Systems 32 Kaissis, G. A., Makowski, M. R., Rückert, D. & Braren, R. F. Secure, privacy-preserving and federated machine learning in medical imaging. , 305–311 (2020). Nat. Mach. Intell. 2 Li, W. et al. in 133–141 (Springer, 2019). Privacy-Preserving Federated Brain Tumour Segmentation Shokri, R. & Shmatikov, V. Privacy-preserving deep learning. In (2015). Proc. 53rd Annual Allerton Conference on Communication, Control, and Computing (Allerton) https://doi.org/10.1109/allerton.2015.7447103 Li, X. et al. Multi-site fMRI analysis using privacy-preserving federated learning and domain adaptation: ABIDE results. , 101765 (2020). Med. Image Anal. 65 Estiri, H. et al. Predicting COVID-19 mortality with electronic medical records. , 15 (2021). NPJ Dig. Med. 4 Jiang, G. et al. Harmonization of detailed clinical models with clinical study data standards. , 65–74 (2015). Methods Inf. Med. 54 Yang, D. et al. in . (2019). Searching Learning Strategy with Reinforcement Learning for 3D Medical Image Segmentation https://doi.org/10.1007/978-3-030-32245-8_1 Elsken, T., Metzen, J. H. & Hutter, F. Neural architecture search: a survey. , 1–21 (2019). J. Mach. Learning Res. 20 Yao, Q. et al. Taking human out of learning applications: a survey on automated machine learning. Preprint at (2019). https://arxiv.org/abs/1810.13306 Ioffe, S. & Szegedy, C. Batch normalization: accelerating deep network training by reducing internal covariate shift. In , PMLR , 448–456 (2015). Proc. 32nd International Conf. Machine Learning 37 Kaufman, S., Rosset, S. & Perlich, C. Leakage in data mining: formulation, detection, and avoidance. In , 556–563 (2011). Proc. 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining Zhang, C. et al. BatchCrypt: efficient homomorphic encryption for cross-silo federated learning. In , 493–506 (2020). Proc. 2020 USENIX Annual Technical Conference, ATC 2020 . (2020). Nvidia NGC Catalog: COVID-19 Related Models https://ngc.nvidia.com/catalog/models?orderBy=scoreDESC&pageNumber=0&query=covid&quickFilter=models&filters Marini, J. J. & Gattinoni, L. Management of COVID-19 respiratory distress. , 2329–2330 (2020). JAMA 323 Cook, T. M. et al. Consensus guidelines for managing the airway in patients with COVID-19: Guidelines from the Difficult Airway Society, the Association of Anaesthetists the Intensive Care Society, the Faculty of Intensive Care Medicine and the Royal College of Anaesthetist. , 785–799 (2020). Anaesthesia 75 Galloway, J. B. et al. A clinical risk score to identify patients with COVID-19 at high risk of critical care admission or death: an observational cohort study. , 282–288 (2020). J. Infect. 81 Kilaru, A. S. et al. Return hospital admissions among 1419 COVID-19 patients discharged from five U.S. emergency departments. , 1039–1042 (2020). Acad. Emerg. Med. 27 He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In (2016). Proc. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) https://doi.org/10.1109/cvpr.2016.90 Irvin, J. et al. CheXpert: a large chest radiograph dataset with uncertainty labels and expert comparison. , 590–597 (2019). Proc. AAAI Conf. Artif. Intell. 33 Wang, R., Fu, B., Fu, G. & Wang, M. Deep & Cross network for Ad Click predictions. In Article no. 12 (2017). Proc. ADKDD’17 Abadi, M. et al. TensorFlow: asystem for large-scale machine learning. In , USENIX Association 265–283 (2016). 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16) . (2020). NVIDIA Clara Imaging https://developer.nvidia.com/clara-medical-imaging Stekhoven, D. J. & Bühlmann, P. MissForest–non-parametric missing value imputation for mixed-type data. , 112–118 (2012). Bioinformatics 28 McMahan, H., Moore, E., Ramage, D., Hampson, S. & y Arcas, B. A. Communication-efficient learning of deep networks from decentralized data. (2017). http://proceedings.mlr.press/v54/mcmahan17a.html Hsieh, K., Phanishayee, A., Mutlu, O. & Gibbons, P. B. The non-IID data quagmire of decentralized machine learning. In PMLR 119 (2020). Proc. 37th International Conf. Machine Learning Robin, X. et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. , 77 (2011). BMC Bioinformatics 12 rekonesans The views expressed in this study are those of the authors and not necessarily those of the NHS, the NIHR, the Department of Health and Social Care or any of the organizations associated with the authors. MGB thank the following individuals for their support: J. Brink, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, MA; M. Kalra, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, MA; N. Neumark, Center for Clinical Data Science, Massachusetts General Brigham, Boston, MA; T. Schultz, Department of Radiology, Massachusetts General Hospital, Boston, MA; N. Guo, Center for Advanced Medical Computing and Analysis, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, MA; J. K. Cramer, Director, QTIM lab at the Athinoula A. Martinos Center for Biomedical Imaging at MGH; S. Pomerantz, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, MA; G. Boland, Department of Radiology, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA; W. Mayo-Smith, Department of Radiology, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA. UCSF thank P. B. Storey, J. Chan and J. Block for implementing the UCSF FL client infrastructure, and W. Tellis for providing the source imaging repository for this work. The UCSF EMR and clinical notes for this study were accessed via the COVID-19 Research Data Mart, . atravè Fakilte a Medsin, Chulalongkorn Inivèsite danje Ratchadapisek Sompoch Endowment Fund RA (PO) (no. 001/63) pou koleksyon ak jesyon nan done klinik ak echantiyon biolojik ki gen rapò ak COVID-19 pou Task Force Rechèch la, Fakilte a Medsin, Chulalongkorn Inivèsite. NIHR Cambridge Biomedical Research Centre danje A. Priest, ki se sipòte pa NIHR (Cambridge Biomedical Research Centre nan Cambridge University Hospitals NHS Foundation Trust). National Taiwan University MeDA Lab ak MAHC ak Taiwan National Health Insurance Administration danje MOST Joint Research Center for AI technology, All Vista Healthcare National Health Insurance Administration, Taiwan, Ministry of Science and Technology, ak Taiwan National Research Center for Theoretical https://data.ucsf.edu/covid19 Papye sa a se disponib nan natirèl anba lisans CC by 4.0 Deed (Attribution 4.0 entènasyonal). Papye sa a se anba CC by 4.0 Deed (Attribution 4.0 entènasyonal) lisans. Disponib nan natirèl
