Umbhali: Ittai Dayan Holger R. Roth Aoxiao Zhong Ahmed Harouni Amilcare Gentili Anas Z. Abidin U-Andrei Liu Anthony Beardsworth Costa Bradford J. Wood Chien-Sung Tsai Chih-Hung Wang U-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 Ukucinga Jason C. Crane Jesse Tetreault Jiahui Guan John W. Garrett Joshua D. Kaggie Jung Gil Park U-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 Umbhali: Ukucinga Holger R. Roth I-Aoxiao Zhong U-Ahmad Harouni Amilcare Gentili U-Anas Z. Abidin U-Andrei Liu U-Anthony Beardsworth Costa U-Bradford J. Wood I-Chien-Sung Tsai I-Chih-Hung Wang U-Chun-Nan Hsu C. K. Lee Ukucinga Ruan Ngathi Ukucinga U-Edy Huang UFELIPE CAMPOS KITAMURA I-Griffin Lacey I-Gustavo César de Antônio Corradi U-Gustavo Nino Ukucinga I-Hirofumi Obinata Ukucinga U-Jason C. Crane U-Jesse Tetreault I-Jiahui Guan UJohn W. Garrett UJoshua D. Kaggie I-Jung Gil Park U-Keith Dreyer I-Krishna Juluru U-Christopher Kersten U-Marcio Aloisio Bezerra Cavalcanti Rockenbach U-Marius George Linguraru I-Masoom A. Haider Ukucinga Nicole Rieke U-Pablo F. Damasceno U-Pedro Mario Cruz e Silva Ukucinga Wang Ukucinga I-Shuichi Kawano U-Sira Sriswasdi I-Soo Young Park uThomas M. Grist Umbhali I-Watsamon Yantarabenjakul Ukucinga Wang I-Won Young Tak Ngathi Xihong Lin I-Young Joon Kwon Ukucinga U-Andrei Feng Andrew N. Priest I-Baris Turkbey Benjamin Glicksberg U-Bernardo Bizzo Ukucinga kweKim UCarlos Tor-Díez I-Chi-Cheng Lee I-Chia-Jung Hsu Ukucinga I-Chu-Ling Lai uChristopher P. Hess U-Colin Compas I-Deepeksha Bhatia U-Eric K. Oermann U-Evan Leibovitz I-Hisashi Sasaki I-Hitoshi Mori U-Isaac Yang U-Jae Ho Sohn I-Krishna Nand Keshava Murthy U-Chen Fu U-Mateus Ribeiro Furtado de Mendonça Mike Fralick I-Kyu Kang Mohammad Adil U-Natalie Gangai Iimveliso ze-Vateekul uPier Elnajjar U-Sarah Hickman Sharmila Majumdar Shelley L. McLeod U-Sheridan Reed Stefan Gräf U-Stephanie Harmon Tatsuya Kodama Ukucinga U-Tony Mazzulli U-Vitor Lima de Trabor I-Yothin Rakvongthai Yu Rim Lee Ngathi wen U-Fiona J. Gilbert I-Mona G. Flores Ukucinga Li Ukucinga I-Federated Learning Setup (FL) yedatha yindlela esetyenziselwa ukuqeqesha iimodeli ye-intelligence ye-artificial intelligence kunye needatha ezininzi kwithuba ukugcina i-anonymity yedatha, ngoko ukunciphisa iingxaki ezininzi kwi-data sharing. Ngiya kusetyenziswa idatha evela kumaziko angu-20 emhlabeni wonke ukuqeqesha iimodeli ye-FL, ebizwa ngokuba yi-EXAM (i-electronic medical record (EMR) chest X-ray AI model), leyo ibonise iimfuno ye-oxygen yexesha elandelayo zeengxaki ze-COVID-19 usebenzisa iingxaki zeengxaki zeengcali, iinkcukacha ze-laboratory kunye ne-crystal X-rays. I- Ukucinga Iimpawu zenzululwazi, zenzulwazi, zonyango kunye nezifundo zenzulwazi zihlanganisa ngokushesha iintlobo ezintsha zenzululwazi (i-AI) ezisetyenziswa ngokukhawuleza kunye nokhuseleko, kunye nokukhuthaza ukuxhaswa kwedatha kunye nokufundisa iimodeli ngaphandle kweengxaki zokusetyenziswa kwe-privacy kunye ne-data ownership ze-collaborations ezivamile. , Iinkonzo ze-Healthcare, i-Researchers kunye ne-Industry ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye zibe. , , , , , , Ukubalwa kwizifundo zonyango lwezilwanyana lula kwaye lula iinkqubo zokusebenza zendawo kunye ne-spirit of international cooperation , , Izifundo ze-Data Analytics kunye ne-AI ziquka ngokuzenzekelayo iinkqubo ezivumileyo kunye ne-collaborative, zihlanganisa iinkcukacha ezifana ne-open-source software, i-replicable research, i-data repositories kunye nokufumana iinkcukacha ze-anonymized ngokubanzi. , I-pandemic yandisa ukuba kufuneka ukuqhagamshelane ngokushesha iinkolelo zeendatha eziholise izixeko ze-clinical kunye ne-scientific ekuphenduleni iingxaki zehlabathi ezincinane kunye ne-pandemic. I-data sharing inesibopheleni ze-ethical, i-regulatory kunye ne-legal eyenziwe, kwaye ingakumbi i-complicated, ngexesha elidlulileyo yeengcali ze-tech kwihlabathi ye-healthcare data. , , . 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Umzekelo olufanelekileyo yintlobo zokusebenziswano lwezinto iinkqubo zethu ezidlulileyo kwi-AI-based SARS-COV-2 clinical decision support (CDS) model. Le model ye-CDS yenzelwe e-Mass General Brigham (MGB) kwaye ilawulwe kwiinkqubo ezininzi zempilo. Iingxaki ze-CDS ziye zithunyelwe kwiimifanekiso ze-X-ray (CXR) ye-chest, iingxaki ze-vital, iinkcukacha ze-demographic kunye neeyunithi ze-laboratory ezidlulileyo zibonakalisa imiphumo yeempembelelo ze-COVID-19 , , , . CXR was selected as the imaging input because it is widely available and commonly indicated by guidelines such as those provided by ACR , the Fleischner Society I-WHO I-National Thoracic Societies I-COVID Handbooks kunye ne-Radiology Societies ehlabathini lonke Ukusuka kwimodeli ye-CDS i-score, ebizwa ngokuba yi-CORISK , that corresponds to oxygen support requirements and that could aid in triaging patients by frontline clinicians , , Iinkonzo zonyango ziye zaziwa ukuba zithanda iimodeli ezidlulileyo ezidlulileyo kwiinkcukacha zayo. Kwixesha elide, iimodeli ezininzi ze-AI, kubandakanya imodeli ye-CDS ebangelwa kwangaphambili, zilungiswa kunye nokuvumelekiswa kwi-data ye-"ngqongileyo" ezininzi ezininzi. , , okukhusela ekugqibeleni kunye nokunciphisa ngokugqithisileyo. Oku kungenziwa ukunciphisa ngokulungiselela idatha ezahlukeneyo ezivela kwiindawo ezininzi ngaphandle kokucubungula idatha ukusetyenziswa kwezindlela ezifana transfer learning , okanye FL. I-FL yindlela esetyenziselwa ukuqeqesha iimodeli ze-AI kwiimveliso zeendaba ezahlukileyo, ngaphandle kokuthunyelwe okanye zithunyelwe ngaphandle kwindawo yayo yokuqala. Nangona ifumaneka kwiimveliso ezininzi, i-FL yaziwa ezidlulileyo kwi-investigation ye-healthcare ye-interinstitutional . 18 19 20 21 22 23 24 25 26 27 28 29 30 27 31 32 33 34 35 36 Ukufundwa kweFederated inikezela ukuqalisa ngokukhawuleza iingcebiso ze-centrally orchestrated kunye nokusetyenziswa kweedatha kunye nokusetyenziswa kwezingxaki ze-algorithmic kunye ne-impact Enye indlela ye-FL, eyaziwa ngokuba yi-client-server, inikeza isampuli ye-"untrained" kwiinkonzo ezininzi (i-"nodes") ezinikezela imisebenzi ze-training, ngokugqithisileyo inikeza iziphumo ziye ziye zithunyelwa kwi-central (i-"federated") server. Lokhu kubandakanyeka njenge-process ye-iterative kuze kube nokugqithwa kwe-training. . 37 36 I-Governance yeendaba ye-FL ibekwe kwi-local, ukunciphisa iingxaki ze-privacy, kunye neengxaki ze-model kuphela okanye i-gradients ezihambelana phakathi kwindawo ze-client kunye ne-server ye-federated. , . I-FL iye ibonelela kwiinkqubo ezidlulileyo zonyango zonyango , , , , kuquka kwi-COVID-19 analysis , , Umzekelo olufanelekileyo umzekelo umdlavuza kwizigulane afanelekileyo nge-SARS-COV-2 esebenzisa iimpawu zonyango, nangona kufutshane ngamanani kwezindlela kunye ne-scale. . 38 39 40 41 42 43 8 44 45 46 Indawo yethu yaba ukuvelisa isampuli eshushu, esebenzayo enokuncedisa enokuncedisa izigulane. Thina siphinde ukuba isampuli ye-CDS ingasetyenziswa ngempumelelo ngenxa yokusetyenziswa kweendaba ezininzi ezininzi ezininzi kwimeko ye-clinical yaye asekuseni kakhulu kwi-operator-dependent assessments of patient condition (njenge-impressions ze-clinical or symptoms). Ngoko ke, iziphumo ze-laboratory, iingcebiso ze-vital, i-imaging study kunye ne-demographic eyenziwe ngokubanzi (i-age), ezisetyenziswa. Ngoko ke, siqhelise isampuli ye-CDS kunye needatha ezininzi usebenzisa indlela ye-client- I-hypothesis yethu yaba ukuba i-EXAM iyafanelekileyo kunokuba iimodeli zangaphakathi kunye nokuvala ngakumbi kwiinkqubo zempilo. Imiphumo I-Exam Model Architecture I-EXAM model isekelwe kwi-CDS model ethandwa phezulu Ukusetyenziswa kwimodeli, i-20 iimpawu (i-19 evela kwi-EMR kunye ne-CXR) ziye zisetyenziswa njenge-input. I-etiquette ye-resultat (i-‘ground truth’) ziye zithunyelwa ngokutsho kwe-patient oxygen therapy emva kwexesha le-24 kunye ne-72 ngehora ukusuka kwisebe se-emergency (ED). Uluhlu oluthile lwezimpawu kunye ne-results ezivela kwi-Table . 27 1 Izixhobo ze-oxygen therapy ziquka i-0, 0,25, 0,50 kunye ne-0,75 ngokuxhomekeke ne-oxygen therapy eyenziwe ngexabiso. Izixhobo ze-oxygen therapy ziquka i-room air (RA), i-low-flow oxygen (LFO), i-high-flow oxygen (HFO)/ventilation non-invasive (NIV) okanye i-ventilation mechanical (MV). Ukuba umdla we-anticipation window, i-etiquetage ye-resultat yenzelwe yi-1. Oku kwenza ngexabelo ezimbini kwi-interval 0-1 (i- 24 kunye ne-72 iiyure). Kwimfuneko ye-EMR, kuphela izibalo ezimbini ezivela kwi-ED ezisetyenziswa kwaye i-data preprocessing iquka i-deidentification, i-imputation ye-value ezikhoyo kunye ne-normalization ukuya kwi-zero-median kunye ne-unity variance. Kwiimifanekiso ze-CXR, kuphela iimifanekiso ezimbini ezivela kwi-ED ezisetyenziswa. Umzekelo umzekelo umzekelo umzekelo umzekelo we-EMR kunye ne-CXR, usebenzisa i- 34-layer convolutional neural network (ResNet34) ukufumana iimpawu kwi-CXR kunye ne-Deep & Cross network ukuze i-concatenate iimpawu kunye ne-EMR features. I-model output yi-risk score, ebizwa ngokuba yi-EXAM score, nto leyo i-value epheleleyo kwinqanaba ye-0-1 ngalinye kwi-24 kunye ne-72 iiyure ze-predictions ezinxulumene ne-labels ezaziwayo. Ukucinga Ukuqhagamshelwano Model Imodeli ye-EXAM yenzelwe ukusetyenziswa kwe-cohort ye-16148 iimeko, okwenza nje phakathi kweemodeli ezimbini ze-FL ze-COVID-19 kodwa nangona inkqubo enkulu kunye ne-multicontinental yokusebenza kwi-AI esebenzayo kwi-clinically (i-Fig. Data phakathi iindawo awungabonakaliwa phambi kokuveliswa, kwaye, phantse imeko real-life zonyango lwezobugcisa, i-harmonization eziqhelekileyo zentlawulo idatha awungasebenzi abenzi (i-Fig. ). 1A,B 1C,D , World map ibonisa i-20 iindawo ezahlukeneyo ze-client ezinikezela uphando lwe-EXAM. , Inani lwezimali ezisetyenziswa ngalinye le-institution okanye i-site (i-client 1 ibekwe kwi-site ebonakalisa inani elikhulu lwezimali). , X-ray X-ray Intensity ukusabalalisa ngalinye indawo client. Ukubonisa ubude obuncinane kunye ne-maximum (i-asterisks), ubude obuncinane (i-triangles) kunye ne-deviation standard (i-bars horizontal). Inani lwesampuli ye-customer site ifumaneka kwi-Table Supplementary . a b c d 1 Iimveliso ze-FL zihlanganisa iimveliso ze-locally trained kunye ne-global FL model kwiimveliso ze-client. Iimveliso ze-FL zihlanganisa ukusebenza okuphumelela kakhulu ( « 1 × 10-3, i-Wilcoxon signed-ranking test) ye-16% (ngokusetyenziswa yi-AUC ephakathi xa isebenze i-model kwiiseti ze-test ye-local: ukusuka kwi-0,795 ukuya kwi-0,920, okanye i-12.5 iiponyo yentlawulo) (Imi. ). Kwakhona kubandakanya ukuveliswa kwe-38% ye-generalizability (ngokuqhelekileyo yi-AUC ephakathi xa isebenze i-model kwiinkqubo ezininzi ze-test sets: ukusuka kwi-0.667 ukuya kwi-0.920, okanye i-25.3 iiponyo yentlawulo) ye-model engcono yehlabathi yokubonisa ukwelashwa kwe-oxygen ye-24 iiyure kunokuba yi-models ezidlulileyo kuphela kwi-data ye-site (i-Fig. Ukulungiselela iziphumo ze-72 iiyure ze-oxygen treatment, ukuqeqeshwa kwimodeli yehlabathi engcono yandisa i-18% yentlawulo yentlawulo malunga neemodeli ehlabathini, kwaye ukuphuculwa kwimodeli yehlabathi yandisa ngentlawulo ye-34% (I-Extended Data Fig. I-stability yeempembelelo zethu ilawulwa ngokuguqula iintlobo ezintathu ze-local kunye ne-FL kwiingxaki zeendaba ezahlukeneyo ze-randomized. P 2A 2B 1 , Ukusebenza kwimvavanyo yeenkcukacha yeenkcukacha yeenkcukacha yeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha zeenkcukacha ze , I-Generalizability (i-performance average on other sites' test data, njengoko i-AUC average) njenge-function of a client's dataset size (no. of cases). I-green horizontal line ibonisa i-performance generalizability ye-best global model. I-performance ye-18 of 20 clients ifumaneka, njengoko i-client 12 iye yaba imiphumela kuphela kwi-72h oxygen (i-Extended Data Fig. ) kunye ne-client 14 iingxaki kuphela kunye ne-RA ukwelashwa, ngoko i-metric ye-evaluation (kwi-AUC) ayisetyenziswa kwiiyure ezininzi ( Izixhobo ze-client 14 ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye zibe. a b 1 Ukucinga Iimodeli zendawo ezidlulileyo usebenzisa i-cohorts ezisebenzayo (isib. iimeko ezininzi ezincinane ze-COVID-19) ziphumelele kakhulu ukususela kwi-FL, kunye nokuphuculwa kakhulu kwimveliso ye-prediction AUC ephakathi kwizigaba ezininzi. Oku kubonakala kwi-client site 16 (i-dataset ezisebenzayo), kunye neempawu ezininzi ezinxalenye kunye namaxesha ezincinane ezininzi. Iimodeli ye-FL iye yenza i-true-positive rate ephakeme kwiiyure ezimbini ezinzima kunye ne-false-positive rate ezincinane kunokuba yi-modeli ye-local, zombini ezibonakalayo kwi-receiver operating characteristic (ROC) plots Izixhobo ze-Data Fig. Ngokuqhelekileyo, ukuphucula kwimodeli ye-FL kubandakanya kakhulu kwimodeli ebandayo. 3A 2 , ROC at client site 16, with unbalanced data and mostly mild cases. , ROC of the local model at client site 12 (a small dataset), mean ROC of models trained on larger datasets corresponding to the five client sites in the Boston area (1, 4, 5, 6, 8) and ROC of the best global model in prediction of 72-h oxygen treatment for different thresholds of EXAM score (left, middle, right). The mean ROC is calculated based on five locally trained models while the gray area denotes the ROC standard deviation. ROCs for three different cutoff values ( ) of the EXAM risk score are shown. Pos and neg denote the number of positive and negative cases, respectively, as defined by this range of EXAM score. a b t Kwiisayithi ze-client kunye ne-datasets ezincinane, i-FL model engcono kakhulu, ngaphandle kwe-local model kuphela, kodwa kwizithuba ezininzi ze-datasets ezincinane ze-client sites kwi-Boston area ye-USA (i-Fig. ). 3b The global model performed well in predicting oxygen needs at 24/72 h in patients both COVID positive and negative (Extended Data Fig. inguqulelo 3 Validation kwiindawo eyahlukileyo Emva kokufunda lokuqala, i-EXAM yaye ilawulwa kwindawo ezintathu zokusetyenziswa: i-Cooley Dickinson Hospital (CDH), i-Martha's Vineyard Hospital (MVH) kunye ne-Nantucket Cottage Hospital (NCH), zonke e-Massachusetts, e-USA. I-modeli ayikho kwindawo ezininzi kwaye isetyenziswa kuphela ngenxa yokufunda. I-cohort size kunye ne-model inference iimiphumo zithunyelwe kwi-Table. , kunye nemibala ye-ROC kunye ne-matrix ye-confusion ye-dataset eninzi (kusuka kwi-CDH) ibonisa kwi-Fig. . The operating point was set to discriminate between nonmechanical ventilation and mechanical ventilation (MV) treatment (or death). The FL global trained model, EXAM, achieved an average AUC of 0.944 and 0.924 for 24- and 72-h prediction tasks, respectively (Table ), which exceeded the average performance among sites used in training EXAM. For prediction of MV treatment (or death) at 24 h, EXAM achieved a sensitivity of 0.950 and specificity of 0.882 at CDH, and a sensitivity of 1.000 specificity of 0.934 at MVH. NCH did not have any cases with MV/death at 24 h. In regard to 72-h MV prediction, EXAM achieved a sensitivity of 0.929 and specificity of 0.880 at CDH, sensitivity of 1.000 and specificity of 0.976 at MVH and sensitivity of 1.000 and specificity of 0.929 at NCH. 2 4 2 , , Performance (ROC) (top) and confusion matrices (bottom) of the EXAM FL model on the CDH dataset for prediction of oxygen requirement at 24 h ( ) kwaye iiyure ze-72 ( ). I-ROC yeentlobo ze-cutoff ( ) of the EXAM risk score are shown. a b a b t For MV at CDH at 72 h, EXAM had a low false-negative rate of 7.1%. Representative failure cases are presented in Extended Data Fig. , showing two false-negative cases from CDH where one case had many missing EMR data features and the other had a CXR with a motion artifact and some missing EMR features. 4 Ukusetyenziswa kwe-differential privacy I-motivation yokuqala ye-healthcare institutes ukuba usebenzise i-FL yinto ukugcina ukhuseleko kunye ne-privacy yeedatha zabo, kunye nokuxhomekeka kwimeko ye-data compliance. Kwi-FL, kunokuba i-risk ye-model 'inversion' ifumaneka. or even the reconstruction of training images from the model gradients themselves . To counter these risks, security-enhancing measures were used to mitigate risk in the event of data ‘interception’ during site-server communication Izixhobo ze-technology zokuvimbela i-interception ye-FL data, kunye ne-feature ye-security eyenza ukuba sinokufunda iinkalo ezininzi ukuyisebenzisa i-FL. Ngokwenza oku, sinikezela iziphumo ezidlulileyo ezibonisa ukuba i-weight sharing, kunye nezinye i-differential privacy techniques, ingasetshenziswa ngempumelelo kwi-FL. Ukusetyenziswa kwe-Scheme of Partial Weight-Sharing , , Ukubonisa ukuba iimodeli ziyafumaneka ukusebenza efanelekileyo nangona kuphela i-25% ye-weight updates zithunyelwe (Extended Data Fig. ). 47 48 49 50 50 51 52 5 Ukudibanisa This study features a large, real-world healthcare FL study in terms of number of sites and number of data points used. We believe that it provides a powerful proof-of-concept of the feasibility of using FL for fast and collaborative development of needed AI models in healthcare. Our study involved multiple sites across four continents and under the oversight of different regulatory bodies, and thus holds the promise of being provided to different regulated markets in an expedited way. The global FL model, EXAM, proved to be more robust and achieved better results at individual sites than any model trained on only local data. We believe that consistent improvement was achieved owing to a larger, but also a more diverse, dataset, the use of data inputs that can be standardized and avoidance of clinical impressions/reported symptoms. These factors played an important part in increasing the benefits from this FL approach and its impact on performance, generalizability and, ultimately, the model’s usability. For a client site with a relatively small dataset, two typical approaches could be used for fitting a useful model: one is to train locally with its own data, the other is to apply a model trained on a larger dataset. For sites with small datasets, it would have been virtually impossible to build a performant deep learning model using only their local data. The finding, that these two approaches were outperformed on all three prediction tasks by the global FL model, indicates that the benefit for client sites with small datasets arising from participation in FL collaborations is substantial. This is probaby a reflection of FL’s ability to capture more diversity than local training, and to mitigate the bias present in models trained on a homogenous population. An under-represented population or age group in one hospital/region might be highly represented in another region—such as children who might be differentially affected by COVID-19, including disease manifestations in lung imaging . 46 Imiphumo ye-validation ibonise ukuba iimodeli yehlabathi ibonakalayo, ebonakalisa i-hypothesis yethu ukuba iimodeli ezidlulileyo ze-FL ziquka zokusetyenziswa kwiinkqubo zonyango. Zifumaneka kwimeko emangalisayo yokusetyenziswa kwamalgorithms e-COVID-19 kwizigulane, kunye nokusetyenziswa kwe-FL kwiimodeli kunye nokuVavanyelwa. Ngokusetyenziswa kwisifundo se-MGB, iisayithi ze-client ziye ziye zithunyelwa kwi-EXAM, ukuze ziye zithunyelwe ngakumbi ngaphambi kokuqiniswa kwimeko ye-regulatory okanye ukuqhutywa kwizigulane ze-clinical. Iziqonga zangaphakathi z , kwakunye kwiindawo ezahlukeneyo ezikhoyo kwi-EXAM training. 53 Over 200 prediction models to support decision-making in patients with COVID-19 have been published . 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 I-Patient Cohort Identification kunye ne-Data Harmonization ayikho iingxaki ezintsha kwi-research kunye ne-data science Ukuphucula iinkqubo ze-information ze-clinical kufuneka ukuphucula ukucwangcisa ukucwangcisa iinkqubo ze-data, okuholela ekuphumelela kwinethiwekhi ze-website ezihlangene ne-FL. Oku, ngokuxhomekeke ne-hyperparameter engineering, inokukwazi i-algorithms 'ukufunda' ngokukhawuleza kwi-datasets ezininzi kunye nokuguquka iiparametres ze-model kwi-site eyodwa ukucwangcisa ngakumbi - umzekelo, nge-fine-tuning ezininzi kwi-site elandelayo I-system ebonakalisa ukufakelwa kwimodeli kunye nokuthuthukiswa kwimiphumo ngokufanelekileyo, ngokufanelekileyo kunye ne-real-time model processing iya kuthatha ukunyaniseka kwaye "ukutshintshela i-lock" ukusuka kokufunda kwimodeli yokusebenza. 54 39 Ngenxa yokuba idatha ayinxalenye, akayi kufumaneka ngokukhawuleza. Ngoko ke, nayiphi na ukucaciswa kwiziphumo, ngaphezu kwimiphumo ezidlulileyo kunye nokufaka, iya kufumaneka. Njengokuba nezinye iimodeli ze-machine learning, i-EXAM ibandakanya umgangatho yedatha ze-training. Izakhiwo ezinikezele ekubunjweni le-algorithm ye-clinical care kufuneka zibonele iingxaki zokusebenza. Umzekelo, ama-label ezisetyenziswa njenge-ground truth ekubunjweni kwimodeli ye-EXAM zithunyelwe ukusetyenziswa kwe-oxygen kwizigidi ze-24 kunye ne-72 h kwizigidi; ibonakala ukuba i-oxygen ebonakalayo kwinqanaba le-oxygen. Nangona kunjalo, kwi-phase-early of the COVID-19 pandemic, abadlali abaninzi bafumana i-high-flow oxygen prophylactically ngaphandle kweemfuno Since our data access was limited, we did not have sufficient available information for the generation of detailed statistics regarding failure causes, post hoc, at most sites. However, we did study failure cases from the largest independent test site, CDH, and were able to generate hypotheses that we can test in the future. For high-performing sites, it seems that most failure cases fall into one of two categories: (1) low quality of input data—for example, missing data or motion artifact in CXR; or (2) out-of-distribution data—for example a very young patient. Kwiminyaka elandelayo, sinokufuneka ukhawulezela inkqubo ye-population drift ngenxa yeefazi ezahlukileyo zonyango. Sinokufuneka ukuba, ngenxa yokuhlanganisa kwizindawo ezininzi ze-20, le mthengiso ingaphantsi. Iimpawu elithuthukise iintlobo zokusebenza ezininzi zokusetyenziswa yi-customer site ekubuyiselwa ukuphuculwa kwimodeli ye-FL jikelele. Oku kunceda ukwahlula i-customer site, kunye nokuphucula iintloko zokusetyenziswa kwedatha kunye neengxaki zokusetyenziswa. Lezi zilandelayo zihlanganisa kakhulu ngenxa yeengxaki ezininzi kunye neengxaki zeengxaki ezininzi zokusetyenziswa kweengxaki ze-consortium ezininzi, kwaye iyakwazi ukufumana iingxaki ze-diversity ngaphezu kwinqanaba elininzi le-data samples. Future approaches may incorporate automated hyperparameter searching , neural architecture search and other automated machine learning iindlela ukufumana iiparamitha zokufundisa engcono ngalinye indawo ye-client. 55 56 57 Iingxaki ezaziwayo ze-batch normalization (BN) kwi-FL yenza nathi ukuguqulwa kwimodeli yethu yokufakelwa kwicandelo ye-image feature Ukuze ukunciphisa i-divergence phakathi kwizithuba ze-client. Izixhobo zangaphambili ziquka iintlobo ezahlukeneyo ze-normalization zenzululwazi ukufaka iimodeli ze-AI kwi-FL ngokukhawuleza xa idatha ye-client ayikho-independent kwaye ifakwe ngokufanelekileyo. 58 49 Recent works on privacy attacks within the FL setting have raised concerns on data leakage during model training . Meanwhile, protection algorithms remain underexplored and constrained by multiple factors. While differential privacy algorithms , , Ukubonisa ukhuseleko elungileyo, kunokukhawuleza ukusebenza kwimodeli. I-algorithms ye-encryption, njenge-encryption ye-homomorphic Umgangatho we-privacy, ukugcina ukusebenza kodwa unokwandisa kakhulu ubungakanani umxholo kunye nexesha le-training. Umgangatho we-privacy enokuthintela ukhetho olufanelekileyo ekubeni iiparametres ezincinane ze-privacy ezininzi ezininzi ezininzi ezininzi, nangokuthintela ukusebenza kwe-clinically ekubeni. , , . 59 36 48 49 60 36 48 49 Emva kokuphumelela kwakhona, sincoma ukusetyenziswa kwimodeli ye-EXAM kwimeko ye-ED njengoko indlela yokuhlola ingozi kwinqanaba ye-per-patient kunye ne-population, kwaye ukunika i-clinicists i-reference point eyongezelelweyo xa kwenza umsebenzi elininzi elidlulileyo yokuhlola izigulane. Kwakhona sincoma ukusetyenziswa kwimodeli njenge-population-level metric eyongezelelweyo ukunceda ukunciphisa iinkcukacha phakathi kwizixeko, izixeko kunye nemasipala. Sincoma yethu ukuba iimpumelelo ye-FL efana ne-data silos kwaye kukuvumela ukuvelisa iimodeli ze-AI ezininzi ezininzi ezininzi ezininzi. Methods Ethics Ukuvumelana 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). Iingxelo ze-MI-CLAIM zokusetyenziswa kwimodeli ze-AI ye-clinical (i-Supplementary Note) ) 2 Ukucinga setting Ukuhlolwa kubandakanya idatha evela kumaziko 20 (i-Fig. ): MGB, MGH, Brigham and Women's Hospital, Newton-Wellesley Hospital, North Shore Medical Center and Faulkner Hospital; Children's National Hospital in Washington, DC; NIHR Cambridge Biomedical Research Centre; I-Self-Defense Forces Central Hospital eTokyo; National Taiwan University MeDA Lab and MAHC and Taiwan National Health Insurance Administration; Tri-Service General Hospital in Taiwan; Kyungpook National University Hospital in South Korea; Faculty of Medicine, Chulalongkorn University in Thailand; Diagnosticos da America SA in Brazil; University of California, San Francisco; VA San Diego; University of Toronto; National Institutes of Health in Bethesda, Maryland; University of Wisconsin-Madison School of Medicine and Public Health; Memorial Sloan Kettering Cancer Center in New York; kunye , , . Data from three independent sites were used for independent validation: CDH, MVH and NCH, all in Massachusetts, USA. These three hospitals had patient population characteristics different from the training sites. The data used for the algorithm validation consisted of patients admitted to the ED at these sites between March 2020 and February 2021, and that satisfied the same inclusion criteria of the data used to train the FL model. 1A 61 62 63 Ukucinga data I-20 iindawo ze-client ziye zibonise i-16148 iimeko (zobugcisa kunye ne-negative) ngenxa yobugcisa, ukuvalwa kunye nokuhlolwa kwimodeli (i-Fig. Iinkcukacha zonyango zithunyelwe malunga neempawu ezihambelana nezifundo zophando. Iinkcukacha ze-client zithembisa zonke iimeko ze-COVID-positive ukususela ekuqaleni kwe-pandemic ngoDisemba ka-2019 kwaye ngexesha lokugqibela ukuqeqeshwa kwizifundo ze-EXAM. Zonke iinkcukacha ze-local zilungiselela ngexesha le-30 Septemba 2020. Iinkcukacha zithunyelwa namanye iimveliso ze-RT-PCR ezininzi ngexesha elifanelekileyo. Ngenxa yokuba iinkcukacha ze-SARS-COV-2 ezininzi kunezingxaki ze-positive kunezingxaki ze-SARS-COV-2 ezininzi, sincoma inani leempawu ze-negative 1b A ‘case’ included a CXR and the requisite data inputs taken from the patient’s medical record. A breakdown of the cohort size of the dataset for each client site is shown in Fig. I-Distribution kunye ne-patterns ye-intensity ye-image ye-CXR (i-pixel values) zihlanganisa kakhulu kwiindawo ezahlukeneyo ngenxa yeengxaki ezininzi ze-patient kunye ne-site-specific, njengeempawu ezahlukeneyo ze-device kunye ne-imaging protocols, njenge-Fig. . Patient age and EMR feature distribution varied greatly among sites, as expected owing to the differing demographics between globally distributed hospitals (Extended Data Fig. inguqulelo 1b 1c,d 6 Iingcebiso ze-Inclusion yePasient Iingcebiso zokusetyenziswa kwama-Patient ziye: (1) i-Patient ebonakalayo kwi-ED okanye i-equivalent ye-hospital; (2) i-Patient ebonakalayo i-RT-PCR ebonakalayo ngexesha elinye phakathi kwe-ED kunye nokukhutshwa kwe-hospital; (3) i-Patient ebonakalayo i-CXR kwi-ED; kunye (4) i-Patient's Record ebonakalayo amayunithi e-EMR ebonakalayo kwi-Table. , 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 Ukusetyenziswa kweempawu ze-EMR ezisetyenziswa kwimodeli. I-etiquette ye-resultat (i.e., i-ground truth) iye yahlelwe ngokutsho imfuneko yeempawu kwama-patient emva kwexesha le-24 kunye ne-72 ngehora ukusetyenziswa kwe-ED yokuqala. Inqaku oluthe ngexabiso yeempawu ze-EMR kunye ne-results ezinikezele kwi-Table . 1 Ukudibanisa ukwelashwa kwe-oxygen usebenzisa izixhobo ezahlukeneyo kwiindawo ezahlukeneyo ze-client ibonisa kwi-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 Inani lwezimo ze-COVID-19 ezibonakalayo, efunyenwe yi-RT-PCR esisodwa esithathwe ngexesha elinye phakathi kokufumana kwi-ED kunye nokuphumelela kwi-hospital, ifumaneka kwi-Table Supplementary . Yonke indawo ye-client ibizwa ukuba yahlule i-dataset yayo ngexesha ezintathu: i-70% ye-training, i-10% ye-validation kunye ne-20% ye-testing. Kwiimodeli ze-24 kunye ne-72h ye-output prediction, i-split ye-random yeenkcukacha ze-local kunye ne-FL ze-training kunye ne-evaluation ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye zibe. 1 Ukuqeqeshwa Model Development Kukho ingxaki ezininzi kwinkqubo yesikliniki yabasetyhini ezinxulumene neempawu ze-COVID-19, kunye nabanye abasetyhini ukuphazamiseka ngokukhawuleza kwimfuneko ye-respiratory kufuneka iingxaki ezahlukeneyo yokuvimbela okanye ukunciphisa i-hypoxemia. , Umxholo owenziwe ngexabiso lwezigulane ngexabiso lwezigulane ngexesha lokuqala yokuxhumana, okanye kwi-ED, ukuba ungenza iimpendulo okanye iingxabiso ezininzi ezidlulileyo okanye eziluncedo (njenge-MV okanye i-anticorps monoclonal), kwaye ngoko kufuneka ufumane i-therapy ezincinane kodwa efanelekileyo, i-therapy kunye ne-risk-benefit ratio ebandayo ngenxa yeengxabiso, okanye i-level ephezulu yeengxabiso, njenge-admission kwi-intensive care unit Ngokuqhelekileyo, umdla owenziwe ngexabiso eliphantsi yokufuneka i-oxygen therapy ingasetyenziswa kwicwangciso ye-intensive, njenge-chamber, okanye ngexabiso esuka kwi-ED ukuze isebenze i-self-monitoring ekhaya. . I-EXAM yenzelwe ukunceda ukwahlukanisa izigulane ezininzi. 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. EXAM score I-EXAM yenzelwe ukusetyenziswa kwe-FL; inikeza i-risk score (i-exam score) efana ne-CORISK (Extended Data Fig. ) and can be used in the same way to triage patients. It corresponds to a patient’s oxygen support requirements within two windows—24 and 72 h—after initial presentation to the ED. Extended Data Fig. illustrates how CORISK and the EXAM score can be used for patient triage. 27 9a 9b Iimifanekiso ze-X-ray ze-chest ziye ziye zithunyelwe ngokukhawuleza iimifanekiso ye-position ye-front kunye nokunciphisa iimifanekiso ze-side view, kwaye ziye zithunyelwe kwi-resolution ye-224 × 224. Njengoko ibonisa kwi-Extended Data Fig. , the model fuses information from both EMR and CXR features (based on a modified ResNet34 with spatial attention Pre-trained kwi-CheXpert dataset) and the Deep & Cross network . To converge these different data types, a 512-dimensional feature vector was extracted from each CXR image using a pretrained ResNet34, with spatial attention, then concatenated with the EMR features as the input for the Deep & Cross network. The final output was a continuous value in the range 0–1 for both 24- and 72-h predictions, corresponding to the labels described above, as shown in Extended Data Fig. . We used cross-entropy as the loss function and ‘Adam’ as the optimizer. The model was implemented in Tensorflow using the 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. inguqulelo 9a 66 67 68 9b 69 70 27 Uhlobo imputation kunye normalization I-Algorithm ye-MissForest Ukusetyenziswa ukucacisa iimpawu ze-EMR, ngokusekelwe kwedatha ze-training yendawo. Ukuba iimpawu ze-EMR ayikho ngokupheleleyo kwedatha ze-site ye-client, i-medium value ye-function, eyakhulwe kuphela kwiidatha ze-MGB client sites, isetyenziswa. Emva koko, iimpawu ze-EMR ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye 71 Imibuzo ye-EMR-CXR data fusion usebenzisa inethiwekhi ye-Deep & Cross Ukusetyenziswa kweengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki zeengxaki. . Binary and categorical features for the EMR inputs, as well as 512-dimensional image features in the CXR, were transformed into fused dense vectors of real values by embedding and stacking layers. The transformed dense vectors served as input to the fusion framework, which specifically employed a crossing network to enforce fusion among input from different sources. The crossing network performed explicit feature crossing within its layers, by conducting inner products between the original input feature and output from the previous layer, thus increasing the degree of interaction across features. At the same time, two individual classic deep neural networks with several stacked, fully connected feed-forward layers were trained. The final output of our framework was then derived from the concatenation of both classic and crossing networks. 68 Iinkcukacha FL Ngokuqhelekileyo i-FL ye-form eyenziwe kakhulu yi-implementation ye-algorithm ye-average ye-federated efunyenwe nguMcMahan et al. , okanye iintlobo zayo. Le algorithm ingasetyenziswa usebenzisa i-customer-server setup apho zonke iindawo eziqhelekileyo zenza njenge-customer. Umntu angathanda i-FL njenge-methode elidlulileyo yokunciphisa i-global loss function ngokunciphisa i-set ye-local loss functions, eyenza kwi-site eyodwa. Ngokunciphisa i-local loss ye-customer site nangokuthi ukutshintshela i-customer site we-weights ezaziwayo kwi-centralized aggregation server, umntu angakwazi ukunciphisa i-global loss ngaphandle kokufuneka ukufikelela kwi-dataset epheleleyo kwi-centralized location. Yonke i-customer site ufundisa kwi-local, inguqulelo 72 9c A pseudoalgorithm of FL is shown in Supplementary Note . In our experiments, we set the number of federated rounds at = 200, with one local training epoch per round ngalinye kubathengi. Inani lwabathengi, , i-up to 20 ngokuxhomekeke ne-network connectivity ye-client okanye iinkcukacha ezisetyenziselwa ixesha elifanelekileyo le-result (24 okanye 72 iiyure). Inani le-iterations ye-training ye-local, , depends on the dataset size at each client kwaye isetyenziswa ukuchitha iingcebiso ka-client ngamnye xa ibandakanya iingcebiso ze-model kwi-average ye-federated. Kwi-FL training task, zonke iisayithi ze-client zibonise i-model yayo engcono kwi-model yi-validation set yayo ye-local. Kwakhona, i-server ibekwe i-model engcono kwi-global ngokusekelwe kwi-validation average scores ezithunyelwe ukusuka kwi-client site kwi-server emva kwe-FL round. Emva kokuphumelela kwi-FL training, iimodeli engcono ze-local kunye ne-model engcono ze-global zithunyelwe ngokuzenzakalelayo kumasayithi ze-client kwaye zithunyelwa kwi-data yayo ye-test. 1 T t K nk k When training on local data only (the baseline), we set the epoch number to 200. The Adam optimizer was used for both local training and FL with an initial learning rate of 5 × 10–5 and a stepwise learning rate decay with a factor 0.5 after every 40 epochs, which is important for the convergence of federated averaging Ukuguqulwa kwe-affine ye-random, kuquka i-rotating, i-translations, i-scaling, i-scaling kunye ne-random-intensity noise kunye ne-shifts, ziye zithunyelwe kwiimifanekiso yokwandisa idatha ngexesha lokufunda. 73 Ngenxa ye-sensitivity ye-BN layers Xa usebenzise abathengi ezahlukeneyo kwi-non-independent kunye ne-identically-distributed setting, sinokufuneka ukuba ukusebenza kwimodeli engcono kunokufumaneka xa ukugcina i-ResNet34 ephakamileyo kunye ne-spatial attention. iiparamitha ezihambelana ngexesha lokufunda kwe-FL (yintoni, ukusetyenziswa kwinqanaba le-learning ye-zero kwiingqungquthela zayo). I-Deep & Cross network ebandakanya iimpawu ze-image kunye neengquthela ze-EMR ayikho iingquthela ze-BN kwaye ngoko ke ayikho kwimeko ze-BN ye-instability. 58 47 Kule nophando, sinqathwe inkqubo yokukhusela ukhuseleko le-privacy eyenza iimodeli ezininzi phakathi kwisayithi ze-server kunye ne-client. Iimodeli ze-weight zithunyelwe ngexesha lokugqibela ngalinye ngama-contribution, kwaye kuphela i-percent of the largest weight updates zithunyelwa kwi-server. Ukunyaniseka, iimodeli ze-weight (eyaziwa ngokuba yi-gradients) zithunyelwa kuphela ukuba i-value ye-absolute yaye ngaphezu kwinqanaba le-percentile, (i) (i-Extended Data Fig. ), leyo ifumaneka ukusuka zonke iingxaki engaphakathi, Δ , kwaye kungenziwa ezahlukeneyo kumakhasimende ngamnye ngalinye kwi-FL round I-Variations ye-scheme yaye ingakumbi kubandakanya i-cliping ye-gradients ezinkulu okanye i-differential privacy schemes Oku kwandisa i-noise kwi-gradients, okanye kwakhona kwi-data ebomvu, ngaphambi kokufumana kwi-network. . k 5 Ukugqithisa k t 49 51 Statistical analysis Thumela i-Wilcoxon signed-ranking test yokubonisa i-significance ye-improvement epheleleyo kwimveliso phakathi kwimodeli ebandayo kunye ne-FL kwimodeli ye-24 kunye ne-72 iiyure (i-Fig. Izixhobo ze-Data Fig. ). The null hypothesis was rejected with one-sided « 1 × 10–3 kwiimeko ezimbini. 2 1 P I-Pearson correlation yasetyenziswa ukucacisa i-generalizability (i-robustness ye-average AUC value to other client sites' test data) yama-models ezidlulileyo kwi-locally trained in relation to the respective local dataset size. I-moderate correlation kuphela ibonakala ( = i-0.43, = 0.035, degrees of freedom (df) = 17 for the 24-h model and = 0.62, = 0.003, df = 16 kwimodeli ye-72h). Oku kubonisa ukuba ubungakanani we-dataset kuphela akuyona umzekelo wokufuneka ubunzima we-model ukuya kwi-data ebonakalayo. r P r P Ukubala izibani ze-ROC ezivela kwi-global FL model kunye ne-models ezivela kwiindawo ezahlukeneyo (I-Extended Data Fig. ), sisetyenziswa iiyunithi ezili-1000 ukusuka kwedatha kunye nokucacisa i-AUC efanelekileyo. Emva koko siqhathanisa i-difference phakathi kwama-series ezimbini kunye ne-standardization usebenzisa i-formula = (AUC1 – AUC2) / Yintoni Yintoni i-standard differential, I-AUC1 kunye ne-AUC2 ziquka i-bootstrapped AUC series. Ukubala Ngokusetyenziswa okuzenzakalelayo, sisebenzisa i- Iimpawu ezibonakalayo kwi-Table Supplementary Iimiphumo zibonakalisa ukuba i-hypothesis ye-null ilawulwa ngexabiso eliphantsi kakhulu iindidi, ebonisa ubunzima we-statistical significance ye-superiority ye-FL imiphumo. Ukubala iindidi ziye zithunyelwe kwi-R kunye ne-pROC library . 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. I-ANOVA (I-Variation Analysis) ye-analysis ye-unidirectional (i-ANOVA) ye-analysis ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics ye-analytics). -statistic, calculated as the variation between the sample means divided by variation within the samples and representing the degree of dispersion among different groups, was used to quantify the models. Our results show that the -izinga ze-5 iindawo ezahlukeneyo ze-local ziquka 245.7, 253.4, 342.3, 389.8 kunye ne-634.8, nangona le model ye-FL yi-843.5. -izinga zibonisa ukuba amaqela ziquka kakhulu, iingcaciso evela kwimodeli yethu ye-FL zibonisa ngokucacileyo ukuxhaswa kwimigodi ezine zeengcaciso zeengcaciso. Ngaphezu kwalokho, i ixabiso lwe-ANOVA kwimodeli ye-FL i-<2 × 10-16, ebonisa ukuba iziphumo ze-FL ze-prediction zihlukile ngokucacileyo phakathi kwezilwanyana ze-prediction. 10 F F F P Reporting Summary Ulwazi oluthe ngexabiso malunga ne-research design ifumaneka kwi linked to this article. Izifundo zeNature Reporting Data Ukuphepha I-dataset ye-20 iiyunithi ezidlulileyo kwi-studi ye-CAMCA ibekwe phantsi komgangatho yayo. Lezi zibonelelo ziye zisetyenziselwa ukuqeqeshwa kwizindawo ezininzi ze-local kwaye ayathunyelwa kwamanye izakhiwo ezidlulileyo okanye kwi-server ye-federated, kwaye ayikho kwi-publishment. Iinkcukacha ezivela kwiindawo ze-validation ezivela kwi-CAMCA, kwaye ukufumana ukufikelela kunokufumaneka kwi-Q.L. Ngokusekelwe ku-CAMCA, ukuxhaswa kwe-data-sharing kunye ne-amendment ye-IRB kwiimeko zophando kunokelela yi-administration ye-MGB kunye ne-IRB kunye ne-policy ye-MGB. Code availability Zonke i-code kunye ne-software ezisetyenziswa kwisifundo le-NGC ziyafumaneka ngokubanzi kwi-NGC. Ukufumana ukufikelela, ukufikelela njenge-guest okanye ukuvelisa i-profile, nqakraza enye kwi-URL ezantsi. Iimodeli eziqeqeshiweyo, izixhobo zokulungisa idatha, ikhowudi yokulungisa, ukuvalwa kwimodeli, ifayile ye-readme, isixhobo se-installation kunye neefayile ze-license ziyafumaneka ngokubanzi kwi-NVIDIA NGC iimveliso The federated learning software is available as part of the Clara Train SDK: . Alternatively, use this command to download the model “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 References 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 Ukubuyiselwa Izixhobo ze-Clinic ezidlulileyo kwi-Data ye-EMGB zibonisa iingcebiso zeengcali, kwaye ezininzi zeengcebiso ze-NHS, i-NIHR, i-Department of Health and Social Care okanye iinkonzo ezininzi ezihambelana neengcebiso. I-MGB ibandakanya iindidi ezilandelayo ngenxa yayo: J. Brink, i-Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, MA; M. Kalra, i-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, i-Department of Radiology, Massachusetts General Hospital, Boston, MA I-Faculty of Medicine, iYunivesithi yaseChulalongkorn sincoma i-Ratchadapisek Sompoch Endowment Fund RA (PO) (nr. 001/63) yokufaka kunye nokulawula iinkcukacha zonyango kunye neengxaki zonyango ezinxulumene ne-COVID-19 kwi-Research Task Force, i-Faculty of Medicine, i-Chulalongkorn University. I-NIHR Cambridge Biomedical Research Center sincoma u-A. Priest, ebizwa yi-NIHR (i-Cambridge Biomedical Research Centre kwi-Cambridge University Hospitals NHS Foundation Trust). I-National Taiwan University MeDA Lab kunye ne-MAHC kunye ne-Taiwan National Health Insurance Administration sincoma i-MOST https://data.ucsf.edu/covid19 Le nqaku lula phantsi kwe-CC by 4.0 Deed (i-Attribution 4.0 International). Le nqaku lula phantsi kwe-CC by 4.0 Deed (i-Attribution 4.0 International).
