The Covid pandemic highlighted the urgent need for collaborations in the healthcare sector to empower clinical and scientific communities in responding to global challenges. In this context, the ICU4Covid project joined research institutions, medical centers, and hospitals all around Europe in a European Telemedicine Network, allowing for sharing of capabilities, knowledge, and expertise distributed in such a network. Nevertheless, healthcare data sharing has ethical, regulatory, and legal complexities imposing restrictions on access and use. In addition, data and knowledge are very often unevenly distributed at the diferent nodes of the network depending on their geographical location and dimension. To address these issues, a federated learning architecture is proposed to allow for distributed machine learning within the cross-institutional healthcare system without moving data outside its original location. The approach has been applied for the early prediction of high-risk hypertension patients. The experimentation carried out shows that the knowledge of single nodes is spread within the federation, improving the ability of each of them to perform predictions also on not previously treated cases. The performance evaluation of the computed predictions in terms of accuracy and precision is over 0.91 confirming the encouraging results of the proposed FL approach.
this context, the ICU4Covid project provides valuable
ground for learning from real-world health data that has
SARS-CoV-2 pandemic highlights the need for improving proven to be efective in multiple healthcare applications,
cooperation and knowledge sharing to prevent disease resulting in improved quality of care [
The ICU4Covid project [
D2.3: TECHNICAL HANDBOOK FOR THE CPS4TIC
The two actors involved in the telemedicine consultation are the telemedicine recipient requiring large data s(uestusa.lWlyferosmhoinwsidhe othwe htohsepiktanl)oawndlethdegteelemicedailcicnoecpkropviitd,era(utseulealmly efrdoimcionuetsicdoenthseole installed in every owned by a single orhgoasnpiitzaal)t,iwonhoiscasnparelsaodbea mreofenrrgedalalsthineternaplearnipdheexrtearlnhalopsrpaicttaitli,oanecros,nrnesepcetcotrivpellya.tform, and smart bedmembers of the federHaotwioenverb,ythimeypruosuvailnlyg btehleonrgeltioabtihlietysamseidIeCUhuHbus.bThhoespICitaUl.4TChoevitdelepmroejdeiccitneis aimed to deploy the of the local model ascoanspurlteadtiiocntiwoonrktfleoswt.is Aas qfoulloawnst:itative CPS4TIC in many hospitals across Europe as a global netand qualitative estimati1o.nTohfe itemlepmreodviceimneernectipbiernoturgeghisttebrys in twheotrekle.mICedUic4inCeocvoindsopler o(Mjeocntaa)adnvdaonpceenssthe CPS4TIC to largethe federated process overthtehveirltouaclatlelnemodedeiscinise rporoomv.ided, scale experimentation and deployment with full-scale reporting an enhanceme2n.t oThfethteelepmeerdfiocirnme arenccipeieunptctoont3a8c%ts.theptaelretmiceidpiacitnioenproofvihdeera.lthcare staf, hospitals, and end-users. Finally, a comparison betwTheeetnelemedicine provider logs aonndto the web interface ofatfhoeretemleemnetdiiocinneedcoicnknpiot.vative technology en3. federated learning Adapting the the centralized approac4h. sTwhhaoeitwitneglseimnthethadeticritnoheoemprfo(ewvdiidteherrMastoeenledac)t,asaptnh-decrheaocbispleeisesnthct’esonrheotsesppmeitcpatilov,esreacerasylla.IvnatileanblseivpehyCsicairaen Units transformation proach prevents data p5r.ivacy and security issues into a structure that operates as one ICU Hub consisting
Optional: the telemedicine recipaietnat shares the screen. slight performance loss.6. Optional: the telemedicine recipient haongfsounpetoceconnttrianuliezethde hseossspiointainl caonontnheercted to its peripheral
The rest of the paper is ororogman.ized as follows. Section hospitals in a geographical area. Each ICU Hub is com2 introduces the ICU4CO7.VTIDhe EteulermopedeiacninePrreocjipeicetn.tSaencdtpioronvideproensdedthoefcaallc.entral ICU and interconnected peripheral hos3 shows an application s8c.enHaerailothopfrofefedsesrioantaelds (lteealermneindigcinfoerrecippiietnatlasnedmprpolvoidyeirn) glogteolfefm.edicine and telemonitoring techenhancing healthcareTkhenovwidleeodgceo.nSsuelctatitoionn 4tapkreessepnlatscethPeeer-ntoi-qPueeers t(Ph2aPt) avsisaistWhebeRaTlCth,cwahreichstaafrien patient screening, performance evaluati oinnte.rFcionnanlelcyt,eidn. TSheectteiolenm5e,dcicoinnecmluosdioulnesis indteinadgendotosiennga,balencdomtrmeuantimcaetinont.bEetawcehenICU Hub is equipped and future works aresptarfef soennstiteeda.t the hospital and external physiwciainthslosctaatteed-oouf-ttshidee-tahrethtoescphitnaloi nloagctyiv,esuch as a 5G module, rooms (video call sessions). The active roomsracdanaronsleynbseorssta,rtaenddbyAaI pchhyispicsia,naantda controlled by a contelemedicine console (Mona) terminal.
trol station called Integration Center. The ICU4Covid architecture allows for the coexistence of both already Page 10 eofst2a5blished and new ICUs as one ICU node. In fact, the system is independent of the hospital’s infrastructure enThe pandemic caused by the SARS-CoV-2 and subse- abling highly encrypted telemedicine and digitalization quently by the COVID-19 virus showed that the uneven of ICUs with collective technological eforts. distribution of capacities and resources of Intensive Care Units (ICUs) located in rural and urban areas remains a big challenge. Hence, real-time information sharing 3. Federated Learning for and cooperation between hospitals, healthcare work- enhancing healthcare eCrosv,iadn-1d9tshperepaudbilnigc aanred henigshulryinsgighnigificha-nqtuatolitcyohnetaalitnhicnagre knowledge services. The Cyber-Physical System for Telemedicine To allow for sharing knowledge between each ICU Hub and Intensive Care (CPS4TIC) aims at expanding Infor- without incurring data privacy and security risks, we mation Technology-based operations and information propose to equip each node of the CPS4TIC system with sharing from the central ICU Hubs to peripheral or ru- a federated learning (FL) architecture. The working sceral hospitals while substantially minimizing the infec- nario consists of three participants respectively named tion risk for healthcare staf (see a conceptual view in Hospital 1, Hospital 2, and Hospital 3 with the same data Fig.1). The CPS4TIC framework comprises a telemedstructure of the CPS4TIC hosted in each hospital, as illustrated in Fig. 2. The data of a participant is private and owned by the hospital, and it is used for local training to learn a local model. Each participant uses a local hybrid network, where hybrid means that the network is composed of diferent deep learning networks. Each local model updates from each participant are sent to an aggregator server that combines them into a global consensus model. This global model is then returned to each participant for further local training. The participants connect to the aggregator server through remote procedure calls via a transport layer security network connection. Sensitive information such as model, optimizer weights and aggregated metrics move between the participant and the aggregator server over this encrypted hitecturechVannael.liFdor tahetsaikoeonfexperimentation, these three participants are emulated on local clusters. cause the knowledge of the CPS4TIC-1 model has been included in the aggregated model according to the federated learning process.
At the end of the federated learning process, all models can correctly classify the validation sample since the aggregation server shares the updated parameters round
TCP/IP by-round with the clients involved in training, so collectAggregator Server ing the knowledge of all nodes.
IP:192.168.4.11:8080 Table 1 reports the qualitative accuracy trend during the training process showing how the accuracy changes in the diferent stages of the training.
Locally at CPS4TIC 1, the accuracy consistently achieves good values from the beginning of the training, while the other nodes and the aggregated model expose low Figure 2: Federated CPS4TIC nodes accuracy values ranging from 52% to 61%. As the rounds 5 progress, CPS4TIC 1 and the aggregator models improve
The proposed approach supports the decision-making their performance to 89% and 75%, respectively. Followprocess of a network of federated hospitals, as those pro- ing the FL algorithm, the aggregator model’s knowledge vided by the ICU4Covid project. In order to validate the is spread to CPS4TIC 2 and CPS4TIC 3, and consequently, approach, it is applied to a use-case scenario for predict- at the end of the training process, their performance iming high-risk hypertensive patients. proved, achieving 84% and 81%, respectively.
An ECG sample of a patient is taken from the dataset Without the federated learning process, Hospital 2 or belonging to CPS4TIC of client 1 (CPS4TIC-1) stored Hospital 3 would have classified as a low-risk only on CPS4TIC-1, named . Such a sample patient and consequently, no healthcare protocol would is labelled as high-risk class. Therefore, CPS4TIC-2 and have been adopted for that patient. Conversely, Hospital CPS4TIC-3 do not learn anything specifically from that 2 and Hospital 3 can correctly classify allowing patient. Fig. 3 shows the progression of the federated them to adopt appropriate healthcare since their local training process. In the beginning, only CPS4TIC-1 is models are updated with the knowledge of the aggreable to correctly classify the validation sample, while gated model.
CPS4TIC-2 and CPS4TIC-3 can not. At this stage, also The improvement of the local models due to the federated the aggregated model fails to correctly classify the sam- process is shown in Table ?? reporting an increase from ple since experimental results show that more training the beginning to the end of the training equals 38% and rounds are necessary to successfully merge the knowl- 35%, respectively. edge from the local models. In the middle of the federated learning process, CPS4TIC-1 still correctly classifies the validation sample, as well as the aggregated model be
CPS4TIC 1 0.85 0.89 0.92
CPS4TIC 2 0.52 0.62 0.84
CPS4TIC 3 0.52 0.61 0.81
This section provides a performance evaluation of the federated learning approach compared with the classical centralized one. The evaluation was conducted using the SHAREE [27] database and considering three local nodes for the federation. The SHAREE database contains 169 electrocardiographic (ECG) records of hypertensive patients monitored with an epicardial holter for 24 hours with an attempt to record major cardiovascular events. Patients who experienced dangerous events were marked as high-risk, while the rest were marked as low-risk. In such an experimental evaluation, the dataset is used to train machine learning models to identify subjects at a higher risk of developing fatal cardiovascular events. The learning process uses multivariate time series (MTS) data, whose raw signal comes from three electrodes placed on the subject’s chest during monitoring.
Five-minute segments of input data are randomly selected as samples. Hence, the training set contains more than 14,000 samples evenly distributed between the two classes, high-risk and low-risk. Training set samples were equally distributed among the three local nodes of the federation. The test set, instead, is defined using the holdout approach. Thus, 700 samples are used on each client node, with two-thirds marked as low-risk and one-third as high-risk.
Table 2 shows the performance achieved in terms of accuracy by each local model on the test set at the end of the learning process. As it is possible to note, the aggregated model achieves slightly higher performance than the local models. Indeed, the accuracy of the aggregated model is equal to 90% higher than each local model, 87%, 88%, and 88%, respectively.
Acc
0.87
0.88
0.88
0.90
Federated Scenario Centralised Scenario
0.90+-0.0019 0.98+-0.005
0.91+-0.0059 0.98+-0.002
performance, with Accuracy and Precision values of 98%.
Despite the better performance of the centralized approach, the performance results achieved by the federated model can be considered satisfactory in terms of Accuracy and Precision, considering the advantages coming from the adoption of a federated approach in the healthcare domain. The centralized mode will require moving all data from its stored nodes to the node performing the learning process. Thus, data security and privacy are compromised by this action. Fig. 4 clearly shows that in the centralized approach, 100% of data are moved across the nodes to be collected in a unique node. On the contrary, the federated approach prevents privacy risks since no data are moved; only the federated model parameters are transferred.
Table 2 ated approach can introduce communication costs issues Comparison between local models and the aggregated model while sharing the neural network parameters, issues evaluated and addressed in [25, 26].
Table 3 compares the federated and classical central- Our results highlight a trade-of between performance ized approaches in terms of accuracy and precision. As and security, reported in Fig. 5, which visually describes it is possible to note, the Federated approach achieves the relationship between these two aspects for a generic an accuracy of 90% and a Precision of 91%, respectively. predictive model. The figure defines a qualitative visual