DCC - UFMG
+55 31 3409 5860
ciia-saude@dcc.ufmg.br

Examples of Research Lines

The examples described in this section describe some research lines to be developed within the Center in the first 5 years of the project. These are lines that are already being developed by the CIIA-Health team, and will be extended throughout the project, or new lines that we intend to develop. These examples are part of the health axes and research lines in AI, described above. The objective is to illustrate the enormous possibilities and potential that a synergistic development of research in health and AI may bring. It is worth mentioning that the examples are not exhaustive, not limiting the possibilities of what we can achieve in the mid and long term in this project, given the ambitious objectives and the quality of the CIIA-Health research team.


Prevention of chronic diseases

Context: In the context of prevention and quality of life, the importance of Chronic Non-Communicable Diseases (NCDs) stands out, accounting for 75% of annual deaths, in addition to the high prevalence of risk factors in the Brazilian adult population: 55% of overweight, 20% of obesity, 10% of smoking, which may result in 38.8% of the loss of DALYs (years of life lost due to premature death and disability).133 It is important to identify individuals at risk of illness and areas of greatest vulnerability with a view to preventive measures.

Questions: Which individuals or population groups will evolve with chronic NCDs in the course of life? How to estimate health indicators and risk factors for small geographical areas?

Bases / methods: Databases and national surveys, wearable devices and data from virtual sensors, cohorts, genetic databases; integrate pre-existing databases (e.g., information collected from social networks, interaction with clinical centers, mobility data, census data) to complement SUS data. Develop embedded wearable devices monitoring individual indicators; Implement physical and virtual sensors to collect environmental data from small geographic areas; Propose heterogeneous data fusion techniques; Develop AI models to predict the risk of illness and death from NCDs for individuals and populations. In order to generate estimates for small areas, we will use a set of AI modeling strategies, such as “generalized stacking” to generate predictions of the variable of interest across the geographic area. When building these predictions, we will use Bayesian and covariate geostatistical models by census sector, in order to estimate indicators for small areas.

Results: Identify populations and areas at risk, aiming at the implementation of public health promotion policies; Provide managers with the best evidence on the outcomes of adopting more effective public health promotion policies, improvements in health services and implemented care. At the individual level, predict risks and provide personalized recommendations that facilitate healthy choices and better possibilities of care and self-care.

back


Telemonitoring and mobile technologies

Context: Web-based and / or mobile / wearable technologies have been increasingly used in mental health and shown to be moderately effective, particularly when guided by a human coaching, which depends on minimally trained professionals, often unavailable or insufficient. The lack of customization and the inability to adapt to the user’s needs or preferences results in worse adherence and results. The storage of mental health data is often performed in a single centralized point, presenting vulnerabilities of robustness and low fault tolerance.

Questions: How to identify risks and symptoms of mental disorders through the use of mobile technologies, telemonitoring / telemetry / wearables? How to use individual data and collective bases to estimate risks and support early diagnosis and interventions in mental disorders?

Bases / methods: The approach will involve the selection of a set of diagnostic instruments for the most prevalent mental disorders, compatible with digital platforms and the development of new sensors; evaluation of new storage architectures, which, at the same time, maintain privacy and security; development of a set of interactivity tools and AI models with human on the loop, and optimization through usability tests. It will include a double-blind, randomized controlled clinical study, comparing it to traditional diagnostic tools for mental disorders. The application of adaptive machine learning analysis to mobile applications has the potential to create a new generation of tools, which may revolutionize the way these processes are conceptualized, designed and implemented. At the same time, the application will be made compatible with the standards established by GDPR and LGPD.

Results / products: (1) Individualized recommendations based on individual habits – AI algorithms for adapting the message to the media (bots, SMS, e-mail, cell phone pop-ups), including cultural and regional aspects. (2) Technology for secure data storage – Storage will be encrypted, supporting authentication and access control, carried by individuals or guardians. In this way, it will be possible to feed the AI ​​with anonymized clinical data and give patients or guardians access to the data in a safe manner. (3) Integrated digital platform to aid in the diagnosis and management of patients– The identification and validation of clinical and physiological parameters and indicators, in a digital, safe and simple to use platform, has the enormous potential to facilitate and optimize diagnosis and care of patients. (4) The investment in efficiency that technology may provide should expand the reach and results of mental health care, offering care that is absent or of low quality in SUS, increasing the chances of reliable diagnoses, regardless of the individual’s location; New communication standards and protocols; new sensors; new AI algorithms for health. These innovations could be extended to a wider range of tools, based on the web and directed to other mental health problems.

back


Diagnosis guided by AI

Context: Improving access to the health system involves automation and improving the accuracy and reproducibility of routine diagnostic methods, although the challenges are different for different exam modalities. The electrocardiogram is a routine method used for the diagnosis of cardiovascular diseases, whose access has increased in Brazil, due to the miniaturization of equipment, expansion of primary health care and the Minas Gerais Telehealth Network, which is currently present in 2500 points of care. attention throughout Brazil.134 Conventional automatic reading systems have limited acuity and the initial experience of our group points to a better accuracy of the diagnosis with the use of AI.135 Another scenario where AI may bring health benefits is in the automation of the screening of rheumatic heart disease by tele-echocardiogram.136 Rheumatic heart disease is a major heart disease in the Brazilian population, impacting quality of life and survival: we found subclinical RHD in 0.5% of young people in Brazil, but the screening performed by doctors is impossible to be extended to the entire population.137 Recognition of typical findings in cardiac ultrasound videos obtained by non-physicians using telemedicine would allow the screening process to be used routinely. The same applies to the recognition and quantification, in slides of histopathology, of hepatic steatosis, which can be used for risk stratification and management of patients with fatty liver diseases or for the quantification of tumor infiltrating lymphocytes, which are associated with the immune response of the tumor. Thus, the AI-guided diagnosis may facilitate this task and make health resources (human and physical) available for therapeutic and management actions. On the other hand, automation should support privacy, security and interpretability, in order to lead to interaction with the specialist and avoid systematic errors in the diagnosis.

Question: Is it possible to use AI techniques to assist the pattern recognition in different complementary health exams, such as ECG, echocardiogram and histopathology slides, in order to automate the diagnosis, define prognosis and guide therapy, with accuracy, safety and transparency ? 

Material and methods

Bases: (1) ECG: CODE base, with 2.5 million tests, including the ECG signal, clinical data, automatic ECG measurements and 78 diagnostic classes. This base is paired with SUS mortality data, with a follow-up of up to 7 years (2010-17), and it is expected to be paired with the SUS hospitalization base. (2) Echocardiogram: 18 thousand screening echocardiograms for rheumatic heart disease, obtained by portable devices by non-doctors, in young people aged 7 to 18. Each exam includes short videos, which contain problems with the quality of the records and insufficient standardization. (3) Histopathology: The Cancer Genome Atlas (TCGA http://cancergenome.nih.gov/), with more than 30 thousand high-resolution images of tissues, 30 types of cancer.

Methods: Integration of ECG database, hospitalization, and mortality by probabilistic record linkage. Definition of risk scores (ECG). Standardization and segmentation of images (echo). Use of predictive models based on DNN / CNNs and clustering algorithms. Obtaining (1) explainable models, but with a high level of learning performance (prediction accuracy) and (2) human users  understanding and effective cooperation with artificially intelligent agents (predictive models). Develop tools for computer-assisted diagnostics with implementations hosted on cloud computing architectures and / or embedded systems on computers and mobile devices.

Results: ECG / echo: Preliminary results already published;9,114,135 automatic ECG classification system; risk prediction algorithms. Echo: simplified portable ultrasound tracking system for rheumatic heart disease. Recent results show the efficiency of quantifying steatosis droplets in slide images and correlation with clinical data, such as the degree of the disease138. Products: Cloud systems for automatic ECG and echo diagnostics, software to be shipped on ECG and echo devices. Automated system for classification and quantification of tumor infiltrating lymphocytes in digitized tissues, extending and generalizing the previously used techniques138.

back


Personalized medicine in oncology

Context: AI may play an important role in advancing knowledge in personalized medicine applied to oncology, both in sporadic tumors that present themselves as multifactorial diseases (with somatic and environmental genetic determinants) and in hereditary cancer, a set monogenic diseases where a germline genetic variant determines greater predisposition. Challenges for AI in the field of sporadic tumors include the identification of genetic variants related to the process of carcinogenesis (variants driver) and proposing new potential therapeutic targets from these variants and / or complex interactions among different genes involved in each of the different stages of tumorigenesis. An additional challenge in the Brazilian population is the influence of the mixed nature of individuals on the expression of the disease (phenotype modifying factors) and the identification of drivers of mutations. Finally, AI has a promising role in the integrated analysis of omic tumor data in order to identify new biomarkers for early diagnosis and therapeutic targets139,140. In hereditary cancer, challenges for AI include the search for causal pathogenic variants in families with a clearly hereditary phenotype, but without identifiable germline changes in the analysis of canonical genes (missing heredity), and in the determination of pathogenicity or causality of Variants considered of Uncertain Meaning (VUS).

Questions: How can AI: (1) identify new therapeutic targets for solid tumors? (2) integrate environmental and trans-omic risk factors to develop risk prediction classifiers in common tumors and to discover new biomarkers for early cancer diagnosis?; (3) determine the pathogenicity of virus in genes predisposed to cancer?

Materials and Methods: We will develop classifiers of genetic variants from functional annotation data, structural protein data, evolutionary conservation data between species, metabolic pathways and pathogenesis profiles. The training of the pathogenicity /inference methodology drivers will be carried out from public databases of exomes of patients with different ancestry, and of population-based and pathogenic mutations. Databases to be considered include Exome Aggregation Consortium. gnomAD, Clinvar, OMIM, and Orphanet. Supervised learning algorithms, such as artificial neural networks, will be trained and optimized for the development of the VUS classifier. We will implement a learning strategy online to make it possible to update the classifiers from new data. The management of databases will be carried out in part using the MASSA platform (Map-Reduce System for SNVs Annotation), which will be improved throughout the project. For the discovery of biomarkers and potential therapeutic targets based on multi-omic data, model combination strategies will be explored, such as learning ensemble and hybrid models, and data fusion methods capable of significantly combining data of different scale and nature. and consistent. We will use data from cBioPortal (including TCGA), COSMIC, and other public databases related to cancer genomics.

Preliminary and expected results: The group has already developed a classifier based on neural networks that predicts the stage of development of Prostate Cancer based on tests provided by SUS and performed an epigenomic analysis on pancreatic cancer that identified differentially methylated signaling pathways that may indicate new antineoplastic therapies141. Expected results include (1) identification of new biomarkers for early diagnosis and therapeutic targets by computational strategies, (2) proposition of strategies for training models from heterogeneous data, based on the case of multi-omic data in cancer, and (3) training of classifiers to identify pathogenic variants and mutations based driver on results of genetic sequencing obtained from public databases.

back


AI-guided surgery

Context: Surgical procedures, especially those of greater complexity, require high expertise. Despite all the planning that is carried out, important decisions have to be made in real time during the execution of the procedures. In addition, the surgeon is often unable to visualize the entire surgical field and has to access regions that are difficult to access and / or ergonomically difficult.

Question: How can the different Artificial Intelligence Techniques142, especially machine learning, computer vision and robotics, be used to assist the surgeon and his team in the various phases of surgical procedures?

Methodology: The three-dimensional reconstruction of organs, from data obtained by different exams and their visualization through virtual reality or 3D printed models, will allow the surgeon to plan the surgical procedure, trying to anticipate the main difficulties. Radiological images, echocardiograms and images in the visible spectrum will be processed using Computer Vision (VC) techniques such as Structure From Motion143 and Optical Flow144 to infer the three-dimensional structures of the regions and / or organs of the body under study. In the intraoperative period, the models generated in the preoperative period can be used in conjunction with augmented reality systems, allowing the visualization of hidden and / or difficult-to-access regions145. An example is the removal of liver tumors: from preoperative imaging exams, with the application of VC algorithms, a 3D model of the liver may be constructed with the location of the tumor, blood vessels, etc., and this information would be available to the surgeon, in real time, through the use of augmented reality glasses such as Hololens. These systems can also be augmented with machine learning algorithms. From multisensory data (images, vital signs, surgeon’s myoelectric signals, motion detection, data from previous surgeries et.c) it is possible to use AI algorithms in the development of systems that alert the surgeon in real time about potentially risky actions during procedure146. Finally, an application of AI that is already a reality in surgical centers is Robotic Surgery, where robotic devices, such as da Vinci, assist the doctor in performing procedures, allowing less invasive surgery and easier recovery, compared to conventional surgeries and bringing more comfort to the doctor. In this context, AI-based control algorithms will also be investigated in order to bring additional benefits to robotic surgery, such as assisting the surgeon’s movements, greater immersion in the environment with haptic feedback, among others147. It is worth noting that these systems can be implemented and tested with the robotic surgery simulator148 to be acquired by the UFMG robotic surgery center through a FINEP (Brazilian agency)  project. Through the use of this simulator, it will be possible to develop, test and analyze the algorithms and their applicability to surgical procedures.

Expected Results: (1) Development of 3D virtual and augmented reality systems that, based on multimodal sensory data, may assist the surgeon both in preoperative planning and during surgery; (2) Application of machine learning algorithms that can model the behavior of the doctor and staff during surgery. The obtained models may be both used in real time, to issue alerts of potentially risky actions and during the training of new surgeons; and (3) Implementation of intelligent control algorithms that may assist the surgeon during robotic surgery procedures.

back


Prevention and mitigation of epidemics

 Context: For several reasons, acute events of great intensity, such as epidemics and disasters, cause enormous stress to people and health systems. There is an immediate need to plan actions to assist victims and mitigate risk. In many instances, these events cannot be predicted. In the case of COVID-19, despite a similar previous event (SARS caused by SARS-CoV-1) and evidence of circulation of new viruses in the region, there was no expectation of a pandemic of the intensity and size with which we are living today. On the other hand, there are epidemics that are repeated frequently, following patterns that are only partly expected and predictable. Dengue, for example, happens only in times of rain and heat, since it requires the growth of the vector mosquito Aedes aegypti, and follows a cycle every 2 or 4 years, depending on the city or country under study. If, on the one hand, predicting new epidemic events and disasters is not simple, AI provides a unique opportunity to use the knowledge gained from previous events to guide health actions to deal with subsequent events. In the context of COVID-19, mathematical models for modeling the spread of the disease or models based on sequencing data103 report on the introduction of the virus in the country, its spread and the impact of non-pharmaceutical intervention measures. In the context of dengue, we have shown that the use of Tweets with other epidemiological data performs well in predicting the entry and spread of dengue in an urban center100

Questions: How to build models, in the context of Covid-19 and other respiratory syndromes, that take into account the great social, demographic and environmental differences, in an extensive and diverse territory like Brazil? Being a tropical continental country and probably with the greatest biodiversity in the world, how to predict the risk of new epidemics by monitoring new genetic sequences, including viruses, in vectors and animals, especially mammals, that surround us? In the context of epidemics with cyclical events only partially predictable, such as Dengue, how to obtain data that allow a better geographical definition of the distribution of the disease in large urban centers (regions or neighborhoods)? How to obtain alarm systems that allow a good separation between dengue and other arboviruses, such as zika and chikungunya, and that possibly foresee the introduction of new viruses? How to train and inform the manager, at the health tip, so that data generated by AI models are understood and used in order to modify health actions? 

Bases / methods: Using available databases that include sequences, formal notifications, epidemiological data, population data, in addition to other data sources such as the web, social networks and from georeferenced devices, the construction of models requires the integration of these data, which they are sparse, skewed, unbalanced and multimodal. The models, built from these heterogeneous data, seek to represent the evolution of epidemics in time and space. Among other strategies, an “ensemble” of models will be used to compose the description, fusing the different types of data. These models can be built from knowledge transfer and learning. Epidemic processes share some “invariants” that, once determined for one process, remain valid for others. Examples of this can be found in the current effort to understand the transmission processes of COVID-19: if we build reasonable models, for example, for the transmission of the disease in specific environments (public transport, schools, restaurants, etc.) we can expect that such models are reused for other diseases that have similar transmission mechanisms in the same environments. 

Expected results: The systems to be built will focus on the production of information capable of supporting the understanding of the phenomena under observation, allowing the formulation and testing of hypotheses and the simulation of scenarios. Through the analysis of the interaction, the system itself must be able to evolve, prioritizing strengthening of the analytical capacity and providing the necessary instruments for decision making by health professionals. The expected results include: follow-up or action protocols in case of catastrophic events that could quickly obtain data to guide actions by health managers and other civil society bodies; early warnings about the emergence of epidemics by both known and undetermined pathogens; geographical delimitation of locations affected by an epidemic outbreak; short-term projections of the evolution of epidemics, as well as the simulation of scenarios for different containment measures, taking into account regional differences, including differences in access to health. In particular, it is expected to promote, with the implementation of the developed tools, the empowerment of health professionals and managers.

back


Integration of SUS Databases and supplementary health

Context: In terms of management, SUS information systems are strategic in the formulation of health policies. Examples include Ambulatory information systems (SIA / SUS); Hospitals (SIH / SUS); National Register of Health Establishments (CNES), National Supplementary Health Agencies (ANS), Mortality Agencies (SIM) and Live Births (Sinasc). Despite all efforts, these systems do not work in an integrated way, nor can they be enriched with data from external systems such as hospital systems, demographic / statistical data, heterogeneous data generated by sensors. Despite the introduction of the Brazilian National Health Card (CNS), individuals still appear multiple times or in a non-integrated way in different systems that cover a specific dimension of health, not allowing them to be easily tracked. In this context, record linkage, described in Section 4.2, is a central problem in this integration. In SUS, this type of problem has been treated through the recognition of similarities in noisy data (missing, incomplete, incorrect) or through the linkage of specific systems, such as the longitudinal analysis of patients in Renal Replacement Therapy149 or affected by cancer150. The absence of linkage may, for example, “inflate” certain statistics, (e.g., multiple repeated occurrences of the same disease or pathology due to the error of not identifying the same person affected by the disease), causing the problem to be over-dimensioned and leading to erroneous decisions by managers. On the other hand, the linkage and identification of patient outcomes have a huge benefit potential, from better management of critical points of care, assessment of provider outcomes by area, cost-effectiveness assessment of treatments, suggestion of treatment, optimal rental equipment in the national territory, with pre-established forms of access for highly complex services.

Questions: How to use AI methods to achieve better health outcomes, better indicators, cost-effective use of drugs, longer survival? How to obtain better results in management, how to increase efficiency, better use and occupation of beds, waiting lines, monitoring the results of other health units? How to use AI to support efficient, effective and scalable linkage of dozens or hundreds of SUS medical bases that are incomplete, multimodal and heterogeneous, and diverse? How to check the quality of the linkage (semi-) automatically?

Materials and Methods: 

Bases: all aforementioned databases, as well as other possibilities that arise throughout the project.

Methods: The new AI-based linkage methods that will be developed in this project will explore (1) active learning119 and co-training120; to create labeled training examples and to evaluate the matches found, with retroactive feedback to the system (human-in-the-loop), requiring a minimum effort from the specialist and thus ensuring scalability to the techniques; (2) meta-attributes118 together with deep neural networks based on transformers to improve similarity functions, thus ensuring effectiveness; (3) and meta-blocking93 to ensure efficiency. Some of the techniques mentioned were developed for other related AI problems, such as Natural Language Processing and Textual Classification. The adaptation and integration of these techniques in a unified framework for the linkage problem guarantees an aspect of novelty and originality to the proposal of this research line.

Expected results: Data Lake of integrated and “clean” medical bases, easily extendable to accommodate new bases; API interface for management applications that use integrated data; Digital platforms with vertical information on the installed capacity of SUS and supplementary health at macro and micro level, agile analysis and related to demographic and mortality trends; Validated algorithms for extracting information from medical records for the construction of performance indicators and the outcome of specific health conditions; Automated systems for audits of procedures and examinations; Integrated database for monitoring technologies incorporated in SUS.

back


References:

9. Ribeiro ALP, Paixão GMM, Gomes PR, Ribeiro MH, Ribeiro AH, Canazart JA, et al. Tele-electrocardiography and bigdata: The CODE (Clinical Outcomes in Digital Electrocardiography) study. J Electrocardiol [Internet]. 2019 Sep 7; Available from: http://dx.doi.org/10.1016/j.jelectrocard.2019.09.008

93. dal Bianco G, Gonçalves MA, Duarte D. BLOSS: Effective meta-blocking with almost no effort. Inf Syst. 2018 Jun 1;75:75–89.

100. Candido D da S, Claro IM, de Jesus JG, Souza WM de, Moreira FRR, Dellicour S, et al. Evolution and epidemic spread of SARS-CoV-2 in Brazil [Internet]. Infectious Diseases (except HIV/AIDS). medRxiv; 2020. Available from: https://www.medrxiv.org/content/10.1101/2020.06.11.20128249v1

101. Guzzetta G, Marques-Toledo CA, Rosà R, Teixeira M, Merler S. Quantifying the spatial spread of dengue in a non-endemic Brazilian metropolis via transmission chain reconstruction. Nat Commun. 2018 Jul 19;9(1):2837.

102. Marini G, Guzzetta G, Marques Toledo CA, Teixeira M, Rosà R, Merler S. Effectiveness of Ultra-Low Volume insecticide spraying to prevent dengue in a non-endemic metropolitan area of Brazil. PLoS Comput Biol. 2019 Mar;15(3):e1006831.

103. Marques-Toledo CA, Bendati MM, Codeço CT, Teixeira MM. Probability of dengue transmission and propagation in a non-endemic temperate area: conceptual model and decision risk levels for early alert, prevention and control. Parasit Vectors. 2019 Jan 16;12(1):38.

114. Nascimento BR, Martins JFBS, Nascimento ER, Pappa GL, Sable CA, Beaton AZ, et al. Deep learning for automatic identification of rheumatic heart disease in echocardiographic screening images: data from the ATMOSPHERE-PROVAR study. J Am Coll Cardiol. 2020 Mar 24;75(11, Supplement 1):3577.

118. Canuto S, Salles T, Rosa TC. Similarity-Based Synthetic Document Representations for Meta-Feature Generation in Text Classification. Proceedings of the 42nd [Internet]. 2019; Available from: https://doi.org/10.1145/3331184.3331239

119. Cardoso TNC, Silva RM, Canuto S, Moro MM, Gonçalves MA. Ranked batch-mode active learning. Inf Sci . 2017 Feb 10;379:313–37.

120. Magalhães LFG, Gonçalves MA, Canuto SD, Dalip DH, Cristo M, Calado P. Quality assessment of collaboratively-created web content with no manual intervention based on soft multi-view generation. Expert Syst Appl. 2019 Oct 15;132:226–38.

121. Freire J, Chirigati F. Provenance and the different flavors of computational reproducibility. IEEE Data Engineering Bulletin. 2018;41(1):15.

122. Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, Chou K, et al. A guide to deep learning in healthcare. Nat Med. 2019 Jan;25(1):24–9.

123. Waring J, Lindvall C, Umeton R. Automated machine learning: Review of the state-of-the-art and opportunities for healthcare. Artif Intell Med. 2020 Apr;104:101822.

124. Phillips KA, Trosman JR, Douglas MP. Emergence of Hybrid Models of Genetic Testing Beyond Direct-to-Consumer or Traditional Labs [Internet]. Vol. 321, JAMA. 2019. p. 2403. Available from: http://dx.doi.org/10.1001/jama.2019.5670

125. Palazzo S, Spampinato C, Kavasidis I, Giordano D, Schmidt J, Shah M. Decoding Brain Representations by Multimodal Learning of Neural Activity and Visual Features. IEEE Trans Pattern Anal Mach Intell. 2020;1–1.

126. Deepak S, Ameer PM. Brain tumor classification using deep CNN features via transfer learning. Comput Biol Med. 2019 Aug;111:103345.

127. Neto JLD, Yu S, Macedo DF, Nogueira JMS, Langar R, Secci S. ULOOF: A User Level Online Offloading Framework for Mobile Edge Computing. IEEE Trans Mob Comput. 2018 Nov;17(11):2660–74.

128. Teodoro G, Hartley TDR, Catalyurek U. Run-time optimizations for replicated dataflows on heterogeneous environments. Proceedings of the 19th [Internet]. 2010; Available from: https://dl.acm.org/doi/abs/10.1145/1851476.1851479

129. Cesar TF, Vieira LFM, Vieira MAM, Neto OPV. Cellular automata-based byte error correction in QCA. Nano Commun Netw. 2020 Feb 1;23:100278.

130. Matheus L, Pires L, Vieira A, Vieira LFM, Vieira MAM, Nacif JA. The internet of light: Impact of colors in LED-to-LED visible light communication systems [Internet]. Vol. 2, Internet Technology Letters. 2019. p. e78. Available from: http://dx.doi.org/10.1002/itl2.78

131. Boito FZ, Inacio EC, Bez JL, Navaux POA. A checkpoint of research on parallel i/o for high-performance computing. ACM Computing [Internet]. 2018; Available from: https://dl.acm.org/doi/abs/10.1145/3152891

132. Moura HD, Macedo DF, Vieira MAM. Wireless control using reinforcement learning for practical web QoE. Comput Commun. 2020 Mar 15;154:331–46.

133. Malta DC, Felisbino-Mendes MS, Machado ÍE, Passos VM de A, Abreu DMX de, Ishitani LH, et al. Risk factors related to the global burden of disease in Brazil and its Federated Units, 2015. Rev Bras Epidemiol. 2017 May;20Suppl 01(Suppl 01):217–32.

134. Alkmim MB, Silva CBG, Figueira RM, Santos DVV, Ribeiro LB, da Paixão MC, et al. Brazilian National Service of Telediagnosis in Electrocardiography. Stud Health Technol Inform. 2019 Aug 21;264:1635–6.

135. Ribeiro AH, Ribeiro MH, Paixão GMM, Oliveira DM, Gomes PR, Canazart JA, et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nat Commun. 2020 Apr 9;11(1):1760.

136. Lopes EL, Beaton AZ, Nascimento BR, Tompsett A, Dos Santos JP, Perlman L, et al. Telehealth solutions to enable global collaboration in rheumatic heart disease screening. J Telemed Telecare. 2018 Feb;24(2):101–9.

137. Nascimento BR, Beaton AZ, Nunes MCP, Diamantino AC, Carmo GAL, Oliveira KKB, et al. Echocardiographic prevalence of rheumatic heart disease in Brazilian schoolchildren: Data from the PROVAR study. Int J Cardiol. 2016 Sep 15;219:439–45.

138. Roy M, Wang F, Vo H, Teng D, Teodoro G, Farris AB, et al. Deep-learning-based accurate hepatic steatosis quantification for histological assessment of liver biopsies. Lab Invest [Internet]. 2020 Jul 13; Available from: http://dx.doi.org/10.1038/s41374-020-0463-y

139. ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium. Pan-cancer analysis of whole genomes. Nature. 2020 Feb;578(7793):82–93.

140. Olivier M, Asmis R, Hawkins GA, Howard TD, Cox LA. The Need for Multi-Omics Biomarker Signatures in Precision Medicine. Int J Mol Sci [Internet]. 2019 Sep 26;20(19). Available from: http://dx.doi.org/10.3390/ijms20194781

141. Gregório C, Soares-Lima SC, Alemar B, Recamonde-Mendoza M, Camuzi D, de Souza-Santos PT, et al. Calcium Signaling Alterations Caused by Epigenetic Mechanisms in Pancreatic Cancer: From Early Markers to Prognostic Impact. Cancers [Internet]. 2020 Jun 30;12(7). Available from: http://dx.doi.org/10.3390/cancers12071735

142. Hashimoto DA, Rosman G, Rus D, Meireles OR. Artificial Intelligence in Surgery: Promises and Perils. Ann Surg. 2018 Jul;268(1):70–6.

143. Schonberger JL, Frahm J-M. Structure-from-motion revisited. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016. p. 4104–13.

144. Weinzaepfel P, Revaud J, Harchaoui Z, Schmid C. DeepFlow: Large displacement optical flow with deep matching. In: Proceedings of the IEEE international conference on computer vision. 2013. p. 1385–92.

145. Fida B, Cutolo F, di Franco G, Ferrari M, Ferrari V. Augmented reality in open surgery. Updates Surg. 2018 Sep;70(3):389–400.

146. Hung AJ, Chen J, Gill IS. Automated Performance Metrics and Machine Learning Algorithms to Measure Surgeon Performance and Anticipate Clinical Outcomes in Robotic Surgery. JAMA Surg. 2018 Aug 1;153(8):770–1.

147. Taylor RH, Menciassi A, Fichtinger G, Fiorini P, Dario P. Medical Robotics and Computer-Integrated Surgery. In: Siciliano B, Khatib O, editors. Springer Handbook of Robotics. Cham: Springer International Publishing; 2016. p. 1657–84.

148. Whittaker G, Aydin A, Raison N, Kum F, Challacombe B, Khan MS, et al. Validation of the RobotiX Mentor Robotic Surgery Simulator. J Endourol. 2016 Mar;30(3):338–46.

149. Queiroz OV de, Guerra Júnior AA, Machado CJ, Andrade ELG, Meira Júnior W, Acúrcio F de A, et al. A construção da Base Nacional de Dados em Terapia Renal Substitutiva (TRS) centrada no indivíduo: relacionamento dos registros de óbitos pelo subsistema de Autorização de Procedimentos de Alta Complexidade (Apac/SIA/SUS) e pelo Sistema de Informações sobre Mortalidade (SIM) – Brasil, 2000-2004. Epidemiol Serv Saúde. 2009;18(2):107–20.

150. Saldanha R de F, Xavier DR, Carnavalli K de M, Lerner K, Barcellos C. [Analytical study of the breast cancer patient flow network in Brazil from 2014 to 2016]. Cad Saude Publica. 2019 Jul 22;35(7):e00090918.