Artificial intelligence (AI) is already ubiquitous in everyday tasks like finding the best traffic route, choosing the most affordable travel package, or providing customer service. Now, it is gradually making inroads into healthcare. The World Health Organization (WHO) sees AI as holding “enormous potential for improving the health of millions of people around the world.” According to WHO’s “Ethics and Governance of Artificial Intelligence for Health” report, AI “can be, and in some wealthy countries is already being used to improve the speed and accuracy of diagnosis and screening for diseases; to assist with clinical care; strengthen health research and drug development; and support diverse public health interventions, such as disease surveillance and health systems management.”
AI refers to the ability of electronic devices to mimic the way humans perceive and respond to different situations, make choices, and solve problems. The software is the logical component of the device—its “brain”—consisting of sequences of instructions called algorithms. Initially algorithms followed pre-programmed instructions, but today they are trained to recognize patterns on their own using the data they are fed. This is known as machine learning (see the glossary of technical terms).
Algorithm: a well-defined computer model that transforms inputs into outputs.
Artificial intelligence: a field that encompasses a variety of techniques for solving problems, including heuristics, probabilistic methods, and machine learning.
Machine learning: a subfield of artificial intelligence that involves training a computational model using data, often by providing an expected response given particular inputs. Over multiple iterations, the model learns by adjusting its internal parameters to minimize the difference between its predictions and true outcomes. Each machine-learning method defines what parameters are to be learned and how they transform inputs into outputs.
Deep learning: a modern machine-learning approach that concatenates multiple learning layers with each other. Deep-learning models often have a large number of parameters that are learned from data, and require large amounts of it to be trained effectively.
Big data: the processing of vast amounts of structured and unstructured data, often associated with deep learning.
Transfer learning: a machine-learning technique in which a model that has been trained in one domain is adapted or repurposed for another. In the first stage of transfer learning the model is trained to complete partially populated datasets in a self-supervised manner, using only input data and no ground-truth labels. The second stage involves fine-tuning the model for a specific set of inputs and outputs.
Sources Marcelo Finger (USP) and Alexandre Falcão (UNICAMP)
Machine-learning algorithms can quickly access and analyze vast amounts of data, identify patterns, and suggest solutions with a higher accuracy rate than humans can. Experts are quick to note, however, that healthcare will continue to demand highly trained professionals, and AI systems are not being developed to replace doctors. “Instead, AI’s powerful analytical capabilities will be used as a tool to aid physicians in decision-making,” says Alexandre Dias Porto Chiavegatto Filho, an economist and professor of artificial intelligence in healthcare, who heads the Laboratory for Big Data and Predictive Analytics in Healthcare (LABDAPS), created in 2017 at the School of Public Health at the University of São Paulo (USP), with funding from FAPESP.
Alphabet, the technology giant and parent company of Google, is currently the world’s leading investor in AI research for healthcare. The Massachusetts Institute of Technology (MIT), Stanford, and Harvard, in the US, and Oxford and Cambridge, in the UK, are other leading research players in this field. In Brazil, USP, the University of Campinas (UNICAMP), and the Federal University of Minas Gerais (UFMG) have some of the largest research programs in AI.
According to the “Artificial Intelligence Index Report 2022” from Stanford University, private-sector investment in AI research for medicine and healthcare was US$11.3 billion worldwide in 2021, an increase of 40% on the previous year. In the past five years, research funding has totaled US$28.9 billion, making this segment the largest recipient of private AI investment, surpassing traditional information technology applications such as fintech and retail. Deep-learning computer vision systems capable of detecting and segmenting organs, lesions, or tumors are some of the applications that have attracted the most interest from the medical community.
One of the most acclaimed recent developments in AI for life sciences is AlphaFold, a software system created by Google subsidiary DeepMind. Named by Science as the 2021 “Breakthrough of the Year,” the program uses deep-learning techniques to model and solve what is known as the “protein folding problem.” Proteins are made up of amino acid chains that fold spontaneously into three-dimensional (3D) structures. Because the 3D shape of a protein determines its biological function, an understanding of protein structure is crucial for a range of applications, from understanding the molecular basis of life to discovering new drugs. With AlphaFold, researchers can simply input the amino acid sequence of a protein and the program will output its predicted folded structure. Over half a million researchers have used the program to create a diverse set of solutions, from addressing antibiotic resistance to tackling plastic pollution.
Despite the substantial investment and recent progress in AI research, a number of technical, ethical, and legal hurdles must still be overcome to increase uptake in healthcare. One concern is the potential liability for misdiagnoses involving AI software and algorithms. Another concern is the potential unethical use of information, such as breaches of patient privacy or the use of algorithms that reinforce biases or perpetuate health inequalities.
Moreover, AI must gain the trust of the medical community as being safe and reliable. “It is still very incipient. It will take time for AI technology to be widely adopted in the healthcare sector in Brazil and globally,” says Chiavegatto. The first challenge needing to be addressed is improving the quality of data used to train the algorithms. Healthcare systems generate large amounts of raw data, but very little curated information. “Incomplete and inconsistent data leads to bad algorithms,” says the researcher.
For everyday things like travel routes or movie recommendations, a misplaced AI recommendation can be frustrating; in healthcare, it can be life-threatening. Every new AI system designed for medical applications needs to demonstrate its algorithms are reliable in order to earn the trust of users.
Another challenge is proving that AI is effective in improving clinical decisions. “There is currently no scientific evidence that medical prescriptions issued with the aid of AI are necessarily more accurate than those made without it,” says Chiavegatto.
One of his research programs at LABDAPS is developing new approaches to curating training data. One of the lab’s recent studies found that algorithms trained using local data performed better than those using a broader dataset gathered from populations with diverse genetic and demographic traits.
The study was designed specifically to develop algorithms to predict mortality rates among patients with Covid-19. The researchers collected data on 16,236 people admitted to 18 hospitals in different regions of Brazil from March to August 2020. Using this data, they tested eight different strategies to develop predictive models. The best results came from models that used local data, and the researchers found that incorporating data from patients in other regions made these models’ predictions less accurate.
“We found that an algorithm trained in New York may not be as well suited for use in Miami or San Francisco,” explains Chiavegatto. A paper about the study has been submitted for publication in a scientific journal. “These findings have global implications. Many algorithms developed in one country may need to be retrained using transfer-learning techniques,” he predicts. Transfer learning is a research problem in machine learning (ML) that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem.
One recent initiative to expand AI deployment in healthcare in Brazil was launched as a partnership between FAPESP, the Brazilian Ministry of Science, Technology, and Innovation (MCTI), the Ministry of Communications, and the Brazilian Internet Steering Committee (CGI.br). Following a public call for project proposals, six Applied Research Centers (CPA) in AI were established in 2021, focusing on applications in healthcare as well as agriculture, manufacturing, and smart cities. Each center will receive R$1 million in funding per year for up to 10 years. This will be matched by private partners for a total of R$20 million in funding per center.
Two of the new CPAs will specifically focus on healthcare, hosting researchers who are already engaged in developing AI solutions for the sector. The Center for Innovation in Artificial Intelligence for Health (CIIA-Saúde), based at the UFMG campus in Belo Horizonte, brings together eight other institutions and four partner companies under its umbrella. The center will pursue research across five thematic areas: prevention and quality of life; diagnosis, prognosis, and tracking; personalized therapeutic medicine; healthcare systems and management; and epidemics and disasters. “We plan to launch our first call for projects by the end of the year,” says Antonio Luiz Pinho Ribeiro, the center’s vice director, who also heads the Telehealth Center at UFMG’s teaching hospital.
Data collected from the roughly 4,000 electrocardiograms (ECGs) performed per day at the Telehealth Center have been used in two studies on AI applications in cardiovascular disease recently reported in Nature Communications. In their first paper, published in 2020, the research team described an automated diagnosis system for standard ECGs, known as 12-lead ECGs. The system was trained to recognize six types of ECG abnormalities using deep learning, with a sensitivity (the ability to accurately identify when a patient has a disease) of 80% and a specificity (the ability to correctly identify patients without the disease) of 99%.
The second study, published in 2021, used AI to create a measure of cardiovascular health called ECG-age. The system analyzes an individual’s ECG and determines their likely age based on heart health. People with an ECG-age that exceeds their chronological age are believed to be at a higher risk of death from cardiovascular disease. “This can be useful in disease prediction and management,” says Ribeiro. “But the AI system won’t be ready for commercial use until it is shown to effectively support medical decisions and patient well-being.”
The Brazilian Institute of Data Science (BI0S), another newly created CPA, brings together 12 public and private partners under the coordination of UNICAMP. Research at BI0S will be focused on the areas of healthcare, agribusiness, and “method,” or applied research in data science for multiple applications. Rodolfo Pacagnella, a professor of obstetrics at UNICAMP who leads the healthcare department at BI0S, says three lines of research have been defined: women’s health, skin cancer diagnosis, and lung cancer diagnosis. Sandra Eliza Fontes de Ávila, at UNICAMP’s Institute for Computing Science, has developed a software system to detect melanoma, the most aggressive type of skin cancer—one of the most common forms of cancer in the world (see Pesquisa FAPESP issue no. 305).
In the area of women’s health, Pacagnella is leading research on the use of AI to identify patterns indicating an increased risk of maternal death and premature births, supporting the design of predictive strategies. “Early detection of health issues not only allows for better disease management and treatment, but can also improve the management of healthcare systems. For example, we can predict when more medical staff or intensive care unit beds will be needed in maternity hospitals,” explains Pacagnella.
Two FAPESP-supported startups have developed AI technology to aid in cancer diagnosis. Onkos Diagnósticos Moleculares, a startup incubated at the Supera Innovation and Technology Park in Ribeirão Preto, São Paulo State, has developed a genetic test called mirTHYpe to improve the classification of indeterminate thyroid nodules. Traditional testing methods, such as fine needle aspiration (FNA) biopsies, are unable determine whether a tumor is malignant or benign in approximately 25% of cases. As a precaution, all patients with indeterminate nodules are referred for diagnostic surgery, but 80% of these surgeries are ultimately unnecessary because the nodules turn out to be benign. This can have long-term consequences for patients, such as the need for lifelong hormone replacement.
Launched in 2018, mirTHYpe uses AI to analyze microRNA—small RNA (ribonucleic acid) molecules that regulate other genes in the human genome—in the nodules to determine whether the lesion is benign or malignant. This helps to avoid unnecessary surgery (see Pesquisa FAPESP issue no. 264). The Onkos research team conducted a study on 440 patients at 128 labs that are currently using the method, and found that mirTHYpe influenced 92% of medical decisions and reduced unnecessary surgeries by 75%. The study was published in The Lancet Discovery Science (eBioMedicine). “With the larger database we have built in recent years, we are now retraining our algorithm and believe we can avoid up to 89% of unnecessary diagnostic surgeries,” says Onkos founder Marcos Tadeu dos Santos.
Artificial Intelligence must gain the trust of the medical community as being safe and reliable
The second startup, Harpia Health Solutions, based in the São José dos Campos Technology Park, has developed an online platform called Delfos that hosts AI and computer vision solutions for analyzing medical imaging data. “Our platform acts as a source of a second opinion for radiologists,” says Daniel Aparecido Vital, a biomedical engineer and cofounder of Harpia.
The platform receives medical images via an application programming interface (API) and responds within five minutes, uploading its findings directly into the image archiving and sharing system (PACS/RIS) used by the radiologist. In those five minutes, the platform automatically identifies and ranks abnormal findings by order of priority for the radiologist to focus on. The platform also uploads an image showing the abnormalities.
“This method increases diagnostic productivity and accuracy by freeing up radiologists to focus on pre-detected abnormalities and reducing errors caused by fatigue from long workdays,” says Vital. Harpia’s first commercially available service is Delfos, a mammography solution that has already processed over 72,000 mammograms. The company is also developing a solution for chest X-rays, which is currently in the validation phase.
Artificial intelligence systems are also being used to improve hospital management. One example is Neonpass, a system that seamlessly tracks a patient’s journey within a healthcare facility. The technology was developed by Hoobox with FAPESP grant funding, and has been successfully implemented at the Albert Einstein and Sírio-Libanês hospitals in São Paulo.
“At most hospitals, patient and visitor movements are tracked using different systems, which each have their own requirements and often do not communicate with each other effectively,” says Hoobox cofounder Paulo Pinheiro. One issue is that patients must provide their identification at the reception desk, and then may be asked for it again at later stages of the patient journey. This disjointed process can lead to loss of data on the duration, volume, and reasons for patient visits, undermining patient care.
“Neonpass is designed to be a single hub that seamlessly orchestrates all patient interactions, and does so more efficiently and without any data loss,” explains Pinheiro. Some of the tasks that can be mediated by Neonpass include check-in (both in person and online), visitor registration, and terminal use. Real-time tracking information allows managers to see how many patients and visitors are in each area, which reception areas are busiest, and the average waiting time for consultations or medical procedures.
Hoobox is also integrating a new module called Sadia into Neonpass. This system uses computer vision and artificial intelligence to detect the risk of patients falling out of bed or developing pressure ulcers—a risk that increases when a person remains in the same position for more than two hours on end. Information is captured by a bedside camera, and machine-learning techniques are used to identify the patient, predict fall risk, and activate appropriate nursing care.
The system also generates other information for use by hospital management, including the type of professional performing bedside procedures and the total number of hours of care. The solution is compliant with Brazil’s new General Data Protection Regulation (GDPR): data is processed and then promptly deleted to ensure patient privacy. Over 100 patient beds are currently being monitored using the Hoobox system.
Clinical decision support algorithms need to be transparent and avoid biases
In a 2021 report on the use of artificial intelligence to improve healthcare delivery and support drug development, the World Health Organization (WHO) cautions developers on the need to put ethics and human rights “at the center of the design, development, and deployment of AI technologies for health.”
The report, titled “Ethics and Governance of Artificial Intelligence for Health,” points to the risk that healthcare data could be used in unethical ways, and especially the risk that algorithms could encode biases, or information could be disclosed that breaches patient privacy. It also addresses cybersecurity risks and the need for AI-generated data to be transparent, explainable, and intelligible.
“As accurate as an algorithm might be, it can still make mistakes. To be trusted, an algorithm needs to transparently present to the physician not only the decision itself, but also the rationale behind it,” says Antônio Luiz Pinho Ribeiro, deputy director at the Center for Innovation in Artificial Intelligence for Health (CIIA-Saúde) in Belo Horizonte, Minas Gerais State.
Rodolfo Pacagnella, who heads the healthcare department at the Brazilian Institute of Data Science (BI0S), notes the importance of using representative data in training algorithms. “Will an algorithm trained to detect skin cancer in Norwegians work just as well for Brazilians?” he asks.
He also highlights the importance of training algorithms using comprehensive data covering populations with different profiles, and preventing racial, social, and economic biases from contaminating algorithms, which can skew the data unnoticeably in AI systems.
Another concern is the way data is collected for machine learning. “An algorithm trained with data collected under ideal conditions in academic laboratories may not accurately reflect conditions on the ground in healthcare systems. An AI program should be validated in real-world use cases before it is launched for commercial use. This is often neglected,” he notes.
1. Bios – Brazilian Institute of Data Science (no. 20/09838-0); Grant Mechanism Engineering Research Centers (CPE); Agreement MCTI/MC; Principal Investigator João Marcos Travassos Romano (UNICAMP); Investment R$2,180,218.21.
2. Center for Innovation in Artificial Intelligence for Medicine (Ciia-Saúde) (no. 20/09866-4); Grant Mechanism Engineering Research Centers (CPE); Agreement MCTI/MC; Principal Investigator Virgílio Augusto Fernandes Almeida (UFMG); Investment R$1,659,839.04.
3. Molecular classification of indeterminate thyroid nodules via microRNA profiling (no. 15/07590-3); Grant Mechanism Innovative Research in Small Businesses (PIPE); Principal Investigator Marcos Tadeu dos Santos (Onkos); Investment R$832,545.15.
4. Computational methods using machine learning for automatic identification and classification of breast lumps and microcalcifications in digital mammography exams (no. 19/16514-0); Grant Mechanism Innovative Research in Small Businesses (PIPE); Principal Investigator Daniel Aparecido Vital (Harpia); Investment R$101,040.80.
5. Unlimited imaging of skin lesions using Generative Adversarial Networks (no. 19/19619-7); Grant Mechanism Doctoral (PhD) Fellowship; Supervisor Sandra Eliza Fontes de Avila (UNICAMP); Beneficiary Alceu Emanuel Bissoto; Investment R$267,948.89.
RIBEIRO, A. H. et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nature Communications. Apr. 9, 2020.
LIMA, E. M. et al. Deep neural network-estimated electrocardiographic age as a mortality predictor. Nature Communications. Aug. 25, 2021.
SANTOS, M. T. et al. Clinical decision support analysis of a microRNA-based thyroid molecular classifier: A real-world, prospective and multicentre validation study. The Lancet Discovery Science (eBioMedicine). June 30, 2022.
JUMPER, J. et al. Highly accurate protein structure prediction with AlphaFold. Science. July 15, 2021.