Since its emergence in late 2019 and early 2020, SARS-CoV-2 has sickened millions of people and killed thousands, testing the limits of science, research, and public health systems. During this time, there has been an international effort to understand this emerging virus and its associated disease, COVID-19. In the race against the pandemic, some of the tools and techniques that have held emerging promise in healthcare, including artificial intelligence (AI), machine learning (ML), and deep learning (DL), (Figure 1) are being repurposed for battling coronavirus. But are these just hype, or is there real promise that like telehealth (NODE.Health article, June 18, 2020), the pandemic could dramatically accelerate their use in healthcare?
|Artificial Intelligence(AI): Science of computer systems that can perform tasks normally requiring human intelligence.||A general concept of machine intelligence which includes ML and DL algorithms. One of the first examples were expert systems, which were adopted in the 1970s by Stanford researchers to help assist medical diagnoses of severe bacterial infections (MYCIN).|
|Machine Learning(ML): Study of algorithms that use sample data to learn a model to predict.||Many categories, including supervised, unsupervised, and transfer learning. Some common learning models include regression, random forest and Naive Bayes.|
|Deep Learning(DL): ML techniques inspired by neural networks in the brain.||Rose to prominence in the last decade due to successes in image recognition among others. Some common models include convolutional neural network models (CNNs) and recurrent neural networks (RNNs).|
The emergence of AI-based models and techniques in healthcare preceded COVID-19. By some estimates, the global healthcare market for AI in 2020 is $4.9B, and is expected to grow to $45.2B by 2026.(1) COVID-19 has had the unexpected effect of accelerating research on the use of AI tools in healthcare. How these tools are researched and implemented during the pandemic will greatly advance their understanding and utility in healthcare and help accelerate our understanding of whether they may measure up to the considerable hype they have had in the last few years.
Is there evidence for the safety and efficacy of these tools as it pertains to COVID-19? Several important domains in which AI, ML, and DL are emerging in healthcare include patient risk stratification, lung imaging, and both vaccine and drug development.
The gold standard for diagnosis of COVID-19 remains real time reverse transcription polymerase chain reaction (rT-PCR) following nasal or pharyngeal swab. However, deciding who and when to test has largely been based on clinical judgment. AI models have been developed to help better assess the pre-test probability that a patient may have infection based on a variety of parameters including age, race, ethnicity, body mass index, smoking status, gender, zip code, symptoms, pneumonia and influenza vaccination status, medication use, and certain lab values. A study of thousands of patients conducted by the Cleveland Clinic in one population and validated in a geographically distinct population, had good concordance between predicted and actual infection (free online risk calculator available here: https://riskcalc.org/COVID19/).(2) Such models can be useful in assisting clinicians in the deciding on the most appropriate testing. Furthermore, by “enhancing” the likelihood that the population being tested actually has infection (namely, increasing the prevalence of infection in the tested population relative to the general population), the positive predictive value of the test increases, an important approach that may help reduce false test results.
AI models have also been used to predict the clinical trajectory of patients diagnosed with COVID-19, including those who are at greatest risk of developing acute respiratory distress syndrome (ARDS). In many patients, development of ARDS can occur quickly and precipitously. Knowing which patients are likely to deteriorate quickly in advance can assist clinicians in triaging hospital beds (ICU, stepdown, or acute care) and resources to those most at risk. In a small study of 52 patients from Wenzhou, China, researchers at New York University used AI models to predict development of ARDS in COVID-19 patients based on patients’ initial presentation, including lab tests and symptoms. The predictive models achieved an accuracy of 70-80% in predicting severe cases.(3)
A similar study deployed AI tools that drew upon Epic’s electronic health record (EHR) Deterioration Index (DI) to validate the algorithm for prediction accuracy using 3 hospital systems and 327,000 patients. Led by Dr. Ron Li, the team at Stanford used the DI to research the change in DI as an important factor in validating the ability of the model to predict patient Intensive Care Unit (ICU) transfers, rapid response events, and death in order to identify sick patients at risk of rapid deterioration in the hospital. (4)
AI techniques have gained their widest adoption in healthcare in medical imaging. Many of the algorithms being explored pre-COVID are being repurposed for identifying COVID features on imaging. For example, DL models such as convolutional neural networks (CNN), are used for lung segmentation to more accurately represent lung changes due to COVID-19 on lung imaging studies such as chest radiographs (CXR) and lung computed tomography (CT) scans. (5) One challenge applying DL methods in medical imaging, however, is that the models need large, uniform training datasets, which are generally lacking for COVID-19. To bypass the challenge, an approach is to train DL models on pre-existing larger datasets for other diseases that appear similar on imaging studies to findings of patients with COVID-19. For example, in a study published in the Journal of Thoracic Imaging, John Hopkins University researchers used a pre existing DL model trained on a larger tuberculosis (TB) dataset for classifying chest radiographs (CXRs) of patients with suspected COVID-19. In the proof-of-concept study, they found that their algorithm correctly predicted positive COVID-19 cases 89% of the time on a test set of 88 CXR’s. (6) Another team created the first open source DL solution, a CNN named COVID-net, which was designed to help radiologists identify changes on CXR from different types of pneumonias. (7) The vision for such open source DL solutions is to invite other researchers to assist with and strengthen the proposed models and algorithms.
A different study tried to answer the difficult question of whether a deep learning model trained on lung CT images could diagnose pulmonary changes specific to COVID-19, and distinguish from, e.g., community acquired pneumonia (CAP). The multi center study examined 4356 CTs from 3322 patients and created a CNN model, COVID-19 detection neural network (COVNET), in order to differentiate COVID-19 and CAP from chest CT images. (8) The study claimed a high level of specificity and sensitivity for COVID positive findings. The drawback of the study was that it was compared with CAP only and not tested with other types of pneumonias that can more clearly establish more disease etiology. The extraction technique also could not establish disease severity.
Using an approach that additionally leverages medical records, researchers at Mt. Sinai Medical Center created an ML algorithm based on CT imaging combined with clinical data such as findings from lab tests, symptoms and exposure history, as a clinical decision support tool to determine the probability of a patient having COVID-19. (9) As the diagnosis of pneumonia of any source is not a purely radiologic diagnosis, this approach mirrors typical diagnostic algorithms with multiple inputs weighted with radiology findings. Further, this approach confirms positive COVID-19 diagnoses in the cases with an absence of radiological features on chest CT scans. (10) Although the study sample was small (905 participants), it showed the potential to diagnose COVID-19 patients rapidly if other, standard methods of detection were unavailable.
An important drawback of such studies is that the trained algorithms are not specific to COVID-19 and therefore cannot necessarily distinguish COVID-19 from other diseases that have similar imaging findings. Although AI models have been deployed on lung imaging for their potential to spot COVID-related findings and augment physician opinion, the American College of Radiology issued a position statement in March 2020, stating that it did not recommend using imaging for screening or diagnosis of COVID-19, indicating that viral testing remains the only specific means of diagnosis. (11) As of the date of this article, that position statement remains unchanged, suggesting that AI for imaging in the context of COVID-19 may be more demonstrative of AI capability rather than its current clinical utility.
AI tools may help shorten the time and costs involved to find effective drug targets and treatments for COVID-19, when compared to traditional experimental methods, especially when supplemented by a coordinated global effort for open source data.
One open source effort is the Global Initiative on Sharing All Influenza Data (GISAID), which is a public-private partnership between the US, Singapore, and Germany in order to identify and sequence all influenza subtypes, including SARS-CoV-2.(12) Another initiative run by the Research Collaboratory for Structural Bioinformatics protein data bank (RCSB PDB) shares experimentally determined and computationally predicted protein structures, whose molecular details are important to help depict the structure of SARS-CoV-2. (13) Both of these open databases are helping to fuel drug discovery by providing the necessary input data for AI tools.
Using these open source tools databases, Alphabet’s Deepmind repurposes their open source AI-based algorithm, Alphafold, originally designed for 3D protein structure discovery, for analyzing COVID-19 proteins (14). Such analyses are crucial for identifying targets for a potential COVID-19 drug.
Another open source project, AI Cures, is a collaboration of life science and computational researchers based out of Massachusetts Institute of Technology (MIT), that draws on data from existing drug molecules and applies AI methods for fast prediction of the efficacy or repurposing of existing drugs. (15)
AI techniques may be used by researchers for modeling viruses such as SARS-CoV-2 in order to distinguish which immunogenic components of these viruses can be used to create vaccines (16) A Stanford study supplemented computational biology tools with pretrained DL neural networks, MARIA and netMHCPan in order to identify the virus’ protein fragments on T-cell and B-cell epitopes based on protein antigen presentation and antibody binding properties. (17, 18) The study found that the most promising vaccine candidates are centered around the spike protein of the virus, which gives it a characteristic appearance similar to other coronaviruses as well as enables it to enter certain human cells. The purpose of the study was to identify the top candidates for epitope-based COVID-19 vaccines as well as T-cell responses. (18)
In April, the Harvard School of Public Health partnered with the Human Vaccines Project to create the Human Immunomics Initiative (HII). The purpose of this joint effort is to decode the human immune system which will help speed the process of development of new vaccines. (19) Initially, HII will study how immunity is affected by age. (19) This research is especially important given the way that COVID-19 has been shown to be more dangerous in older populations over 65.
AI, ML, and DL tools are evolving in healthcare, and their application to COVID-19 has accelerated their potential uses, but is still largely in a proof-of-concept stage. One should not expect AI to replace traditional diagnostic approaches yet, but their use for clinical decision support in COVID-19 is beginning to be appreciated. In order for AI to further help the diagnosis and treatment of diseases requiring rapid response, like COVID-19, there needs to be a concerted international effort of data sharing, integrity and standardization similar to GISAID to build and validate AI-based models.
AI-based methods are not standalone solutions. Rather, they can be very useful adjuncts for clinical decision making and drug and vaccine development. In addition to the challenge of feeding the AI-based algorithms large high-quality datasets, crucial AI factors that need to be achieved include performance, robustness, and interpretability. Notably these factors are dependent on the integrity and volume of the training data as well as the right application of AI tools and techniques to a properly framed problem in healthcare.
NODE.Health Foundation is a 501(c)(3) non-profit organization dedicated to education, validation and dissemination of evidence based digital medicine. As the largest professional association in digital medicine, NODE.Health empowers societies, executives and NODES from health systems, payers, life sciences, venture capital, startups and the public sector involved in healthcare digital transformation. NODE.Health does not endorse any specific products or services.
NODE.Health is pleased to cross post this article giving examples of the use of artificial intelligence (AI) in healthcare specifically for COVID-19. NODE.Health encourages its readers to be diligent with selecting such tools and understand the evidence. As more evidence comes out on the use of AI for COVID-19 and beyond, NODE.Health will keep its readers informed about the latest developments. Interested in learning more about the Network of Digital Evidence (NODE.Health)? Click here