AI for COVID-19
Few would disagree that there are big inefficiencies and deficiencies in healthcare systems across the world, many of which were known before this pandemic. Those include lack of accessibility to healthcare services, physician shortage and burnout, and medical errors. Covid-19 has not only exposed, but also augmented all these shortcomings, putting the whole medical system through an unprecedented stress. Because of this extraordinary situation, many are now looking to telemedicine services and the magic bullet that they promise. Traditional telemedicine, however, where a patient is connected to a physician over video chat, is mostly a different medium for the same processes and practices we have had for decades (if not centuries) in the traditional healthcare system. In fact, very importantly, telemedicine does not scale or improve quality by itself. If you have been following us at Curai for some time, you will know that scaling telemedicine through AI and automation was precisely what we were working on when this crisis broke out.
In this post, we will describe how we are tackling some of the immediate COVID-19 needs by introducing new offerings including COVID-19 risk assessment and antibodies testing. We will also discuss some of the longer-term research that we are now adapting to the COVID-19 context to improve the quality and the scalability of our response.
Before we dive into some of our COVID-related efforts though, let us add some context by quickly describing what our general approach to providing quality care is.
About Curai
At Curai we are building an end-to-end medical platform that is always-on, accessible, and affordable for everyone. We combine Artificial Intelligence and human providers to accomplish both scale and quality of care. We call this combination “Augmented Intelligence”. We design our systems to be always-learning from the data and to operate “in-the-wild”. The patient interacts with both AI and human providers in a transparent way that makes it explicit to the user. Human providers see their abilities and efficiency augmented by AI and automation features that are fully integrated into our FHIR-based home-grown Electronic Health Record (EHR). Our AI/automation features combine different approaches and components, from the latest deep learning techniques to expert systems, all of which are connected through our internal knowledge base.
So, how does this work in practice? As you can see in the screenshots below, it is a primarily chat-based service that allows for synchronous and asynchronous communication. The service is supported by two applications: a consumer-facing, and a provider-facing one. AI tools are integrated in both.
Our COVID-related product features
Now that we have some context, let’s focus on our approach to COVID-19. Our current approach includes several AI powered features such as automated Question Answering, Diagnosis and Assessment, and Follow Ups. We also have integrated Care Plan recommendations that include prescriptions and lab testing.
Let’s start with our automated Diagnostic Assessment, which is based on the latest known recommendations and research. On the surface it may look like other COVID-19 symptom checkers, but there is much more to it. We have a dynamic questionnaire that adapts to responses in the flow (see 3rd screenshot where the patient gets follow-up questions related to their reported symptoms). Assessment is also based on personalized features including for example city of residence and incidence of COVID-19 in the area.
Our automated assessments provide real-time recommendations to our patients at the onset of their symptoms and also through daily check-ins that ensure that we monitor their symptoms and enable efficient constant communication. Throughout their disease course, we always offer conversations with our medical providers to address issues or questions that the AI cannot.
As mentioned before, we have added capabilities for integrating at-home COVID testing that enables end-to-end care without the patient leaving the house except in critical cases.
Through our partnership with Quest, our users have access to testing capabilities in the service. This includes COVID-19 antibodies testing, which has been fully integrated in our COVID-19 offering.
Covid-related research
Since we started Curai, we have strongly focused on research related to medical diagnosis, medical NLP, and image processing. You can see most of our recent publications, many of which have been done in collaboration with universities such as Stanford and MIT, in Google Scholar.
Some of that research is now being applied to the COVID-19 pandemic. As a first example, our earlier research, published at Neurips ML4H, showed how we can use a doubly fine-tuned BERT language model to identify similar medical questions in “Domain-Relevant Embeddings for Medical Question Similarity”. The pre-trained language model was fine-tuned on a medical Q&A task and then on medical question-question similarity. We now apply that with some slight modifications to the COVID-19 scenario. We are implementing an automated question answering feature with the goal of providing a much better user experience to people looking for COVID-19 information. Instead of navigating through a long list of questions, the user can input their personal question and be directed to the most relevant ones. To avoid misinformation, our answers are curated, and because a simple answer might not solve the concern, we again offer the opportunity to connect directly with a provider. To try out the model, you can access our COVID-19 QA section in our website.
Finally, we are also very happy to share that we have trained what as far as we know might be the first machine-learned COVID-aware diagnosis model. We have extended the approach described in our “Learning from the experts” paper, which is illustrated in the figure below and discussed in a previous blog post. The basic idea is to use an expert system as a prior and simulator to generate clinical cases. Simulated cases can then be combined with any other data such as, for example, data from EHR.
We have now added data coming from our COVID assessments. This is a great example of the ever-learning feedback loop we mentioned earlier. We now have a fully COVID-aware DX model. It is hard to evaluate the accuracy of the model because of the lack of ground truth data at this point, but we hope to continue improving it as more data is available.
Conclusions
Healthcare needs to scale quickly, as we have unfortunately witnessed quite recently. We strongly believe that effective scaling of healthcare access in order to not only preserves but improves quality of care, is by leveraging technology and AI. However, AI can not be merely dropped into old ways of doing things, however. We need to integrate it in an end-to-end system benefiting both patients and providers. At Curai we are focusing on the long-term vision, but in doing so, we have had the opportunity to be useful in the short-term by quickly applying our approaches to an unforeseen situation. While we are now only operating in California, we hope to scale our solution to the whole world in the future and bring the world’s best healthcare to everyone.
Note: This post is an updated and longer version of a talk we gave at the AI for COVID-19 Conference at Stanford a few weeks back. The short video below can be a good visual summary of what is explained in the post.
