I met Dr. Pooja Rao last year through a very revealing email exchange. I sent Pooja a note to share some recent Qure.ai coverage and invite her to subscribe and she responded with a series of questions about the tools we use to automate this type of outreach. It was at that moment that I realized Dr. Rao is uniquely solutions oriented.
As Qure.ai’s co-founder and Head of R&D, Pooja is usually focused on solving far more important issues than email automation, using her background in medicine, data science, and neuroscience to make healthcare more accessible and affordable through deep learning.
In this first-ever Imaging Wire Q&A, we sat down with Pooja to discuss the current challenges in stroke and head trauma treatment and how AI solutions, such as Qure.ai’s qER product, stand to improve clinical outcomes. Here it is:
What drew Qure.ai to stroke and head trauma AI?
Pooja Rao: Stroke is one of the leading causes of death and long-term disability worldwide. Patient outcomes depend strongly on how quickly stroke is diagnosed and treated, measured as ‘symptom onset-to-needle’ time.
Most patients with a stroke go through an accelerated stroke protocol that includes rapid imaging and review, but there are many others with brain bleeds (stroke-related or otherwise) who are outside of this protocol. For example, a patient who’s already in the hospital for an ischemic stroke, but gets an acute bleed during treatment. That’s where you need AI that works in the background to pick up these scans and prioritize the right patients.
Over 2.5 million people suffer head injuries in the U.S. every year. A fraction of those will require urgent neurosurgical intervention – and imaging is key to making that decision. The use of CT scans in the emergency room has been on the rise for decades, which means that radiologists in turn have long lists of ‘STAT’ scans to review. If AI could scan through these and push the critical ones to the top of the list it would save a lot of valuable time for these patients.
What are the current stroke and trauma guidelines and how does AI fit in?
Pooja Rao: The 2018 American Stroke Association/American Heart Association (AHA/ASA) stroke guidelines say that non-contrast CT provides the information needed to make decisions about acute stroke management in most cases. They also say that the primary role of a head CT scan for patients with stroke symptoms is to rule out a bleed, and that there is no evidence for making treatment decisions based on the subtle CT signs of ischemia.
Further, they advocate for using non-contrast CTs to screen patients because it’s cost-effective. This means that radiologists’ head CT volume continues to grow. High-volume practices can have as many as 20 head CTs an hour in addition to all the other studies they read. Simply flagging critical scans would add a lot of value here.
Stroke centers are also required to score intracranial bleeds by volume. This is another area that AI can save time for radiologists, by marking out brain hemorrhage and measuring its volume.
What has research revealed about the performance of AI solutions for stroke and head trauma?
Pooja Rao: Standalone studies show that the technology works well and is safe and effective enough to be used in clinical practice, and it sounds like regulatory bodies agree, given the recent clearance of AI products to triage critical scans and assist radiologists.
Our own study, published last year in The Lancet showed that qER accurately detects not only bleeds but also other critical head CT scan abnormalities like mass effect (sometimes the only early sign of a tumor), midline shift, and cranial fractures.
What about in clinical use?
Pooja Rao: As we deploy at more hospitals and imaging centers, we’re generating evidence that AI works just as well in the clinical setting as it does in the lab. In addition to proving that the technology generalizes well (performs with high accuracy independent of the CT scanner model or population), we’re also quantifying the clinical benefit to patients, radiologists, and other physicians. When we evaluate the benefits of AI for stroke and head trauma we look at:
- How much time is saved when critical scans are prioritized by AI?
- How does this prioritization impact other studies on the worklist?
- How are patient outcomes impacted?
Where is head trauma and stroke AI being adopted first and who’s finding it most beneficial?
Pooja Rao: There is a lot of AI research coming out of academic centers, where quality of care is the highest and there’s an abundance of the best and brightest doctors. But care and radiology standards aren’t uniform across the world, or even within the U.S.
We’re seeing that the earliest serious AI adopters are community hospitals and remotely located healthcare providers where there may not be reliable, accurate 24×7 radiologist coverage. It also seems that geographies with a shortage of expert care are taking the lead in adopting AI, reflecting where value is truly being added.
Of course, there is still a long way to go and there are a lot of questions that need answers. Is the role of AI to prevent tired doctors from missing critical findings, help save time dictating reports, or to prioritize critical scans on busy worklists? Is it all three?
And how are these solutions benefiting patients and radiologists?
Pooja Rao: For patients, a lot of the benefit of AI is access – just having access to rapid, accurate diagnosis and treatment, and not having to wait hours in the ER.
For radiologists, the benefits of AI differ based on the setting in which they operate. Busy urban practices or teleradiology setups benefit the most from having critical cases automatically flagged for review. Many radiologists also like having bleeds and midline shifts quantified because it saves them time. In places where radiologist coverage is sparse, radiologists and other clinicians find the mobile phone alerts with non-diagnostic preview images particularly useful.
These are exactly the patient and radiologist benefits we targeted with qER.
What’s the next frontier for head trauma and stroke AI?
Pooja Rao: Everyone wants algorithms that can be superhuman and see abnormalities that radiologists can’t, but there are easier problems to solve first.
One of these is incorporating clinical knowledge. In studies that we’ve done, we’ve observed that radiologists are at their most accurate when provided the full clinical context. We’re now training AI to incorporate that clinical context.
Another one is predicting long-term outcomes. qER already measures the volume of the abnormalities it detects to help study progression in patients with traumatic brain injury. We’re now going beyond quantification and progression monitoring to using these measures to predict patient outcomes.
Thank you, Pooja. It’s exciting to watch Qure.ai work with global healthcare providers to address serious conditions like stroke and tuberculosis and we can’t wait to see what’s next.