Navigating the Terrain of Healthcare AI
Artificial intelligence is a rapidly evolving tool in the healthcare space. AI has the potential to positively impact our healthcare experience, ranging from functioning as a safety net providing a second read on imaging scans, to having an immediate impact at the bedside through clinical decision support tools. As healthcare AI evolves, so do the questions and the need to learn from real-world medical experience.
To help you navigate the healthcare AI terrain, we’re working with Dr. Harpreet Dhatt, a Diagnostic Radiologist at Dignity Health. Together with Dr. Dhatt, we’ll explore AI through the eyes of a physician providing direct patient care. He’ll share his experience and learnings in the blog series Navigating the Terrain of Healthcare AI.
The series will cover:
- An introduction to the field of radiology
- A look at the benefits of AI from the physician perspective
- Fundamentals of machine learning in medical imagery
- Machine learning and its role in healthcare
- AI as a tool for preventing missed findings and minimizing malpractice
- Moving from peer review to peer learning and how AI can help
- Will AI replace the radiologist
- Where does AI fit into current workflows
Enjoy the first post in the series, Radiology and Artificial Intelligence Alliance.
Inadvertently wandering into the basement of Stanford Fleischman Labs as a third-year medical student fundamentally altered the trajectory of my career. Peeking into an open room, I saw a large radiology monitor with a rotating 3-D reconstructed image of a full-human body – I was mesmerized and further intrigued as to how one could use this image to diagnose diseases. It was the beginning of my love for medical imaging – a noninvasive way of diagnosing human diseases ranging from infections to cancers and from traumatic internal injuries to fetal anomalies. It’s a broad discipline requiring rigorous training post-medical school to fully understand and accurately diagnose a broad spectrum of diseases utilizing ultrasound, x-rays, computed tomography, magnetic resonance imaging, and nuclear medicine.
Radiologists are doctors supporting patient-facing physicians who can be in the emergency room or the outpatient setting. Our clinical colleagues depend on radiologists to make the initial diagnosis and provide guidance for the next steps along the patient care path. It’s our responsibility as radiologists to diagnose breast cancer, brain bleeds, fractures, or fetal anomalies. The tremendous responsibility and stress of radiology is balanced with the satisfaction of being the first to make life-saving diagnoses. But more than just image interpreters, we are a fraternity of progressive physicians who are often the first to adopt technological advances such as artificial intelligence.
From German Physicist Wilhelm Roentgen’s discovery of x-rays to the advent of CT by Godfrey Hounsfield and MRI by Paul Lauterbur, radiologists have incorporated these amazing technologies to significantly advance the field and patient care. More recently, the adoption of PACS (Picture Archiving and Communicating System) in the early 1990s revolutionized our field – moving from hard copies of medical images to electronic storage and distribution. During this time, computing power has made tremendous progress along with mass access to the internet, radiology images via the web have become increasingly available to physicians and non-physicians. Teleradiology firms are increasingly used along with traditional hospital-based practices to manage the explosive growth in the use of radiology as a diagnostic tool.
In 2019 about 135 million radiology studies were performed, with a price tag of approximately $130 billion. The expected growth is 4% per year over the next 7 years. These staggering numbers have challenged our field, resulting in an expected shortage of radiologists over the next decade. In the near term, radiologists are overworked. Cross-sectional studies have become more complex, and the ever-increasing number of images is difficult to manage. As an example, the CT angiogram of the chest, abdomen, and pelvis of an ER patient presenting with back pain may have more than 3,000 scan images requiring immediate interpretation.
With increased pressures to analyze large volumes of cases, missed diagnoses are inevitable and compounded by the sheer increase in volume. The legal and monetary ramifications of missed malignant lesions and diagnoses on radiologic exams can be devastating to patients, individual doctors, and health systems. One of the most common radiology malpractice complaints is a missed cancer diagnosis on a study performed for a completely separate indication.
Let’s work through the details. The primary care doctor orders a CT of the abdomen and pelvis to evaluate subacute diffuse abdominal pain on a middle-aged patient. Image acquisition of CT for this study includes small portions of the lung bases. The radiologist interprets the exam as negative for acute findings. However, there is a 10 mm nodule at the lung base only visualized on coronal reconstruction and not commented on in the report. The patient returns 5 years later complaining of unexplained weight loss and upon repeat scan, the lung nodule is now 2 cm and turns out to be malignant.
The stress of a missed lung nodule which is potentially malignant percolates in the mind of nearly every radiologist, creating tremendous stress. Today we have the opportunity to become both gatekeepers and first-line adopters of machine learning, a technology that could potentially reduce that stress, further advance our field and positively impact the quality of patient care.
So why are radiologists hesitant to use machine learning? At a first and superficial glance, the advantages of machine learning (or more colloquially, Artificial Intelligence, stated in laymen media) are obvious. Accurate diagnosis at a faster rate, unburdening the radiologists, and potentially reducing the medical legal liability is a no-brainer. Shouldn’t radiologists adopt this technology with the proverbial “open arms”? As always, the reality is murkier. It turns out, both AI technology and malpractice laws aren’t equipped for prime time – just yet.
Looking at the process of AI, convolutional neural networks are a revelation. It’s amazing to think that millions of nodes are working together to learn from image and text data sets, as a path to build AI algorithms. However, what we find is that these algorithms are often narrow and task-specific. Each algorithm is generally focused on achieving only one output. As an example, is there pneumonia on a chest x-ray or is there a lung nodule on the CT scan? Interpretation of a seemingly simple study by a radiologist, such as a chest x-ray, is far more complex and nuanced, hence the requisite training and education.
Taking a broader look at the role of radiology, we find the narrow task-specific AI of today can serve radiologists in a very tangible way.
How you ask?
As a double-check, safety net mechanism running behind the scenes, AI can ensure that no serious and obvious finding goes unnoticed. In such a setting, a product that seamlessly integrates into the radiologist’s workflow will be the most useful. This form of machine learning is in opposition to the narrative headlined by some – the development of front-end technologies making the initial diagnosis, auto-generating reports, and replacing radiologists.
The concern over job loss and replacing radiologists is palpable, not only at the hospital water cooler but also in major radiology online community forums. Replacing radiologists with AI is not an option because radiologists aren’t just tasked with making a singular diagnosis on images, we are responsible for the entirety of the data set. AI is a perfect tool to augment and support the radiologist, not replace them. Let’s use the example of a chest x-ray. An AI algorithm interprets a chest x-ray, identifies a large pneumonia, and sends the report automatically to the referring physician. But what if, in addition to pneumonia, there was subtle “free air under the diaphragm” suggestive of bowel perforation? The AI algorithm was trained specifically to look for pneumonia and misses the subtle finding. which is likely more important in this case than the pneumonia as the mortality from bowel perforation is higher.
The questions asked in this situation are… Who would be legally responsible for such a missed finding? Would it be the enterprise that chose that algorithm? The hospital leadership who implemented the workflow? The AI company that developed the algorithm?
It’s a very complex situation without a simple answer, but one that supports not just the need for radiologists but also the use of AI as a safety net, second look tool.
We will work through these questions in the next posts.
As always, we welcome your thoughts on the use of AI in the radiology space. Feel free to drop a comment or question in the discussion area below.