November 2019, Ai4 Health Care held its second annual conference in New York City. Speakers and attendees shared their experiences, expertise, and perspectives on AI – Artificial intelligence. Excitement over AI was widely shared. It was often pointed out that with countless opportunities for AI in healthcare, the need is real and the market is huge; AI is required to analyze the huge amount of patient data being generated. Although AI already surrounds us, the mood at the conference was that we were all early witnesses at the start of a new era in problem solving.
Of the many topics discussed during the conference, I kept wondering how the ideas presented will eventually lead to new products and processes. In attendance were clinicians, practitioners of medicine and potential end users of new AI based products; data scientists and other persons and companies on the supply side of AI; and corporations who currently supply services and products to patients and the medical community. For successful new product development, ideally an interdisciplinary community can correlate challenges and emerging technologies, and form clearly defined problem statements based on clearly defined needs. From problem statements come design inputs which eventually results in new innovative products. when product development is removed from day to day experience, the results may be a ‘solution in need of a problem.’ Have you ever used a ‘time saving’ product that waste more time than it saves? It’s predicted that amidst all the excitement and allure of AI, there will be large amounts of waste on projects that will fail. Sounds pessimistic, but some failure is inevitable. The advice is to think projects through up front.
At the conference, there was a definite apprehension between product developers and clinicians, data scientists and doctors, and this apprehension was discussed in several talks and panel discussions. The first bit of apprehension concerns data privacy. The topic readily surfaces in almost every discussion. Each new story introduces a new valid and concerning perspective. Data ethics includes the idea that personal data is used responsibly, as opposed to ‘not used,’ because there’s a benefit to the larger community when health data is shared. There was also apprehension concerning product development, outside of the data privacy issue. Anecdotes were shared where developers created product designs without any input from the medical community. Several clinicians spoke of the unrealized promise of EHRs, electronic health records; to have readily available patient data is a great idea, but it takes hardware, software, and time for data entry. In service of EHRs, doctors and nurses now have a tablet and data entry tasks in between them and the patient.
One exciting and active area of productization was image analysis. Several presentations, made by data scientists equipped with a working knowledge of a particular disease, were focused on improving the efficiency of diagnosis using image analysis. This was in follow up from last year’s discussions concerning the enormous amounts of imaginary data, now easily acquired and stored, potentially overwhelming clinicians. The application of AI and data science to image analysis is a good match; in addition to assisting a clinician tasked with reviewing images, early disease detection will improve patient outcomes and lower healthcare costs.
Interpreting what was presented about image analysis, and evaluating it as a potential new product, the problem statement needs to clarify if the new product aims to improve the doctor’s ability to treat the disease or concentrate only on assisting with the diagnosis. Although this seems subtle, it drastically alters the product’s regulatory strategy and will impact numerous other design decisions. Ambiguity complicates communication between product developers and users. Another obstacle is terminology. Despite the fact that quite a few persons at the conference were well versed in both clinical and data science terminology, it appears that even foundational terms such as AI and data science required presenters to clarify what they meant when using these terms. Machine Learning, Deep Learning, AI, and Data Science are all caught up in the allure and hype of AI. Several data scientists made effort to differentiate data science from data wrangling. Within the technical community, some struggle to understand when ‘data’ became ‘the’ science. Without claiming the authority to provide a definitive definition, I assume ‘big data’ made ‘data’ a stand alone field.
Continuing to explore the above mentioned hypothetical image analysis project, it will incorporate machine learning. It will use annotated data to train an AI which is called supervised learning. In this example, the data set will be a specific set images. From this data set, the model acquires expertise. When a new image is fed into the model, an image outside of the original data set that is not annotated, the image is evaluated using descriptors that fit the original data to determine if it contains evidence of the disease. My perspective is that the AI flags new images, for further review by a clinician, based on how the new image compared to the known images in the data set. The AI is performing a similar task as that of an assistant helping take a first pass at the diagnostic images gathered during a patient visit. The emphasis is on assisting the physician, not replacing the physician or suggesting treatment. Since the original data set, used to train the model or AI, is defined, annotated, and accessible, it can be accessed and expanded by adding new annotated images which will be leverage to improve the AI’s ability to flag new images.
The importance of the original, annotated data set quickly comes into focus. The AI is only as good as the data set, so care is required to assemble and annotate this original ‘training’ data set. But, will every competing product have its own data set? Or would it be beneficial to public health if a standardized set of images existed? It’s probably unlikely to have complete industry collaboration such to create a single national or global data set for all facilities to train their models, with a second standardize set to test each model.
Ai4 Healthcare 2019 brought many interesting parties together to discuss the possibilities for AI in Healthcare, and I will add that Ai4 does a great job concerning conference logistics. My concluding thoughts were that image analysis via an AI is inevitable, but I believe it will originate inside a healthcare facilities where physicians, IT staff, data scientists, programmers, as well as administrative personnel all work together to clearly identify the right opportunities where existing processes are inefficient or a particular problem is persistent. This is opposed to a standalone AI provided developing a ‘black box’ product. Perhaps after being prototyped onsite, productization will follow.
Glenn DiCostanzo, March 2020
Founder & Principal, GD Consulting llc