Healthcare and AI: What is needed for them to better interact with each other?
Currently an Associate Professor at Aalborg’s University Human-Center Computing group, Niels van Berkel is a remarkable leader in the research area focusing on human-AI interaction. His various experiences as a student, researcher or professor at universities in the Netherlands and Australia, Microsoft Research center in the USA and Interaction Centre at University College London, exposed him to numerous concepts and ways of thinking that sparked his interest in improving end-user interaction with technology.
One of his most recent publications in ACM Transactions on Computer-Human Interaction, “Measurements, Algorithms, and Presentations of Reality: Framing Interactions with AI-Enabled Decision Support”, is of utmost importance for researchers and practitioners in the Human-Computer Interaction (HCI) community as it describes challenges that were generated as AI technologies were brought into clinical practice. For that, Niels and his colleagues create a MAP (Measurement, Algorithm, Presentation) model through which they contrast the interpretation of situations in the real world as typically carried out by medical personnel with the required steps for an AI system to reach an effective interpretation and present it to clinicians in actional form. This model consists of three steps taken by such a support system to provide care to its end users: Measurement, in which a real-world phenomenon is described and recorded by means of one or more sets of data, Algorithm, through which the collected data is transformed into a useful outcome signal, and Presentation, in which the outcome is presented to the user. The case studies on colonoscopy practice and dementia diagnosis put all these ideas in practice, reinforcing the significant need for representatives from HCI to get involved more deeply in the design of AI systems.
The first element of their model — Measurement — focuses on data and any inference that needs data. For a healthcare professional, gathering data is synonymous with the observation of the patient, with computational systems requiring the conversion of this information into digital data. A real-world application of this concept is represented by heart rate monitors: a measurement of the heart’s electrical activity through ECGs is converted into a digital signal that a trained operator can interpret. This human factor that comes into play when performing the interpretation is the one that causes most challenges when trying to build AI-enabled systems. The selection of appropriate computational representations to identify what is really happening with the patient, integrating pieces of data that come from multiple sources and understanding the completeness or validity of data are considered by Niels and his colleagues as the major challenges in building AI medical systems that value Measurement.
The Algorithm comes next in the analysis of the model they proposed, with it referring to the computations that need to be conducted in day-to-day healthcare in order to consider both in-the-moment and long-term outcomes. The best example related to this element is the process of assessing a patient’s pulse — calculate heartbeat per minute based on observation and reason if the heart rate is normal or not based on the patient’s age and sex. It is easy to identify that reliance on historical data and categorization algorithms results in several concerns when thinking about the transition to AI-based applications. According to Niels et al., these concerns may vary from ensuring that training data reflects the true circumstances of the problem in question to building algorithms that promote replicability as well as the ability to modify computations to optimize results.
The last piece of their model is Presentation, which deals with the capacity of a system to convey information to the user as efficiently as possible. Even though the Human Computer Interaction literature explored the power of Presentation deeply in the last few years, it hasn’t touched too much upon its importance in the healthcare industry powered by AI. With the medical environment being one in which everything happens very fast and unexpectedly, building computations systems for it may cause various challenges to arise: ensuring the interpretability of results, supporting collaboration between healthcare professionals at different levels and building with the remarkable need for timeliness in mind.
Now that we have the MAP model promoted by Niels conceptualized, we can take a look at one of the case studies that applies it in a real world context. It focuses on colonoscopy practice, being part of a project aiming to support endoscopists in identifying polyps in the colon during an inspection that uses an AI enabled application.
As far as Measurement is applied in this study, it was observed that the system can only identify polyps when they are present in the view of the endoscope’s camera. Highlighting potential polyps in the endoscope’s view provides only a partial solution for the problem of polyp miss rates. AI-enabled solutions that aim solely to identify polyps that are already visible and at the surface of the colon may, therefore, not provide a complete measure of real-world phenomena (hidden polyps). Instead, representing the real-world challenges would require an AI-enabled system to also provide support in guiding clinicians in ensuring complete coverage of the colon.
Analyzing the Algorithm, the clinical collaborators involved in the study raised concerns about the possibility and consequences of false positives once AI-enabled decision support systems are integrated into clinical practices. Currently available systems frequently show false positives when presented with non-polyp material present in the colon, such as bubbles, stool, or wrinkles on the surface of the colon. Given the frequent occurrence of these features, the system can become highly distracting to the operator — potentially resulting in the clinician ignoring the system’s correct recommendations or turning the system off entirely. These false positives result from incomplete training datasets, which render the AI system unable to distinguish between anomalies present in the colon and actual polyps.
The Presentation factor in the case of colonoscopy practice was mostly important in two directions. The first one discusses the fact that colonoscopy procedures are team efforts that involve multiple medical professionals. For example, staff nurses and healthcare assistants are responsible for delivering medical instruments through the endoscope when requested by the acting endoscopist, as well as maintaining patient sedation or physically moving the patient [99]. A critical assessment on supporting collaboration between team members when using an AI-support system in the clinic is currently lacking. The second area that is covered highlights that the assessment of individual video frames at near real time speed requires substantial computational resources. Delays in the presentation of the algorithmic assessment are detrimental to the usefulness of the AI system, as clinicians require direct feedback when manipulating the endoscope.
So, the entire paper written by Niels and his colleagues identifies and explores the importance of introducing AI-enables systems in healthcare with a lot of consideration, since the field is really complex and has the most precious thing at stake — human life. The three stage procedure they are promoting through the MAP model formalizes the interpretation of medical reality through the lenses offered by AI, pointing out numerous and diverse concepts that need to be taken in account when building and implementing such systems.
