Some artificial intelligence experts have been worrying publicly that recent high-profile AI programs are starting to get too good at passing as human – which could have the effect of eroding trust in the technology.

Peter Shen, the head of the digital and automation business at Siemens Healthineers, worries about the risk a loss of trust could pose for AI in healthcare in particular – an industry that has been slow to adopt the technology despite the potential to improve patient care, among other things.

We interviewed Shen to get a deep dive into the world of healthcare AI and discuss what it will take to ensure trust in and wider adoption of the technology.

Q. How might a program like ChatGPT change healthcare in the future?

A. Since the artificial intelligence company OpenAI made ChatGPT – a free, interactive chatbot powered by machine learning – available to the public, this interactive tool has generated headlines. It can answer questions, compose emails and generate long-form written content (including stories, news articles and student essays) with startling speed, as well as a level of quality that suggests a human author.

The implications for this technology, which is considered a paradigm-shifter in AI’s evolution, are wide-ranging. They include potential applications in marketing and journalism, as well as healthcare. Concerns have also arisen regarding potential abuse of ChatGPT by students, as well as its reliability with respect to some of the sources it cites.

Some critics of ChatGPT’s potential use in healthcare assume that clinicians would use such a tool to drive a patient diagnosis or critical clinical solution. However, I see a potential benefit for a solution such as this in terms of its ability to consume large amounts of textual data and summarize the results.

That capability is extremely relevant in healthcare, where clinicians are challenged with absorbing incredibly large amounts of emergent digital data in the form of publications and scientific studies regarding their specialty. Armed with information about new, potentially relevant clinical results, the clinician could change diagnosis and treatment paradigms for patients.

ChatGPT represents an opportunity to consume and summarize this publicized, scientifically validated data, helping the physician to remain current. For maximum effectiveness, ChatGPT would need to be fed the most current publications on an ongoing basis. This practical application of the tool would benefit the clinician and healthcare in general while avoiding the landmine of AI making definitive clinical decisions, as trust on that front remains elusive.

Q. Healthcare is an industry that has been slow to adopt AI despite the potential to improve patient care. Why do you think this is?

A. In its early days, AI was the subject of considerable speculation, and its practical value had not yet been established. One thing that seemed certain, if you believed early headlines and professional polls: It would replace radiologists and other clinicians. Perceiving a new technology as a job threat is always counterproductive to adoption.

In recent years, however, those replacement fears have eased, and the healthcare community has a clearer sense of AI’s core benefits. In radiology, the technology has proven capable not only of saving time by automating repetitive tasks, but also of identifying otherwise-overlooked areas of concern by leveraging pattern recognition.

Additionally, AI has begun to provide qualitative visualizations and guidance associated with suspected malignancies.

But new questions have arisen. If AI introduces additional data points that may not even be relevant to a diagnosis, then how much time does it truly save a clinician? And how can this new data be contextualized and incorporated into a reporting style? Also, how do we address the issue of implicit bias, where algorithms are trained on batches of data that fail to include gender, racial and geographical differences?

These questions underscore how much AI still must evolve to provide broader, more substantive value to not only the clinician, but also to the patient and the healthcare institution. Until those questions are answered, some healthcare entities will struggle with cost justification and be ambivalent about wider AI adoption.

Q. How can the healthcare industry increase AI adoption?

A. For its adoption to increase, AI must distinguish itself in the eyes of radiologists and other clinicians as being able to bring something truly new to the table. After all, these professionals already know how to make a diagnosis or treatment decision. They have been doing it throughout their careers.

They need to be able to determine whether AI’s additional information is helpful, relevant and worthy of consideration. They need to know how it changes their diagnosis – if at all.

Developing AI models that include the rationale for their findings will make the tool more useful to radiologists and other clinicians, leading them to be more vocal champions of AI. Those clinicians will also feel more confident about using AI when its implicit bias has been removed through next-generation algorithms that are continuously fed data that is representative of diverse patient populations.

Even more important to AI adoption is the effective, seamless integration of AI’s additional data into the routine clinical workflow. That additional information provided by AI should be just one more easily accessible tool that complements the clinician’s established routine; it should never be intrusive with respect to that routine.

But perhaps the larger, more overarching challenge with respect to AI involves changing our collective mindset about what is and is not AI’s role in healthcare. The prevailing perception in some circles is that AI’s implementation will lead to the tool, rather than the clinician, making decisions.

Even standalone AI solutions from Siemens Healthineers are companion technologies designed to assist the clinician, who makes the ultimate determination concerning patient care. Fully recognizing that AI does not – and should not – bear the burden of making clinical decisions is key to broader acceptance.

Q. Do you have a vision for the next generation of health AI? What changes might we see in the future?

A. Currently we use AI to, for example, spot a potential abnormality on a chest CT image. Taking AI to the next level involves using a multi-data middleware platform to combine that sort of imaging information with other forms of heretofore siloed healthcare data – lab diagnostics, pathology results, genomic information – and overlaying AI across those silos to find correlations.

This use of AI will help drive more informed diagnoses and more personalized treatment decisions.

A hypothetical example: A urologist, based on professional experience, may prescribe 10 weeks of radiation therapy, three times a week, to treat a prostate cancer patient. But if that urologist could examine all available forms of data from that patient – imaging, laboratory, pathology and genomic – and overlay AI to find correlations in that data, the result might be a suggested personalized treatment plan with a scaled-back regimen.

It might consist of only five weeks of radiation delivered just once or twice a week.

This sort of AI-assisted personalized treatment planning has tremendous implications for patient care. Institutions could apply it to an entire cohort of patients who have similar characteristics to achieve greater success. Using AI in this manner represents true population health management, which is a goal of Siemens Healthineers.

Q. How can AI provide broader value to the healthcare system at large?

A. If AI can facilitate not only a precise diagnosis and treatment decision for the individual patient, but also scale up that personalized medicine approach to affect entire patient cohorts, it will prove value to a healthcare system beyond a specific discipline or specialty.

A key component of that approach is an integrated data management layer that can pull together disparate forms of information and bring them into one platform to inform planning and prescription. Already, some progressive healthcare institutions are moving in that direction.

In a related vein, AI may one day prove its broader value by creating highly accurate models of a patient’s anatomical structure. These models could demonstrate, noninvasively, how that anatomy reacts to different forms of treatment.

Ultimately, that “digital twin” of the patient would assist the clinician noninvasively in determining a personalized optimal treatment. More broadly, it would also enable institutions to place the patient in a wellness-focused environment and help determine ways to keep that person healthy. AI’s ability to provide that benefit could radically transform healthcare.

Follow Bill's HIT coverage on LinkedIn: Bill Siwicki
Email the writer: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.