Artificial Intelligence in Healthcare: Lost In Translation?

Abstract

Artificial intelligence (AI) in healthcare is a potentially revolutionary tool to achieve improved healthcare outcomes while reducing overall health costs. While many exploratory results hit the headlines in recent years there are only few certified and even fewer clinically validated products available in the clinical setting. This is a clear indication of failing translation due to shortcomings of the current approach to AI in healthcare. In this work, we highlight the major areas, where we observe current challenges for translation in AI in healthcare, namely precision medicine, reproducible science, data issues and algorithms, causality, and product development. For each field, we outline possible solutions for these challenges. Our work will lead to improved translation of AI in healthcare products into the clinical setting. BackgroundHealthcare systems worldwide suffer from ageing populations and exploding medical costs [1]. On the current trajectory, they will soon become economically unsustainable. A possible solution is the application of artificial intelligence (AI) technology, which has the potential to enable improved healthcare outcomes and to reduce overall costs [2].This can be attributed to the main properties of AI technology. Currently, when we refer to AI, we usually mean techniques of machine learning (ML). ML excels at pattern recognition, especially in big data, and can process a multitude of different input types, ranging from tabular clinical data to medical imaging. These techniques have led to unprecedented advances in areas ranging from computer vision to natural language processing.But where does AI in healthcare stand today? Headlines suggest that AI has already surpassed the performance of human doctors in many medical fields [3,4]. There is no panel on AI in healthcare without the question of when AI will replace human doctors arising. These perspectives stand in stark contrast to clinical reality. Only a few AI tools are available in the clinical setting [5,6] and we are not aware of a single AI in healthcare tool appearing in clinical guidelines as a standard of care.How can this discrepancy be explained? The results gaining media attention are mostly so-called exploratory studies trying to establish proof-of-concepts (PoC) that do not reflect the full process necessary to bring these advancements to clinical reality. Exploratory studies are necessary to identify potentially valuable use cases for AI in healthcare [7]. Here, many scenarios need to be tested and successful PoCs can justify -both on an economic as well as an ethical level -the allocation of funds for follow-up confirmatory studies [7]. These confirmatory studies are needed to safely translate the exploratory results to the clinic and need to adhere to all quality markers of studies establishing efficacy, e.g. adequate power and predefined protocols [7]. Given the costs of confirmatory studies, they will only be performed as part of product development.

AI revolution in healthcare and medicine and the (re-)emergence of inequalities and disadvantages for ageing population

“…In fact, there are only few certified and even fewer clinically validated products available in the clinical setting. Most of the hype around medical uses of AI is related to cases of technology in the exploratory stages of development (proof of concept), which identify potentially valuable use cases, but which are yet to be validated in the clinical use trials (Madai and Higgins, 2021 ). Hence, many experts suggest caution in estimating the real effects of this technology on the future of healthcare for older adults (Berisha and Liss, 2022 ; WHO, 2022 ).…”

The Promise of Explainable AI in Digital Health for Precision Medicine: A Systematic Review

“…Artificial intelligence (AI) is integral to offering solutions to various challenges in healthcare, including the standardization of digital health applications and ethical concerns related to patient data use [ 27 , 28 ]. Precision medicine, a common application of AI in digital health, involves tailoring healthcare interventions to subgroups of patients by using prediction models trained on patient characteristics and contextual factors [ 29 , 30 , 31 ]. However, the reliance on AI in healthcare raises issues regarding transparency and accountability with black-box AI systems whose decision-making processes are opaque [ 32 , 33 ].…”

Demystifying Supervised Learning in Healthcare 4.0: A New Reality of Transforming Diagnostic Medicine

“…The challenges towards healthcare transformation in using SML are data issues, data-snake oil, interdisciplinary team building, reproducibility, personalized medicine, moving into clinical practice, data and algorithms, causal AI, product development, and effectiveness and trust in AI-augmented healthcare [ 107 ].…”

“…Some challenges also translate theoretical fundamentals to practical ones as experience is scarce and the SML method is still in its infancy. Integration of and overlaps of theoretical fundamentals, data science, proper statistical inference, approximation, and attribution, are very active and open research up to questions regarding this type of healthcare research [ 107 ].…”