Artificial intelligence applied to medicine: There is an “elephant in the room”

Abstract

“…..AI algorithms used for diagnosis and prognosis must be explainable and must not rely on a black box…..”. This strong and courageous sentence recently captured our attention: it was stated by S Kundu, a highly reputed young scientist, in her short editorial/letter recently published on Nature Medicine [[1]Kundu S. AI in medicine must be explainable.Nat Med. 2021; 27: 1328Crossref PubMed Scopus (26) Google Scholar]. Dr Kundu has a very intriguing profile being both a physician and a computer/engineering scientist: despite her young age, she received several recognitions being named as “one of Forbes 30 under 30, MIT Technology Review's 35 innovators under 35, a World Economic Forum Global Shaper and a winner of the Carnegie Science Award” as reported in her Wikipedia page [[2]https://en.wikipedia.org/wiki/Shinjini_Kundu.Google Scholar]. Thanks to her rapid career, a broader public increasingly knows her, becoming an ambassador of the need for transparent AI in medicine. She spoke at the recent United Nations AI for Good global summit and inspired one of its priority areas of sustainable development Goals. She is a physician and a researcher at the Department of Radiology of Johns Hopkins University (JHU) in Baltimore, not by chance, mostly working on AI applications in imaging. AI applications to medical images currently represent a paradigmatic example of the “lights” and “shadows” of the introduction of AI in clinical environments. The Dr Kundu’s editorial gave us the opportunity to discuss several aspects that stay at the core of the actual process of AI introduction in medicine, including visions concerning how medical physicists may actively contribute within this complex scenario. Consider a “well trained” AI-based model (model 1) able to estimate the probability that a certain patient affected by a specific type of tumor will not respond to a particular therapy. Suppose that the discriminative power of such a model (incorporating the “hidden” contribution of tens or hundreds of variables) has a declared (after some serious independent validation) area under the curve (AUC) of 0.95. On the other side, for the same clinical situation, an alternative predictive model (model 2), including only a few features/parameters whose meaning is well interpretable or at least understandable, may offer the same prediction with an AUC of around 0.85. Is there someone who doubts what model will be preferentially chosen if the dilemma is proposed to a large group of expert doctors? We may guess that the majority would choose model 2. This simple and hypothetical example well describes clinicians’ absolute preference to the possibility of “understanding” why the model they are using works. Despite this evidence, the image of a medicine more and more driven by AI tools that support (and maybe in the future partly replace) the doctor in making decisions concerning the care of patients is largely promoted in the scientific world and still more in the media and the public. We may argue how many interests and pressures from industry and technological players may act at various levels to push ahead the penetration of AI in the medical field. However, the lack of explainability of black-box AI tools emerged in the recent literature as a relevant issue [3He J. Baxter S.L. Xu J. Xu J. Zhou X. Zhang K. The practical implementation of artificial intelligence technologies in medicine.Nat Med. 2019; 25: 30-36Crossref PubMed Scopus (490) Google Scholar, 4Topol E.J. High-performance medicine: the convergence of human and artificial intelligence.Nat Med. 2019; 25: 44-56Crossref PubMed Scopus (1460) Google Scholar, 5Buch V.H. Ahmed I. Maruthappu M. Artificial intelligence in medicine: current trends and future possibilities.Brit J Gen Pract. 2018; 68: 143-144Crossref PubMed Scopus (109) Google Scholar, 6Stupple A. Singerman D. Celi L.A. The reproducibility crisis in the age of digital medicine. npj Digit.Med. 2019; 2Google Scholar, 7Chen J.H. Asch S.M. Machine learning and prediction in medicine-beyond the peak of inflated expectations.N Engl J Med. 2017; 376: 2507-2509Crossref PubMed Scopus (427) Google Scholar] and is increasingly debated at different levels, including policymakers, ethics experts and philosophers. Notably, the European General Data Protection Regulation (GDPR) [[8]EU General Data Protection Regulation (GDPR): Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation), OJ 2016 L 119/1.Google Scholar] includes a first fundamental step, pointing attention to this issue and including the right for an individual to obtain a meaningful explanation when automated (algorithmic) decision-making is involved. This includes the information on the existence of automated decision-making, meaningful explanation of the logic involved, the significance and envisaged consequences of such processes. In our opinion, the recent emphasis on the road toward “explainable AI” [9Wiens J. Saria S. Sendak M. Ghassemi M. Liu V.X. Doshi-Velez F. et al.Do no harm: a roadmap for responsible machine learning for health care.Nature Med. 2019; 25: 1337-1340Crossref PubMed Scopus (187) Google Scholar, 10Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. Causability and explainability of artificial intelligence in medicine (2019) Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 2019;9(4):art. no. e1312.Google Scholar, 11Tonekaboni S. Joshi S. McCradden M. Goldenberg M.D. What clinicians want: contextualizing explainable machine learning for clinical end use.PMLR. 2019; : 1-21Google Scholar] should be considered as a first attempt to recognize the presence of the “elephant in the room”. On the other hand, we are far from having the solution: the elephant is there, and it is hard to push it out of the room! We propose some considerations, focusing on the position of Medical Physics within Medicine and, likely more, in the traditional domains where medical physicists are mainly asked to deal with AI. Hopefully, this will contribute to better orienteering us in the current initial process of AI implementation. It is out of doubt that several applications of AI may be considered as intrinsically explainable or at least interpretable or usable. Not by chance, they represent areas of likely more rapid implementation of AI-assisted tools that promise to improve quality and efficiency in different domains of medicine. A paradigmatic example is the AI-based automatic segmentation for radiotherapy applications [[12]Unkelbach J. Bortfeld T. Cardenas C.E. Gregoire V. Hager W. Heijmen B. et al.The role of computational methods for automating and improving clinical target volume definition.Radiother Oncol. 2020; 153: 15-25Abstract Full Text Full Text PDF PubMed Scopus (19) Google Scholar] and, more and more, in imaging applications due to the growing use of quantitative information during the diagnostic process. We can also mention the ever-increasing applications of AI in image reconstruction, aimed to obtain images with better contrast and resolution [[13]Wang G, Zhang Y, Ye X, Mou X. Machine learning for tomographic imaging. IOP Publishing, 2020 Bristol, UK.Google Scholar] or the spread of AI-based approaches for radiotherapy plan optimization, shifting an increasing fraction of manual procedures toward automation [14Wang M. Zhang Q. Lam S. Cai J. Yang R. A review on applications of deep learning algorithms in external beam radiotherapy automated treatment planning.Front Oncol. 2020; 19: art 580919Crossref Scopus (33) Google Scholar, 15Pallotta S. Marrazzo L. Calusi S. Castriconi R. Fiorino C. Loi G. et al.Implementation of automatic plan optimization in Italy: status and perspectives.Phys Med. 2021; 92: 86-94Abstract Full Text Full Text PDF PubMed Scopus (1) Google Scholar]. These are only a few and not exhaustive examples of AI implementation, often largely involving medical physicists, that present a high degree of self-explainability/interpretability/usability and that should be more easily accepted by the clinical community. Suppose a new AI-based algorithm is used to generate a CT image: the medical physicist and the radiologist have the concrete possibility to compare the resulting output of AI application against more traditional image reconstruction methods and assess that the new way is better than the older one. Similarly, the resulting segmentation of organs and structures based on previously trained AI-based algorithms can be directly checked (and, if necessary, corrected) by radiation oncologists usually performing this task. In several situations, AI-based tools will clearly translate in the sparing of a large amount of time that doctors or other professionals may dedicate to other, less repetitive, duties: this is the case, for instance, of AI-based segmentation and radiotherapy plan optimization. The recognition that these tools work similarly or better than humans in such applications is expected to be relatively fast, right due to the definite possibility to validate and verify their performances in the clinical environment extensively. Looking at this kind of applications, we may state that we can move the elephant out of the room. This does not mean that the acceptance will be easy and rapid everywhere, primarily due to the associated changes in mentality, professional roles and responsibilities. In this context, “the room” is just our mind and the process to change the mind is never easy, especially if dealing with well-assessed practices. Within this process, the role of medical physicists is of paramount importance, principally as facilitators. Their role in imaging, radiation oncology, nuclear medicine has always been associated with innovation and implementation of new technology or new approaches [16Fiorino C. Jeraj R. Clark C.H. Garibaldi C. Georg D. Muren L. et al.Grand challenges for medical physics in radiation oncology.Radiother Oncol. 2020; 153: 7-14Abstract Full Text Full Text PDF PubMed Scopus (18) Google Scholar, 17Fiorino C. Guckenberger M. Schwarz M. Heide U.A. Heijmen B. Technology‐driven research for radiotherapy innovation.Mol Oncol. 2020; 14: 1500-1513Crossref PubMed Scopus (27) Google Scholar, 18Kevill S. Physics and medicine: an historical perspective.The Lancet. 2011; 379: 1517-1524Abstract Full Text Full Text PDF Scopus (32) Google Scholar]. In line with this tradition, medical physicists will facilitate the implementation of AI through rigorous validations and adaptation to local needs, mainly thanks to their specific skills and ability to understand the clinicians’ language and translate, in a team effort, their requests into optimally tailored solutions [19Bosmans H. Zanca F. Gelaude F. Procurement, commissioning and QA of AI based solutions: An MPE’s perspective on introducing AI in clinical practice.Phys Med. 2021; 83: 257-263Abstract Full Text Full Text PDF PubMed Scopus (7) Google Scholar, 20Beckers R. Kwade Z. Zanca F. The EU medical device regulation: Implications for artificial intelligence-based medical device software in medical physics.Phys Med. 2021; 83: 1-8Abstract Full Text Full Text PDF PubMed Scopus (18) Google Scholar, 21Balagurunathan Y. Mitchell R. El Naqa I. Requirements and reliability of AI in the medical context.Phys Med. 2021; 83: 72-78Abstract Full Text Full Text PDF PubMed Scopus (7) Google Scholar, 22Zanca F. Hernandez-Giron I. Avanzo M. Guidi G. Crijns W. Diaz O. et al.Expanding the medical physicist curricular and professional programme to include Artificial Intelligence.Phys Med. 2021; 83: 174-183Abstract Full Text Full Text PDF PubMed Scopus (15) Google Scholar]. Coming back to the core of the Dr Kundu’s message [[1]Kundu S. AI in medicine must be explainable.Nat Med. 2021; 27: 1328Crossref PubMed Scopus (26) Google Scholar], let us focus on the highly topical field of radiomics. This may be considered a relevant example (among many others) of potential application of AI-based tools dealing with large amount of information extracted from images. Despite the impressively growing number of scientific publications dealing with the topic, resulting in hundreds of AI-based models for diagnosis, prognosis and outcome prediction (mostly in cancer patients), the clinical use of these results is still apparently null. In our view, this is one of the major pieces of evidence of the consequences of ignoring our elephant. Not by chance, in most of the potential applications of AI black-box radiomic scores, doctors should decide if one patient has or does not have a cancer or if they should or should not deliver therapy based on something intrinsically unexplainable. And, as we may expect, this does not happen! The sooner we recognize the issue, the sooner we find solutions. Medical physicists are part of the game; they can contribute to exploring smart ways to exploit high computational and modelling skills without forgetting the indispensable need to drive the picture toward model’s explainability, interpretability and usability. Intending to make AI-based models for medicine more friendly and possibly more explainable, many researchers focus on developing methods and tools that may offer the possibility to capture the inter-dependence among predictors after a “black-box” model is built. These post-hoc actions are aimed to make the black box “a bit less black”. They may give the operators some possibility to explore the logic behind it and, to some extent, to “score” the relative weight of different predictors or of their combinations in assessing the most relevant variables acting on single individuals. In the best situation, these attempts may offer elements of explainability from the point of view of the “technical” building of the algorithm without any possibility to deal with the true issues regarding clinical or biological explanation/interpretation [7Chen J.H. Asch S.M. Machine learning and prediction in medicine-beyond the peak of inflated expectations.N Engl J Med. 2017; 376: 2507-2509Crossref PubMed Scopus (427) Google Scholar, 8EU General Data Protection Regulation (GDPR): Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation), OJ 2016 L 119/1.Google Scholar, 10Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. Causability and explainability of artificial intelligence in medicine (2019) Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 2019;9(4):art. no. e1312.Google Scholar]. This transparency is obviously of paramount importance and any effort in this direction is relevant. However, this approach continues to ignore the elephant, being intrinsically lacking from the point of view of (clinical) explainability. We can find rare exceptions in the literature; when AI tools are forced to remain potentially explainable, resulting in few-feature models [[23]Carrara M. Massari E. Cicchetti A. Giandini T. Avuzzi B. Palorini F. et al.Development of a ready-to-use graphical tool based on artificial neural network classification: application for the prediction of late fecal incontinence after prostate cancer radiation therapy.Int J Radiat Oncol Biol Phys. 2018; 102: 1533-1542Abstract Full Text Full Text PDF PubMed Google Scholar]. The dogmatic credence that AI is “superior” in modeling (large) clinical data is unproven and basically wrong when introducing explainability and interpretability concepts. If “superior” includes the ability of models to be explainable/interpretable (and, as discussed, this should be the case!), we should admit that the “best” models are the ones based on the understanding of basic principles explaining, at least in part, the causes of what we see. The causality, proven or supposed, is a vital characteristic of usable and clinically accepted models. This is clear when few mechanisms may be summarized as those that mainly cause the effect. A quite outdated but still (to some extent) valid example is the linear-quadratic model. We have been applying it for decades in radiotherapy, and it proved robust enough to explain most of the picture in many scenarios when providing estimates of radiation-induced effects on tumors and, in part, normal tissues [[24]McMahon S.J. The linear quadratic model: usage, interpretation and challenges.Phys Med Biol. 2018; 64: 01TR01Crossref PubMed Scopus (121) Google Scholar]. More recent examples include modeling tumor regression as quantified by imaging to provide good estimates of tumor control [[25]Fiorino C. Gumina C. Passoni P. Palmisano A. Broggi S. Cattaneo G.M. et al.A TCP-based early regression index predicts the pathological response in neo-adjuvant radio-chemotherapy of rectal cancer.Radiother Oncol. 2018; 128: 564-568Abstract Full Text Full Text PDF PubMed Scopus (19) Google Scholar] and the quantitative assessing of hypoxia through functional imaging individually predicting tumor radio-resistance [[26]Thorwarth D. Welz S. Mönnich D. Pfannenberg C. Nikolaou K. Reimold M. et al.Prospective evaluation of a tumor control probability model based on dynamic 18F-FMISO PET for head and neck cancer radiotherapy.J Nucl Med. 2019; 60: 1698-1704Crossref PubMed Scopus (24) Google Scholar]. These are well-known examples taken from the radiotherapy field where models based on few, clear and interpretable parameters provide accurate predictions that clinicians use. Why should black-box AI data-driven models convince clinicians to replace them? Models with an intrinsic high degree of explainability align with how science progressed until our days. We may expect that the continuously improving ability to explain central basic mechanisms at the laboratory and the translational research-level would offer more and more alternatives to the black-box approach in many instances in imaging and other areas where medical physicists are deeply involved. And physicists may (and, in our opinion, should!) contribute massively. A major argument pushing to the use of machine learning and AI to model medical data (including medical images, laboratory data, “omics”…) is that, for the first time in the history, humans can access, pool and handle incredibly large amounts of quantitative information. A way to exploit data without forgetting the fundamental issues of explainability and interpretability is the application of machine learning and AI constrained to search the “major players”. We can find examples of such an approach in the TRIPOD guidelines [[27]Collins G.S. Reitsma J.B. Altman D.G. Moons K.G. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): the TRIPOD statement.Ann Intern Med. 2015; 162: 55-63Crossref PubMed Google Scholar] or in the attempts to score radiomic-based models strongly weighting the model explainability [[28]Lambin P. Leijenaar R.T.H. Deist T.M. Peerlings J. de Jong E.E.C. van Timmeren J. et al.Radiomics: the bridge between medical imaging and personalized medicine.Nat Rev Clin Oncol. 2017; 14: 749-762Crossref PubMed Scopus (1746) Google Scholar]. Methods to reduce statistical redundancy, to select variables better explaining the “signal” and to discard features describing the “noise” (i.e. limiting overfit) can be applied in combination with machine learning and deep learning to make models more robust and generalizable as well as potentially more explainable. This approach also, at least, ensures the “control” on the effect/effect-size of treatment variables which play a key role in predictive models used in medicine. Treatment variables usually “cause” effects, they are not “casually associated” to the outcome of interest; so, we should not accept a model predicting a reduced outcome for an increase in the level of the “causing” effect [[23]Carrara M. Massari E. Cicchetti A. Giandini T. Avuzzi B. Palorini F. et al.Development of a ready-to-use graphical tool based on artificial neural network classification: application for the prediction of late fecal incontinence after prostate cancer radiation therapy.Int J Radiat Oncol Biol Phys. 2018; 102: 1533-1542Abstract Full Text Full Text PDF PubMed Google Scholar]. Of course, this “hybrid” approach does not assure that a complete explainability and interpretability of the resulting models is achieved. Still, it looks like a good compromise in many cases. In the case not all selected best predictors are interpretable, hypotheses can be generated. They can prompt researchers to try to explain why the specific (few) variables were found to be predictive [10Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. Causability and explainability of artificial intelligence in medicine (2019) Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 2019;9(4):art. no. e1312.Google Scholar, 29Astaraki M, Yang G, Zakko Y, Toma-Dasu I, Smedby Ö, Wang C. A comparative study of radiomics and deep-learning based methods for pulmonary nodule malignancy prediction in low dose CT images. Front Oncol 2021;11:art. no. 737368.Google Scholar, 30Miller D.D. The medical AI insurgency: what physicians must know about data to practice with intelligent machines.npj Digit Med. 2019; 2: 62Crossref PubMed Google Scholar, 31Van Der Schaaf A. Langendijk J.A. Fiorino C. Rancati T. Embracing phenomenological approaches to normal tissue complication probability modelling: a question of method.Int J Radiat Oncol Biol Phys. 2015; 91: 468-471Abstract Full Text Full Text PDF PubMed Google Scholar]. On the other hand, black-box AI-based models will never be explainable, even in the case of high performances of the resulting models. AI is radically changing a large variety of activities and our societies rapidly, and none can fully predict what this will mean in the next decades. It is pretty natural that medicine will be largely influenced, maybe in a disruptive way, with potentially large benefits for the community and the individuals. This optimistic vision does not contradict our discussion: the doctors’ caution in moving toward AI is related to the intrinsically ethic basement of medicine that always needs to demonstrate that “a progress is a progress” if patients benefit from it. This is a quite obvious but seldomly forgotten statement: the patient is the center both individually and at larger scales (hospital, regional, national, worldwide). Efficiency, rapidity and cost cutting (all promises of large-scale AI applications) do not translate automatically in better and sustainable care, including the reduction of inequalities in the access to adequate cures. Doctors need to trust in something that can be understood and explained without any new dogmatic absolutism of AI over human intelligence and human empathy. This claims an urgent need for a “democratic” development process for AI implementation in medicine [[32]Hengstler M. Enkel E. Duelli S. Applied artificial intelligence and trust: the case of autonomous vehicles and medical assistance devices.Techn Forecast Soc Changes. 2016; 105: 105-120Crossref Scopus (247) Google Scholar]. The process will empower health practitioners against undesired effects of automated decisions, increase the trust of patients and doctors and help people make better decisions. Within this process, medical physicists have to play a relevant role in “opening up the black box” as Dr Kundu writes in concluding her excellent editorial.