European Conference on Interventional Oncology

April 24-27 | Vienna, Austria

April 24-27 | Vienna, Austria

April 24-27 | Vienna, Austria

April 24-27 | Vienna, Austria

April 24-27 | Vienna, Austria

Slider
ProgrammeTopic highlightsCurrent evidence for AI in IO

The current evidence for AI in IO, is there a chance that it can change the way EBM is generated?

We spoke with Prof. Maxime Ronot to learn more about his presentation at ECIO.

You can now watch this session on demand!

Medical imaging is a very active field of research on and development of AI tools. Patient scheduling, image acquisition and reconstruction, image quality analytics, radiation dose estimation, reporting, and data integration, although frequently overlooked, are being profoundly redefined and reorganised by AI. Closer to the medical activity of radiologists, a variety of tasks ranging from simple to complex (detection or interpretation of imaging findings, longitudinal analysis, imaging post-processing, prediction, prognostication, etc.) are subject to considerable ongoing research that is likely to reshape our medical specialty towards augmented-radiology.

In the field of interventional oncology, the potential applications of AI can be divided into pre-procedural (patient selection), peri-procedural (image registration, artefact correction, device selection, and recognition, guidance), and post-procedural (follow-up, prediction of response) [1,2]. Unfortunately, AI in IO is still in its early stages, and evidence supporting its value is scarce. Most results are derived from small-sized retrospective studies, lacking validation or clinical applicability. Nevertheless, these studies can be seen as proofs-of-concept or proof-of-principle, paving the road for more ambitious projects to come and exploring the unexplored in complex patient care possibilities.

While most studies focus on the possible value of AI for patient management (acting as triage, replacement, or add-on technology), one may wonder if its very existence may not lead to more profound changes in medical practice and research conduction, especially regarding the way evidence is gathered and consolidated. In other words, is there a chance that it can change the way evidence-based medicine is generated?

Evidence-based medicine is “the conscientious, explicit and judicious use of current best evidence in making decisions about the care of individual patients” [3]. It is based on a hierarchy of evidence (or levels of evidence) used to rank the relative strength of results obtained from scientific research. It is consensually considered that the highest level of evidence should be obtained from at least one properly designed randomised controlled trial. Unfortunately, most IO evidence is based on observational methods (cohort and case-control studies) that are considered to have little or no value. Consequently, IO studies are often criticised for their lack of solid scientific merit, especially when compared to pharmacological treatments. This perspective fails to acknowledge several well-described limitations of randomised trials (that may be unnecessary, inappropriate, impossible or inadequate), although all limitations can be, at least theoretically, overcome [4]. More importantly, it negates the peculiarities of non-pharmacological treatments. Indeed, an accurate and complete description of most non-pharmacological interventions is challenging. Interventions are usually complex, and each element or step possibly impacts the outcome. This results in difficulty in designing clinically relevant and standardised trials together with poor description and reporting of these interventions [5] that limit the replication of results. This may have severe consequences for patients’ safety and is part of the agenda to reduce waste in clinical research [6]. Ongoing research on composite indicators covering entire intervention processes (e.g., textbook outcome), on adapting reporting systems (e.g., CONSORT-NPT) [7], or the recent increase in the number of prospective IO observational studies or registries, shall be seen as evidence of the growing awareness of these issues in the IO community, rather than mere proof of scientific illiteracy.

Artificial intelligence, especially machine learning, has many strengths that may be used to overcome some of the abovementioned limitations researchers face in the “factory of evidence” in general and in IO in particular. Additionally, it may also improve the way current EMB standards are designed and applied [8]. Machine learning techniques are agnostic and data-driven, with few assumptions around data completeness, accuracy, classification, and independence. They process a considerable amount of already available data sets to identify possible relationships between many variables that are not prespecified and have high diversity. These data can be retrieved from multiple sources (electronic health records, administrative data, imaging data, genomic and proteomic databanks, social media, etc.) via natural language processing and automated data screening. This could not only improve data quality and exhaustivity in observation studies (which are very important for IO), but may also benefit randomised trials. Indeed, algorithms may improve patient selection (reducing population heterogeneity by leveraging electronic phenotyping, prognostic enrichment, predictive enrichment, estimation of individual risks) with direct implications for treatment allocation and stratification in trials [9]. In addition, it may redefine the randomisation process, improve statistical analysis of data, help validate causal inferences, create better reference standards (augmented pathology, imagomics), facilitate appointment scheduling, calculate more appropriate dosing regimens than current algorithms, help compare results from EMB with usual care to assess their applicability and validity in routine practice, facilitate meta-research (living network meta-analyses), or even simulate trials based on existing data. This is, of course, counterbalanced by the well-recognised limitations of AI (dependence of data quality, absence of risk of bias or quality of evidence rating, importance of context, limited explanatory power) which we need to collectively address.

In conclusion, artificial intelligence represents a formidable opportunity to build better IO evidence. Not only will it help address many flaws of observational studies and registries (especially in terms of data exhaustion) we need in IO, but it will help redefine the way high EBM is obtained. In return, to achieve “prime time” clinical application, AI will need to submit to the highest standards of evaluation EBM offers. It is our responsibility to facilitate and structure this virtuous circle.

References

  1. Letzen B, Wang CJ, Chapiro J. The Role of Artificial Intelligence in Interventional Oncology: A Primer. J Vasc Interv Radiol. 2019 Jan;30(1):38-41.e1.
  2. Seah J, Boeken T, Sapoval M, Goh GS. Prime Time for Artificial Intelligence in Interventional Radiology. Cardiovasc Intervent Radiol. 2022 Jan 14
  3. Sackett DL, Rosenberg WM, Gray JA, Haynes RB, Richardson WS. Evidence-based medicine: what it is and what it isn’t. BMJ. 1996 Jan 13;312(7023):71-2
  4. Feinstein AR. An additional basic science for clinical medicine: II. The limitations of randomized trials. Ann Intern Med. 1983 Oct;99(4):544-50
  5. Hoffmann TC, Erueti C, Glasziou PP. Poor description of non-pharmacological interventions: analysis of consecutive sample of randomised trials. BMJ. 2013:10;347:f3755
  6. Cook A, Douet L, Boutron I. Descriptions of non-pharmacological interventions in clinical trials. BMJ. 2013 Sep 11;347:f5212.
  7. Boutron I, Moher D, Altman DG, Schulz KF, Ravaud P; CONSORT Group. Extending the CONSORT statement to randomized trials of nonpharmacologic treatment: explanation and elaboration. Ann Intern Med. 2008 Feb 19;148(4):295-309
  8. Scott IA. Machine Learning and Evidence-Based Medicine. Ann Intern Med. 2018
    Jul 3;169(1):44-46. doi: 10.7326/M18-0115.
  9. Bhatt A. Artificial intelligence in managing clinical trial design and conduct: Man and machine still on the learning curve? Perspect Clin Res. 2021 Jan-Mar;12(1):1-3.

Maxime Ronot

Beaujon Hospital, Paris/FR

Dr. Ronot is a Professor in the Department of Radiology at the Beaujon Hospital (Université de Paris, France). His research interests focus on liver and pancreas diseases, tumours, and interventional abdominal oncology. He is part of the CIRSE AI task force and a member of the LI-RADS AI workgroup. He is a fellow of the ESGAR and serves as president of the French national ethics committee for medical imaging research. Dr. Ronot is associate editor of the Journal of Hepatology, abdominal radiology, section editor of European Radiology, and deputy editor of Diagnostic Interventional Imaging. He authored and co-authored more than 250 articles in peer-reviewed journals. You can follow him on twitter at @maximeronot.