Artificial intelligence-derived imaging biomarkers to improve population health
Year: 2020
Type: article
Abstract: In The Lancet Digital Health, Perry Pickhardt and colleagues1Pickhardt PJ Graffy PM Zea R et al.Automated CT biomarkers for opportunistic prediction of future cardiovascular events and mortality in an asymptomatic screening population: a retrospective cohort study.Lancet Digital Health. 2020; (published online March 2.)https://doi.org/10.1016/S2589-7500(20)30025-XSummary Full Text Full Text PDF Scopus (71) Google Scholar demonstrate the potential of artificial intelligence (AI) to extract prognostically relevant information contained within radiological images obtained for different clinical indications by evaluating a panel of automated imaging biomarkers in a colorectal cancer screening population. The authors demonstrate that the discriminative performance of five combined imaging biomarkers is significantly higher than that of body-mass index and the Framingham risk score, an established clinically used risk score for the prediction of future cardiovascular events. Univariate 5-year area under the receiver operating characteristic curve values for predicting death were 0·743 (95% CI 0·705–0·780) for aortic calcification, 0·721 (0·683–0·759) for muscle density, 0·661 (0·625–0·697) for ratio of visceral to subcutaneous fat, 0·619 (0·582–0·656) for liver density, and 0·646 (0·603–0·688) for vertebral density, compared with 0·499 (0·454–0·544) for body-mass index and 0·688 (0·650–0·727) for the Framingham risk score. Although the prognostic value of these biomarkers has been shown before,2Targher G Bertolini L Rodella S et al.Nonalcoholic fatty liver disease is independently associated with an increased incidence of cardiovascular events in type 2 diabetic patients.Diabetes Care. 2007; 30: 2119-2121Crossref PubMed Scopus (410) Google Scholar, 3Martin L Birdsell L Macdonald N et al.Cancer cachexia in the age of obesity: skeletal muscle depletion is a powerful prognostic factor, independent of body mass index.J Clin Oncol. 2013; 31: 1539-1547Crossref PubMed Scopus (1557) Google Scholar, 4Wiklund P Nordström A Jansson J-H Weinehall L Nordström P Low bone mineral density is associated with increased risk for myocardial infarction in men and women.Osteoporos Int. 2012; 23: 963-970Crossref PubMed Scopus (32) Google Scholar, 5Hoffmann U Massaro JM D'Agostino Sr, RB Kathiresan S Fox CS O'Donnell CJ Cardiovascular event prediction and risk reclassification by coronary, aortic, and valvular calcification in the Framingham Heart Study.J Am Heart Assoc. 2016; 5e003144Crossref PubMed Scopus (136) Google Scholar the authors provide proof of principle of how AI algorithms can automate the extraction of established biomarkers for opportunistic screening purposes. Upon successful prospective validation, this AI algorithm could hold enormous potential to identify individuals at high risk for future events in organised or opportunistic screening settings at low additional costs and high speed. However, implementation of such algorithms could also lead to an information overflow without tangible clinical benefits. Thus, rigorous validation and demonstration of actual clinical consequences are required before these methods can be introduced into clinical practice. Here, we describe the potential of AI algorithms as used by Pickhardt and colleagues with implications for the broader field, and we identify obstacles that could hamper clinical introduction. So far, huge amounts of imaging data are acquired every day, but only a small proportion of the information contained in the scans is used to confirm or rule out a medical condition. The vast majority of information in the image is not quantified because reliable and efficient ways for extraction are missing. The authors' study presents an interesting example of how recent advances in automated AI techniques, both engineered and deep learning based, have introduced the possibility for high-throughput quantification of imaging biomarkers in an automated manner. This approach is in line with the ongoing paradigm shift in radiology from a mostly subjective specialty to a more objective discipline.6Hosny A Parmar C Quackenbush J Schwartz LH Aerts HJWL Artificial intelligence in radiology.Nat Rev Cancer. 2018; 18: 500-510Crossref PubMed Scopus (1160) Google Scholar One potential application is organised screening programmes for a specific disease. For example, prediction of future major adverse cardiac events has been shown to be important for individuals screened for cancer.7Jacobs PC Gondrie MJA van der Graaf Y et al.Coronary artery calcium can predict all-cause mortality and cardiovascular events on low-dose CT screening for lung cancer.AJR Am J Roentgenol. 2012; 198: 505-511Crossref PubMed Scopus (122) Google Scholar By doing so, participants with a high cardiovascular risk profile could be referred for further workup to initiate preventive lifestyle changes before the potentially fatal occurrence of an event. Similarly, these tools could be used in an opportunistic approach by application to inpatients and outpatients undergoing an imaging examination for other indications. In this setting, we see a potentially greater benefit than with organised screening programmes because it might have an immediate impact on clinical decision making and patient management. As long as the body region of interest is covered by the scan and acquisition settings are acceptable, biomarkers could be extracted and used for such an opportunistic risk assessment. Although the clinical value of various imaging biomarkers is well known,8Grundy SM Stone NJ Bailey AL et al.2018 AHA/ACC/AACVPR/AAPA/ABC/ACPM/ADA/AGS/APhA/ASPC/NLA/PCNA guideline on the management of blood cholesterol: a report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines.Circulation. 2019; 139: e1082-e1143PubMed Google Scholar and dedicated software packages are available in the clinic to quantify these measures in a semi-automated fashion, they are not routinely applied. This is not only due to the time-consuming nature of quantification but, more importantly, also due to the unknown risks and benefits of clinical implementation. Pickhardt and colleagues' automated AI algorithms address the former; however, the latter remains largely unknown. Therefore, a proof of clinical value is mandated to comprehensively assess the risks versus benefits of such a new system. This proof should include the demonstration of reclassification of risk into different risk categories with therapeutic implications, accounting for concomitant medical therapy and other information available at the time of imaging (such as other diagnostic tests). Other factors that need to be considered are the additional costs of the AI, even though they might be low, as well as false negative or positive findings. From an imaging standpoint, thorough investigation and description of biases and variance are crucial to ensure clinical validity and robustness. This investigation includes the definition of acquisition and reconstruction protocols to account for differences between vendors and institutions. In addition, the expertise and understanding of the end user regarding the use and output of biomarkers need to be explored to provide recommendations for correct application. Consequently, each biomarker should be accompanied by detailed user guidelines, in which requirements and limitations for reliable and safe implementation are described. Although several obstacles—from a development as well as clinical implementation perspective—need to be addressed before widespread use of AI imaging biomarkers is possible, their implementation could represent a fundamental change in the way radiology data are interpreted and used. Leveraging the power of AI techniques for large-scale screening has the possibility to accelerate improvement of population health in several facets of health care. UH reports grants from KOWA, HeartFlow, Duke University (supported by Abbott), and MedImmune/AstraZeneca, and personal fees from Duke University (supported by US National Institutes of Health) and ReCor Medical, outside the submitted work. HJWLA reports personal fees from Onc.AI and grants from the National Institutes of Health, outside the submitted work. JW declares no competing interests. Automated CT biomarkers for opportunistic prediction of future cardiovascular events and mortality in an asymptomatic screening population: a retrospective cohort studyFully automated quantitative tissue biomarkers derived from CT scans can outperform established clinical parameters for presymptomatic risk stratification for future serious adverse events and add opportunistic value to CT scans performed for other indications. Full-Text PDF Open Access
Source: The Lancet Digital Health
Authors: Jakob Weiss, Udo Hoffmann, Hugo J W L Aerts
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $6,680
OpenAlex ID: W3009413379