Figures
Citation: Zhang J, Mattie H, Shuaib H, Hensman T, Teo JT, Celi LA (2022) Addressing the “elephant in the room” of AI clinical decision support through organisation-level regulation. PLOS Digit Health 1(9): e0000111. https://doi.org/10.1371/journal.pdig.0000111
Editor: Nadav Rappoport, Ben-Gurion University of the Negev, ISRAEL
Published: September 15, 2022
Copyright: © 2022 Zhang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: The authors received no specific funding for this work.
Competing interests: I have read the journal’s policy and the authors of this manuscript have the following competing interests: Leo Anthony Celi is the Editor-in Chief and Heather Mattie is an Associate Editor of PLOS Digital Health.
Consider the following proprietary artificial intelligence (AI) algorithm products: (1) continual monitoring to predict likelihood of acute kidney injury (Dascena Previse, Dascena, USA); (2) predicting significant events for patients on intensive care (CLEWICU, CLEW Medical, Israel); (3) an early warning system for acute inpatient deterioration (Wave Clinical Platform, Excel Medical, USA); and (4) using electronic health record (EHR) data to predict likelihood of sepsis (Epic Sepsis Model, Epic Systems Corporation, USA).
These algorithms provide early signals of potentially treatable events using real-time clinical data. However, the first three are considered software as a medical device (SaMD) under oversight of the US Food & Drug Administration (FDA) [1–3]. In contrast, the last has undergone no visible regulatory scrutiny [4] and demonstrates minimal data or algorithmic transparency [5], yet is actively used in hundreds of hospitals in the United States that employ the Epic EHR [6]. In 2021, an independent evaluation of this sepsis model demonstrated poor performance (relative to vendor reported metrics), failing to identify 67% of patients with sepsis, with a positive predictive value of 12% and substantial alert burden for clinicians [7]. Other technology vendors [8–10] and healthcare providers [11,12], are also known for hosting development and operationalisation of proprietary algorithmic clinical decision support (CDS). It is likely that many AI implementations fly under the radar.
The elephant then, sitting next to the FDA, is the different consideration given to algorithmic devices for market, and proprietary algorithms developed within existing EHR (traditionally outside of FDA scope [13]). With increasing appearance of CDS, the 21st Century Cures Act 2016 introduced statutory SaMD definitions, such that a non-device CDS is defined by provision of recommendations where clinicians can review the basis for predictions. This could arguably be applied to many algorithms classified as SaMD, and proposed 2019 guidance clarified that CDS must only “recommend” (rather than “drive”) decisions, while creating no intention that “the healthcare provider rely primarily on any of such recommendations to make a clinical diagnosis or treatment decision…” [14]. This distinction remains imprecise. Unlike AI for diagnostic imaging that provides a clear signal (e.g. “there is a nodule”), AI algorithms using EHR data are positioned in complex environments amongst many extraneous considerations; the line between “drive” and “recommend” is consequently blurred, regardless of explainability in underlying intuition, and parallel clinician input is almost always obligatory.
We now observe a resultant dichotomy where the same predictive algorithm might receive different categories of oversight depending on context. This situation poses safety risk:
(1) The FDA considers “recommendation” to pose less risk than decision-making SaMD, but this is arguable. Recommendation flags are an unavoidable additional data-point, and incorrect recommendations may tip decisions towards delayed action or create alert fatigue as much as decision-making SaMD. It is notable that a device for detecting sepsis (AWARE, Ambient Clinical Analytics, USA) received FDA classification of moderate-to-high risk (Class II) whereas the Epic sepsis model was deployed without FDA clearance.
(2) AI CDS largely depend on EHR data. By nature, data quality is variable, being dependent on documentation and coding practices. Demographic data such as race-ethnicity may be missing during training and validation. The risk of algorithmic bias is not trivial and cannot be mitigated by clinician “review” of the recommendation.
(3) AI CDS often produce rapid-cycle recommendations on real-time data with dynamic characteristics, introducing need to re-calibrate/re-train algorithms over time. While FDA has introduced lifecycle [15] and adaptive SaMD [16] guidance, these themes of continuous monitoring are equally relevant to unregulated AI CDS.
(4) Clinicians historically use risk scores to guide decisions [17]. In contrast to proprietary EHR CDS, such risk scores are peer-reviewed and when calculated are used situationally. Decisions to employ risk scores in contextually validated and interpretable environments are taken out of clinicians’ hands; deployment is driven, in part, by incentivised system vendors rather than evidence-based guidelines.
(5) Finally, and most importantly—without requirement for oversight, there is no assurance that CDS are accurate in their predictions; no ‘post-market’ evaluation of unintended consequences; and no confidence that risks are suitably handled. EHR vendors cannot simply reassure providers and patients that their opaque, internal procedures to build these algorithms are robust.
The current climate of AI CDS raises patient safety concerns. Based on 2019 FDA non-binding recommendations, moderate-to-high risk, explainable CDS algorithms will likely remain unregulated. The FDA could decide to expand oversight, for example by including all algorithms above a risk threshold. This would be in line with European Union consideration of any Medical Device software which influences therapeutic decisions at a minimum of Class IIa (requiring notified body assessment) [18]. However, for both FDA and EU MDR bodies, the required scalability to handle future volumes of AI CDS is a challenge [19]. But the resulting bottlenecks may stifle innovation, in a period of accelerating AI development [20].
A possible solution is to embrace this dichotomy and regulate according to differences between device manufacturers (who sell focused devices to a wider market), and healthcare provider/ vendor partnerships (who iterate on numerous and diverse CDS for local adoption). Regulators are transitioning to a lifecycle approach for SaMD, with requirements for manufacturers to demonstrate quality management systems across the entire lifecycle, including continuous safety and effectiveness monitoring. This approach should also apply to AI CDS with oversight of the processes employed to create them, rather than the devices.
System views of regulation have been previously discussed [19,21]. In the context of AI CDS, this means defining “AI-ready” organisation/vendor partnerships that can independently deploy AI algorithms onto internal pathways, while maintaining quality and safety. While proposing a detailed framework is outside scope of this piece, any organisation-level approach must consider: (1) maturity of digital infrastructure; (2) functioning relationships with systems suppliers; (3) clear quality systems for evaluation; (4) workforce training and involvement; and (5) transparency in data, development, and outcomes for external audit. These elements are outlined in greater detail in Table 1.
General good practices that may feed into regulation are laid out in the FDA/MHRA joint principles for Good Machine Learning Practice [22].
There are multiple downstream benefits. Trust is placed in organisations, and organisation-vendor partnerships, that have pre-existing duties of care to patients. Requirement for end-user input will benefit workforce development, and tighter integration will reduce distance from concepts to deployment. Reducing reliance on duplicative assessment of individual CDS promotes innovation and limits the scalability problem. Requirements for representative data and processes to guarantee calibration to under-represented groups will result in richer data sources, and will share the burden of detecting and mitigating algorithmic bias across local stakeholders [23].
This approach risks shutting out less digitally advanced organisations. To safely deploy AI CDS, a data pipeline in addition to AI expertise that are typically found in well-resourced, academic, networks are required. Smaller providers serving disadvantaged populations may be left behind. Regardless of how CDS is regulated in the future, pooling resources, data, and expertise through broad and inclusive collaborations, is vital to democratise AI benefits.
Regulating organisations is outside the traditional regulatory scope of the US FDA, the European Medicines Agency, or the UK Medicines and Healthcare products Regulatory Agency. Whether through expansion of reach, or delegation to separate (or new) agencies, organisational-level regulation may be the only feasible approach to ensuring quality and safety in the increasing number of AI CDS in EHRs.
References
- 1.
Black R. Predictive Patient Surveillance System Receives FDA Clearance. Healthcare Executive. 9 Jan 2018. Available: https://www.chiefhealthcareexecutive.com/view/predictive-patient-surveillance-system-receives-fda-clearance. Accessed 16 Jul 2022.
- 2.
Jercich K. FDA issues landmark clearance to AI-driven ICU predictive tool. Healthcare IT News. 4 Feb 2021. Available: https://www.healthcareitnews.com/news/fda-issues-landmark-clearance-ai-driven-icu-predictive-tool. Accessed 16 Jul 2022.
- 3.
Budwick D. Dascena Receives FDA Breakthrough Device Designation for Machine Learning Algorithm for Earlier Prediction of Acute Kidney Injury. Business Wire. 7 Jul 2020. Available: https://www.businesswire.com/news/home/20200707005149/en/Dascena-Receives-FDA-Breakthrough-Device-Designation-Machine. Accessed 16 Jul 2022.
- 4. Price WN II. Distributed Governance of Medical AI. SSRN Journal. 2022 [cited 16 Jul 2022].
- 5. Habib AR, Lin AL, Grant RW. The Epic Sepsis Model Falls Short—The Importance of External Validation. JAMA Intern Med. 2021;181: 1040. pmid:34152360
- 6. Tarabichi Y, Cheng A, Bar-Shain D, McCrate BM, Reese LH, Emerman C, et al. Improving Timeliness of Antibiotic Administration Using a Provider and Pharmacist Facing Sepsis Early Warning System in the Emergency Department Setting: A Randomized Controlled Quality Improvement Initiative. Critical Care Medicine. 2021;Publish Ahead of Print. pmid:34415866
- 7. Wong A, Otles E, Donnelly JP, Krumm A, McCullough J, DeTroyer-Cooley O, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. JAMA Intern Med. 2021;181: 1065. pmid:34152373
- 8. Jason C. Epic Systems, Cerner Lead EHR Vendors in AI Development. EHR Intelligence. 12 May 2020. Available: https://ehrintelligence.com/news/epic-systems-cerner-lead-ehr-vendors-in-ai-development. Accessed 16 Jul 2022.
- 9. Powles J, Hodson H. Google DeepMind and healthcare in an age of algorithms. Health Technol. 2017;7: 351–367. pmid:29308344
- 10. Bell D, Baker J, Williams C, Bassin L. A Trend-Based Early Warning Score Can Be Implemented in a Hospital Electronic Medical Record to Effectively Predict Inpatient Deterioration. Critical Care Medicine. 2021;49: e961–e967. pmid:33935165
- 11. Clinic Mayo. Mayo Clinic: Emerging Capabilities in the Science of Artificial Intelligence. In: Mayoclinic.org [Internet]. 2021 [cited 19 May 2022]. Available: https://www.mayoclinic.org/giving-to-mayo-clinic/our-priorities/artificial-intelligence
- 12. Sousa K. Partners HealthCare and GE Healthcare launch 10-year collaboration to integrate Artificial Intelligence into every aspect of the patient journey. GE Healthcare (press release). 17 May 2017. Available: https://www.ge.com/news/press-releases/partners-healthcare-and-ge-healthcare-launch-10-year-collaboration-integrate
- 13.
Konnoth C. Are Electronic Health Records Medical Devices? 1st ed. In: Cohen IG, Minssen T, Price WN II, Robertson C, Shachar C, editors. The Future of Medical Device Regulation. 1st ed. Cambridge University Press; 2022. pp. 36–46. https://doi.org/10.1017/9781108975452.004
- 14.
US FDA Center for Devices and Radiological Health. Changes to Existing Medical Software Policies Resulting from Section 3060 of the 21st Century Cures Act: Guidance for Industry and Food and Drug Administration Staff. United States Food & Drug Administration; 2019. Available: https://www.fda.gov/media/109622/download
- 15.
US FDA Center for Devices and Radiological Health. Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. US Food & Drug Administration; 2021.
- 16.
US FDA Center for Devices and Radiological Health. Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD). United States Food & Drug Administration; 2019. Available: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-software-medical-device
- 17. Eugene N, Oliver CM, Bassett MG, Poulton TE, Kuryba A, Johnston C, et al. Development and internal validation of a novel risk adjustment model for adult patients undergoing emergency laparotomy surgery: the National Emergency Laparotomy Audit risk model. British Journal of Anaesthesia. 2018;121: 739–748. pmid:30236236
- 18.
EU Medical Device Coordination Group. MDCG 2021–24 Guidance on classification of medical devices. EU Medical Device Coordination Group; 2021.
- 19. Panch T, Duralde E, Mattie H, Kotecha G, Celi LA, Wright M, et al. A distributed approach to the regulation of clinical AI. Lu HH-S, editor. PLOS Digit Health. 2022;1: e0000040.
- 20. Zhang J, Whebell S, Gallifant J, Budhdeo S, Mattie H, Lertvittayakumjorn P, et al. An interactive dashboard to track themes, development maturity, and global equity in clinical artificial intelligence research. The Lancet Digital Health. 2022;4: e212–e213. pmid:35337638
- 21. Gerke S, Babic B, Evgeniou T, Cohen IG. The need for a system view to regulate artificial intelligence/machine learning-based software as medical device. npj Digit Med. 2020;3: 53. pmid:32285013
- 22.
US FDA Center for Devices and Radiological Health. Good Machine Learning Practice for Medical Device Development: Guiding Principles. United States Food & Drug Administration; 2021.
- 23. Zhang J, Symons J, Agapow P, Teo JT, Paxton CA, Abdi J, et al. Best practices in the real-world data life cycle. McGinnis RS, editor. PLOS Digit Health. 2022;1: e0000003.