Artificial Intelligence in Medicare: Utilization, Spending, and Access to AI-Enabled Clinical Software

ABSTRACT

Objectives: In 2018, CMS established reimbursement for the first Medicare-covered artificial intelligence (AI)–enabled clinical software: CT fractional flow reserve (FFR_CT) to assist in the diagnosis of coronary artery disease. This study quantified Medicare utilization of and spending on FFR_CT from 2018 through 2022 and characterized adopting hospitals, clinicians, and patients.

Study Design: Analysis, using 100% Medicare fee-for-service claims data, of the hospitals, clinicians, and patients who performed or received coronary CT angiography with or without FFR_CT.

Methods: We measured annual trends in utilization of and spending on FFR_CT among hospitals and clinicians from 2018 through 2022. Characteristics of FFR_CT-adopting and nonadopting hospitals and clinicians were compared, as well as the characteristics of patients who received FFR_CT vs those who did not.

Results: From 2018 to 2022, FFR_CT billing volume in Medicare increased more than 11-fold (from 1083 to 12,363 claims). Compared with nonbilling hospitals, FFR_CT-billing hospitals were more likely to be larger, part of a health system, nonprofit, and financially profitable. FFR_CT-billing clinicians worked in larger group practices and were more likely to be cardiac specialists. FFR_CT-receiving patients were more likely to be male and White and less likely to be dually enrolled in Medicaid or receiving disability benefits.

Conclusions: In the initial 5 years of Medicare reimbursement for FFR_CT, growth was concentrated among well-resourced hospitals and clinicians. As Medicare begins to reimburse clinicians for the use of AI-enabled clinical software such as FFR_CT, it is crucial to monitor the diffusion of these services to ensure equal access.

Am J Manag Care. 2024;30(Spec Issue No. 6):SP473-SP477. https://doi.org/10.37765/ajmc.2024.89556

Takeaway Points

Medicare has begun paying for the use of artificial intelligence (AI)–enabled clinical software. Our study quantified trends in adoption of CT fractional flow reserve (FFR_CT), the first software to be reimbursed by Medicare, beginning in 2018.

• From 2018 through 2022, utilization of FFR_CT in Medicare grew 11-fold.
• FFR_CT utilization was concentrated among a small number of well-resourced hospitals and clinicians with cardiac specialties.
• Patients receiving FFR_CT were less likely to come from historically underserved patient populations.
• As Medicare begins to reimburse clinicians for their use of AI-enabled clinical software such as FFR_CT, it is crucial to monitor the impact on use and spending.

The newest generation of artificial intelligence (AI)–enabled clinical software typically analyzes imaging output to inform clinician decision-making.¹ In the past decade, the FDA has approved AI-enabled clinical software with applications in numerous clinical settings.^2,3 These technologies promise to bolster clinician productivity, improve care quality, and reduce spending.⁴

In 2018, CMS established reimbursement for the first Medicare-covered AI-enabled technology: CT fractional flow reserve (FFR_CT). Approved by the FDA in 2016, FFR_CT technology assists in determining the presence or severity of coronary artery disease (CAD), which affects approximately 18.2 million adults in the US.⁵ FFR_CT applies deep learning to coronary CT angiography (CCTA) scans to measure stenosis (ie, narrowing) of the coronary arteries. HeartFlow, the inventor and sole provider of FFR_CT technology, charges providers either a per-use or subscription-based fee for the technology. To use HeartFlow Analysis, clinicians must have access to a specialized CT scanner that has enough detectors to create 3-dimensional images of the heart (ie, CCTA scans), which are then sent to HeartFlow (via the cloud) for an off-site analysis by a HeartFlow technician. Analysis results are shared with the clinician within a few hours.

FFR_CT has entered clinical guidelines issued by a number of cardiologist societies⁶ as an appropriate test for the evaluation and diagnosis of patients with chest pain. Preliminary evidence suggests that FFR_CT may improve clinical outcomes compared with other testing options.^7-10 Additionally, FFR_CT is marketed to patients and clinicians as a noninvasive alternative to coronary artery angiography, currently the gold-standard method to diagnose and quantify CAD. If FFR_CT truly offers a higher-quality and lower-cost option, understanding adoption patterns is a critical component of evaluating the impact on spending as well as on access. Although survey findings suggest a growing interest in AI technologies,¹¹ we know of no nationwide estimates of their use in health care.

In this study, we took advantage of Medicare reimbursement to observe adoption patterns for FFR_CT using administrative claims data. We measured trends in FFR_CT billing from 2018 to 2022 in Medicare fee-for-service claims. We compared the characteristics of FFR_CT-adopting and nonadopting hospitals and clinicians, as well as the characteristics of patients who did and did not receive FFR_CT. We found 11-fold growth in utilization (from 1083 to 12,363 claims), concentrated among a small number of well-resourced hospitals and patients.

METHODS

We used 100% Medicare fee-for-service claims data from January 2018 through December 2022 to conduct a retrospective study of FFR_CT billing. Providers can bill Medicare to cover the costs of FFR_CT using a set of Current Procedural Terminology codes (eAppendix Table 1 [eAppendix available at ajmc.com]). In 2018, FFR_CT began to be reimbursed by Medicare only in outpatient settings under its Outpatient Prospective Payment System. In 2022, CMS expanded reimbursement to physicians under the Physician Fee Schedule. We identified FFR_CT claims in the Outpatient and Carrier files to capture billing in both settings. To quantify changes in Medicare billing for FFR_CT, we calculated the total number of FFR_CT claims annually (see eAppendix), total Medicare spending, number of unique hospitals billing for FFR_CT, and number of unique clinicians billing for FFR_CT.

We classified short-term acute care hospitals in 1 of 3 ways: (1) those that billed for at least 1 FFR_CT claim from 2018 to 2022, (2) those that billed for CCTA (the basic imaging technology without the AI overlay) and not for FFR_CT, and (3) those that billed for neither. For each hospital, we assessed characteristics of interest in 2019 to avoid detecting secular changes across hospitals during the study period. For hospital characteristics, we relied on 2 data sources: the 2019 American Hospital Association (AHA) Annual Survey and CMS Healthcare Cost Report Information System (HCRIS) public cost report data. Because hospitals report over different time periods, we used the HCRIS report that was based on the largest number of days in 2019—an approach used in existing literature.^12,13 Hospital characteristics of interest included size (total bed count), membership in a hospital system, type (nonprofit, government-owned, or for-profit), whether they reported a positive profit margin, competitive landscape (Herfindahl-Hirschman Index for each hospital’s hospital referral region [HRR], each hospital’s share of total hospital beds in the HRR), rurality, and cardiac care capabilities (total number of cardiac intensive care unit beds and whether a hospital reported providing any cardiac care services).

Similarly, we classified clinicians as FFR_CT-billing or non–FFR_CT-billing based on the presence of at least 1 FFR_CT claim from 2018 to 2022. We assessed clinician characteristics using Medicare claims and National Plan and Provider Enumeration System data from 2019, accessed via the CMS Virtual Research Data Center. Characteristics of interest included specialty category (defined in eAppendix Table 2), credential, and practice size (ie, number of clinicians billing under the same Taxpayer Identification Number).

We compared the characteristics of FFR_CT-billing vs non–FFR_CT-billing hospitals and clinicians. All categorical hospital and clinician characteristics were transformed into binary indicators and compared using 2-sided t tests, with significance defined at the P < .05 level. Hospital comparisons contrasted FFR_CT-billing facilities vs those billing only for CCTA (ie, hospitals equipped with the underlying technology for using FFR_CT) and facilities billing for FFR_CT or CCTA vs those billing for neither service.

We next compared the characteristics of all patients receiving FFR_CT with those of all patients who received a CCTA without subsequent FFR_CT. To focus on patients likely receiving diagnostic care, we included patients with no diagnosis of CAD in the 31 to 365 days prior to their CCTA or FFR_CT. Patient characteristics were obtained from the Master Beneficiary Summary File (MBSF), including age, sex, race/ethnicity, dual eligibility for Medicare and Medicaid, reason for Medicare entitlement, and number of chronic conditions present (out of 27 chronic conditions reported in the MBSF Chronic Conditions segment).^14,15 Patient characteristics were assessed at the time of FFR_CT or CCTA using 2-sided t tests. Categorical variables were transformed into binary indicator variables.

Our sensitivity analysis was limited to patients of clinicians who had ever billed for FFR_CT, and it compared the characteristics of patients receiving FFR_CT with those of patients who received CCTA without FFR_CT, controlling for their treating clinician. This allowed us to understand whether observed differences were due to different patients visiting different hospitals rather than clinicians using FFRCT on different patients within their panel. Specifically, we used ordinary least squares to regress each patient characteristic on an indicator for whether the patient received FFR_CT and a vector of clinician National Provider Identifiers (ie, clinician fixed effects). From this regression, we reported marginal effects computed using Stata 17.0 (StataCorp LLC).

RESULTS

From 2018 through 2022, hospitals and clinicians billed Medicare for FFR_CT 28,343 times for a total of $18.2 million in Medicare spending. Annual FFR_CT claims volume increased from 1083 claims in 2018 to 12,363 claims in 2022, an increase of 1038% (Figure). The number of unique hospitals billing for FFR_CT increased substantially, from 66 in 2018 to 323 in 2022 (eAppendix Figure). Across the 343 hospitals that ever billed for FFR_CT during our study period, the median number of FFR_CT claims billed per hospital was 30, with an IQR of 9 to 90 claims. The highest-billing 20% of hospitals submitted 67.5% of all FFR_CT claims. The number of clinicians using FFR_CT increased from 557 in 2018 to 4205 in 2022.

Among hospitals providing CCTA, 17% also billed Medicare for FFRCT. Hospitals billing for FFR_CT were different from those billing only for CCTA (Table). Compared with CCTA-only hospitals, FFR_CT-billing hospitals were bigger (438.2 vs 240.9 beds; P < .001) and more likely to be part of a larger health system (86.9% vs 74.8%; P < .001). They were more likely to have nonprofit status (84.8% vs 66.9%; P < .001), report positive profit margins (84.0% vs 79.0%; P = .04), and have larger market shares within their HRR (21.6% vs 17.7%; P < .001). FFR_CT-billing hospitals also were more likely to specialize in cardiac care. They had more cardiac intensive care beds on average (13.1 vs 5.1 beds; P < .001) and were more likely to be a cardiac hospital (88.0% vs 69.8%; P < .001) compared with nonbilling hospitals. We repeated this comparison for adopting hospitals with at least 10 FFR_CT claims and found similar results (eAppendix Table 3).

Approximately 43% of all short-term acute care hospitals with AHA survey data used CCTA (with or without FFR_CT). Compared with hospitals not using CCTA, those using CCTA were more likely to be larger, part of a health system, nonprofit, and located in less concentrated and urban markets; to report positive profit margins; and to provide cardiac care (P < .001 for all comparisons) (Table).

Clinicians billing Medicare for FFR_CT were also different from those billing only for CCTA (eAppendix Table 4). More than half of the clinicians billing for FFR_CT were cardiac specialists (52.2% vs 21.0%; P < .001). FFR_CT billers were also more likely to be nonphysicians (11.4% vs 8.6%; P < .001) and work at larger practices (mean, 650.5 vs 360.2 clinicians; P < .001).

Of the 15,873 patients receiving FFR_CT, 88.6% had no prior CAD diagnosis, suggesting that FFR_CT was being used primarily for diagnostic purposes. Patients receiving FFR_CT represented 2.2% of all patients receiving diagnostic CCTA. FFR_CT-receiving patients were different from non–FFR_CT-receiving patients (eAppendix Table 5). Compared with non–FFR_CT-receiving patients, FFR_CT-receiving patients were more likely to be male (52.6% vs 45.3%; P < .001) and less likely to be dually enrolled in Medicaid (7.5% vs 10.7%; P < .001) or receiving disability benefits (11.7% vs 14.4%; P < .001). FFR_CT-receiving patients were also more likely to be White (87.5% vs 85.5%; P < .001). Patient differences remained when we controlled for the billing clinician (eAppendix Table 6).

DISCUSSION

Use of FFR_CT—the first Medicare-reimbursed AI-enabled clinical software—increased rapidly during the first 5 years of reimbursement. Characteristics of FFR_CT-billing hospitals indicated that they were better resourced than CCTA-only billing hospitals. For example, they were more likely to be part of a multihospital system, report positive profit margins, have a dominant position within their hospital market, and already specialize in cardiac care. Less than half of hospitals billed for CCTA, which is required for FFR_CT, suggesting a limited scope for potential adoption. Similarly, FFR_CT-billing clinicians were more likely to be cardiac specialists working in very large practices compared with all clinicians billing for CCTA scans. Medicare patients receiving FFR_CT for diagnostic CAD were more likely to be male, White, and not dually eligible for Medicaid and Medicare compared with similar patients who did not receive FFR_CT. Although some of the differences in patient characteristics are likely driven by the sorting of patients to clinicians and hospitals, these differences persisted when comparing patients who had the same billing clinician.

Many thought leaders have extolled the promise of AI to reduce health care costs and improve clinical outcomes.¹⁶ As Medicare begins to reimburse clinicians for using AI-enabled clinical software, administrative claims data provide an opportunity to study patterns of use of and spending on these services. Although FFR_CT is only a single technology, it is emblematic of many AI-enabled clinical software applications recently reimbursed by Medicare. HeartFlow touts its technology as a low-cost, high-quality replacement for more traditional tests for CAD. For example, HeartFlow reported cost savings of 26% for health systems using FFR_CT vs standard care for patients with planned invasive coronary angiography (a more expensive testing option) in a preliminary study.^17,18

Despite HeartFlow’s marketed cost savings,¹⁸ we found that adoption was concentrated among well-resourced providers and patients. One barrier to adoption may be the hardware investment required to purchase CCTA-enabled CT scanners. We found that many hospitals were not using CCTA, indicating that inequities in access to newer cardiac testing technology existed before the use of FFR_CT. If adoption remains concentrated among the most well-resourced hospitals, FFR_CT and other AI-enabled medical technologies may have the potential to exacerbate existing inequities in cardiac testing.^19-21 Potential disparities in access to AI-enabled clinical software are particularly concerning for rural patients who may face barriers to accessing specialty care due to clinician or medical facility shortages in their areas. AI-enabled clinical software could help address some of these challenges by allowing generalist clinicians to identify patients in need of specialty care.

Not only do these findings raise concerns about inequitable access to care, but they also pose a challenge to the quality of FFR_CT itself. Because AI-enabled clinical software learn from the data they are trained on, the quality of their output relies on the richness of their training data. Manufacturers such as HeartFlow are therefore well motivated to reduce barriers to adoption, which may include skepticism about efficacy and the costs associated with integration into current clinical workflows.^22,23

Although the advent of Medicare reimbursement for AI-enabled clinical software conveniently creates data to understand adoption patterns, it also emphasizes the need to do so. We hope that quantifying adoption patterns will inform future evidence-based payment policy and research quantifying the effect of AI-enabled clinical software on Medicare spending and quality.

Limitations

Our study had several limitations. First, billing for FFR_CT was not necessarily synonymous with adopting FFR_CT technology. However, Medicare reimbursement provides an incentive to bill for FFR_CT when using it, so we expect trends in billing and adoption to be highly correlated. Second, our analysis was limited to Medicare fee-for-service claims, and we cannot draw conclusions about the utilization of FFR_CT among other payers. Third, information on hospital characteristics was self-reported and not all hospitals reported financial information in HCRIS because some categories of hospitals are not required to do so. Fourth, we cannot comment on the clinical appropriateness of FFR_CT for patients. Finally, our comparisons of hospital, clinician, and patient characteristics were descriptive. We cannot make any causal claims about which characteristics had a direct effect on the decision to bill for FFR_CT, nor can we comment on how Medicare reimbursement affected trends in use of FFR_CT.

CONCLUSIONS

We found that the billing of FFR_CT grew 11-fold (from 1083 to 12,363 claims) during the first 5 years of reimbursement. This growth was concentrated among a small number of well-resourced hospitals and clinicians. Patients receiving FFR_CT were less likely to come from historically underserved patient populations.

Author Affiliations: University of Chicago Booth School of Business (AZ, CB), Chicago, IL; Cardiovascular Division, Department of Medicine, Washington University School of Medicine in St Louis (KEJM), St Louis, MO; Department of Health Care Policy, Harvard Medical School (MEC), Boston, MA; Division of Health Policy and Management, School of Public Health, University of Minnesota (HTN), Minneapolis, MN.

Source of Funding: Arnold Ventures.

Author Disclosures: Dr Joynt Maddox is a member of the Centene Health Policy Advisory Council and has grants pending and received from the National Institutes of Health. Dr Chernew reports grant funding for this work from Arnold Ventures and is a co–editor in chief of The American Journal of Managed Care. The remaining authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Authorship Information: Concept and design (AZ, CB, KEJM, MEC, HTN); acquisition of data (HTN); analysis and interpretation of data (AZ, CB, KEJM, MEC, HTN); drafting of the manuscript (AZ, CB, KEJM, MEC, HTN); critical revision of the manuscript for important intellectual content (AZ, KEJM, HTN); statistical analysis (AZ, HTN); and obtaining funding (MEC).

Address Correspondence to: Anna Zink, PhD, University of Chicago Booth School of Business, 5807 S Woodlawn Ave, Chicago, IL 60637. Email: azink@uchicago.edu.

REFERENCES

1. Trang B. Making sense of AI research in medicine, in one slide. STAT. May 22, 2023. Accessed June 4, 2023. https://www.statnews.com/2023/05/22/ai-artificial-intelligence-research-medicine-health/

2. Zhu S, Gilbert M, Chetty I, Siddiqui F. The 2021 landscape of FDA-approved artificial intelligence/machine learning-enabled medical devices: an analysis of the characteristics and intended use. Int J Med Inform. 2022;165:104828. doi:10.1016/j.ijmedinf.2022.104828

3. Artificial intelligence and machine learning (AI/ML)-enabled medical devices. FDA. Updated December 6, 2023. Accessed February 13, 2023. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices

4. Sahni N, Stein G, Zemmel R, Cutler DM. The potential impact of artificial intelligence on healthcare spending. National Bureau of Economic Research working paper 30857. January 2023. Updated October 2023. doi:10.3386/w30857

5. What is coronary heart disease? National Heart, Lung, and Blood Institute. Updated December 20, 2023. Accessed August 25, 2023. https://www.nhlbi.nih.gov/health/coronary-heart-disease

6. Gulati M, Levy PD, Mukherjee D, et al. 2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR guideline for the evaluation and diagnosis of chest pain: executive summary: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation. 2021;144(22):e368-e454. doi:10.1161/CIR.0000000000001030

7. Danad I, Szymonifka J, Twisk JWR, et al. Diagnostic performance of cardiac imaging methods to diagnose ischaemia-causing coronary artery disease when directly compared with fractional flow reserve as a reference standard: a meta-analysis. Eur Heart J. 2017;38(13):991-998. doi:10.1093/eurheartj/ehw095

8. Lu MT, Ferencik M, Roberts RS, et al. Noninvasive FFR derived from coronary CT angiography: management and outcomes in the PROMISE trial. JACC Cardiovasc Imaging. 2017;10(11):1350-1358. doi:10.1016/j.jcmg.2016.11.024

9. Driessen RS, Danad I, Stuijfzand WJ, et al. Comparison of coronary computed tomography angiography, fractional flow reserve, and perfusion imaging for ischemia diagnosis. J Am Coll Cardiol. 2019;73(2):161-173. doi:10.1016/j.jacc.2018.10.056

10. Mickley H, Veien KT, Gerke O, et al. Diagnostic and clinical value of FFRCT in stable chest pain patients with extensive coronary calcification: the FACC study. JACC Cardiovasc Imaging. 2022;15(6):1046-1058. doi:10.1016/j.jcmg.2021.12.010

11. Kent J. 90% of hospitals have artificial intelligence strategies in place. HealthITAnalytics. March 11, 2021. Accessed February 12, 2023. https://healthitanalytics.com/news/90-of-hospitals-have-artificial-intelligence-strategies-in-place

12. McGlave CC, Nikpay SS, Henning-Smith C, Rydberg K, Neprash HT. Characteristics of short-term acute care hospitals that experienced a ransomware attack from 2016 to 2021. Health Aff Sch. 2023;1(3):qxad037. doi:10.1093/haschl/qxad037

13. Nikpay S, Buntin M, Conti RM. Diversity of participants in the 340B Drug Pricing Program for US hospitals. JAMA Intern Med. 2018;178(8):1124-1127. doi:10.1001/jamainternmed.2018.2015

14. Chronic conditions. Chronic Conditions Data Warehouse. Accessed February 13, 2023. https://www2.ccwdata.org/condition-categories-chronic

15. Gorina Y, Kramarow EA. Identifying chronic conditions in Medicare claims data: evaluating the Chronic Condition Data Warehouse algorithm. Health Serv Res. 2011;46(5):1610-1627. doi:10.1111/j.1475-6773.2011.01277.x

16. Agrawal A, Gans J, Goldfarb A, Tucker C. Introduction to “The Economics of Artificial Intelligence: Health Care Challenges.” In: Agrawal A, Gans J, Goldfarb A, Tucker CE, eds. The Economics of Artificial Intelligence: Health Care Challenges. University of Chicago Press; 2023. Accessed April 3, 2023. https://www.nber.org/books-and-chapters/economics-artificial-intelligence-health-care-challenges/introduction-economics-artificial-intelligence-health-care-challenges

17. Douglas PS, De Bruyne B, Pontone G, et al; PLATFORM Investigators. 1-year outcomes of FFRCT-guided care in patients with suspected coronary disease: the PLATFORM study. J Am Coll Cardiol. 2016;68(5):435-445. doi:10.1016/j.jacc.2016.05.057

18. Clinical evidence. HeartFlow. Accessed August 28, 2023. https://www.heartflow.com/clinical-evidence/

19. Wenneker MB, Epstein AM. Racial inequalities in the use of procedures for patients with ischemic heart disease in Massachusetts. JAMA. 1989;261(2):253-257. doi:10.1001/jama.1989.03420020107039

20. Ferguson JA, Tierney WM, Westmoreland GR, et al. Examination of racial differences in management of cardiovascular disease. J Am Coll Cardiol. 1997;30(7):1707-1713. doi:10.1016/s0735-1097(97)00365-3

21. Gorthala S, Chew K, Poon M, Teressa G. Racial disparities in utilization of noninvasive cardiac testing and correlation with mortality. J Am Coll Cardiol. 2021;77(18)(suppl 1):134. doi:10.1016/S0735-1097(21)01493-5

22. He J, Baxter SL, Xu J, Xu J, Zhou X, Zhang K. The practical implementation of artificial intelligence technologies in medicine. Nat Med. 2019;25(1):30-36. doi:10.1038/s41591-018-0307-0

23. Emanuel EJ, Wachter RM. Artificial intelligence in health care: will the value match the hype? JAMA. 2019;321(23):2281-2282. doi:10.1001/jama.2019.4914