Characterizing the clinical adoption of medical AI through U.S. insurance claims

Kevin Wu, Eric Wu, Brandon Theodorou, Weixin Liang, Christina Mack, Lucas Glass, Jimeng Sun, James Zou
doi: https://doi.org/10.1101/2023.08.26.23294629

Abstract

There are now over 500 medical AI devices that are approved by the U.S. FDA. However, little is known about where and how often these devices are actually used after regulatory approval. In this paper, we systematically quantify the adoption and usage of medical AI in the U.S. by tracking Current Procedural Terminology (CPT) codes explicitly created for medical AI. CPT codes are widely used for documenting billing and payment for medical procedures, providing a measure of device utilization across different clinical settings. We examine a comprehensive nationwide claims database of 16 billion CPT claims between 1/1/2015 to 6/12023 to analyze the prevalence of medical AI based on submitted claims. Our results indicate that medical AI adoption is still nascent, with most usage driven by a handful of leading devices. For example, only AI devices used for assessing coronary artery disease and for diagnosing diabetic retinopathy have accumulated more than 10,000 CPT claims. Furthermore, medical AI usage is moderately over-represented in higher-income zip codes and metropolitan areas. Our study sheds light on the current landscape of medical AI adoption and usage in the U.S., underscoring the need to further investigate barriers and incentives to promote equitable access and broader integration of AI technologies in healthcare.

Introduction

As artificial intelligence (AI) has rapidly progressed in recent years, significant investments have been devoted to developing and commercializing AI in medicine. As of 2023, over 500 medical AI devices have undergone U.S. FDA evaluation and received approval across areas such as radiology, neurology, and pathology1. During an FDA submission, device manufacturers are required to report evidence of the efficacy and safety of their products, providing crucial insight into how AI algorithms are evaluated before being used on patients2. However, after approval, companies rarely share where and when their products are used. As such, despite the proliferation of medical AI approvals, little is known about their real-world usage.

The usage and adoption patterns of medical AI can significantly impact their clinical impact. First, the performances of AI algorithms are notoriously susceptible to changes in healthcare settings and fluctuate during deployment3,4. For instance, despite initial studies indicating up to a 20% improvement in detection rates, computer-aided detection (CAD) products for mammography approved in the early 2000s have been found to provide no tangible benefits to women5. This discrepancy has been attributed to adoption and usage factors such as changes in clinician interaction with the software and the transition from film to digital mammograms6. Consequently, while AI medical devices may demonstrate strong performance under specific evaluation conditions, variations in real-world applications can yield drastically different outcomes. Second, the impact of medical AI devices is mediated by economic forces. After FDA approval, companies need to find sustainable revenue streams for the promises of AI-driven healthcare to be realized. Different reimbursement approaches can affect how often and on whom these devices are used, and it is still unclear which model is optimal for the new AI devices7,8. Studying the empirical usage of medical AI devices is a crucial step in characterizing the landscape of medical innovations and can provide a more holistic view of the translational pipeline from algorithm to patient.

Recently, Current Procedural Terminology (CPT) codes have been created specifically for medical AI devices7,8. CPT codes are designated by the U.S. Department of Health and Human Services under the Health Insurance Portability and Accountability Act (HIPAA) as a national coding set for physicians and other healthcare professional services and procedures to be used by the Centers for Medicare & Medicaid Services (CMS)9. The codes are regularly created, updated, modified by the American Medical Association (AMA), and are the most widely accepted medical nomenclature under public and private health insurance programs9. Healthcare providers use these codes to generate itemized bills detailing the specific services delivered to a patient during a medical encounter. Subsequently, these bills are submitted to insurance companies, who use the coded information to determine the appropriate reimbursement for the services rendered. As such, CPT codes play a crucial role in ensuring the accuracy and uniformity of medical billing, as well as promoting accountability and transparency within the healthcare system.

CMS also provides coverage for medical AI through New Technology Add-On Payment (NTAP) which is specifically designed to encourage healthcare providers to adopt new technologies7. However, NTAP specifically focuses on inpatient payments, whereas CPT codes apply to both inpatient and outpatient settings10,11. In this paper, we focus on CPT codes as they are most widely adopted and standardized across both public and private insurance programs9, whereas NTAP is specifically used within Medicare11, presenting only a partial view of national AI usage. Additionally, due to its extensive and long-term adoption by healthcare payers, CPT is also an informative resource for comparing baseline usage rates of non-AI devices.

While an increasing number of CPT codes have been made available for medical AI devices, these codes are generally spread across various medical domains and reserved for medical coders and insurance companies. As such, there currently does not exist a single database of AI-related CPT codes or a systematic analysis of their usage. In this paper, we identify and organize a comprehensive list of CPT codes that apply to medical AI devices. We analyze the usage of these codes on a large national claims database and present their temporal and geographical trends.

Related works

Previous analyses have focused on translational roadblocks for medical AI stemming from model evaluation, ethics, and reporting2,12. Specific studies have shown how AI algorithms can perform worse in clinical practice despite promising retrospective evaluations13,14. A variety of studies have analyzed the emergence of reimbursement mechanisms for medical AI products. For example, researchers have highlighted Viz.AI’s NTAP payment model and its potential impact on stroke care, as well as the economic challenges of adopting LumineticsCore from a cost-benefit perspective11,15. Current payment models for AI have been previously analyzed along with examples of reimbursable AI devices7. More specifically, a recent study has proposed a framework for analytically determining the value and cost of each unique AI service in order to encourage ethical and optimal deployment8.

While our work is the first to analyze AI usage through CPT codes, several studies have analyzed geographical distributions present in AI development. For example, researchers have analyzed PubMed for the training datasets used in various medical AI algorithms and found that the data are disproportionately located in California, Massachusetts, and New York16. Datasets used in AI skin cancer diagnosis have also been exclusively found to be from Europe, North America, and Oceania17. The usage of CPT codes for digital health technologies like remote physiologic monitoring, eConsults, and eVisits have also been systematically studied by reporting the total number of claims in Medicare data18. Our work focuses specifically on the subset of digital health relevant to artificial intelligence and machine learning.

Methods

Our analysis consists of two main parts: the organization of medical AI CPT codes and the analysis of their usage. First, to find CPT codes used for medical AI devices, we use a combination of official sources, web resources, and insurance company policies. Second, we searched a large national claims database to quantify the usage of each code.

Collecting CPT Codes for Medical AI

Official AMA sources

The AMA develops CPT codes and is responsible for the development of new billing codes for medical AI products. The CPT Editorial Panel has issued guidance for classifying AI applications (Appendix S), which includes assistive, augmentative, and autonomous work19, but only a few examples of AI codes are referenced. For a comprehensive list of new CPT codes, we processed the AMA’s list of Category III codes (accessed 3/1/23 20), which are a set of temporary codes assigned to emerging technologies, services, and procedures21. While these codes are billed like all other codes, Category III codes are intended to be used primarily for data collection to substantiate widespread usage before granting reimbursement. After five years, they are re-evaluated and replaced with a Category I code if deemed qualified. We analyzed each of the AMA’s Category III codes (long descriptors) for the terms artificial intelligence and machine learning and their variants. Next, for Category I and II codes, we perform a comprehensive search using Codify by AAPC, a search engine for CPT codes22.

CPT code long descriptors provide limited information on the underlying technology behind the procedure and the product name. As such, we complement the CPT code descriptions with details provided by insurance companies in policy documents. Such documents provide detailed descriptions of a given procedure, as well as any medical evidence that might support the case for its reimbursement. Additionally, the policies often reference specific product names that the CPT codes refer to. We analyzed the policies of Premera, Amerigroup, and Blue Cross and noted products that were referred to as artificial intelligence or machine learning devices23–25.

Determining AI devices

We determine whether each candidate CPT code bills for an AI medical device if either of the following criteria is met: (1) the device manufacturer makes explicit marketing claims that its product uses AI and/or machine learning (ML), or (2) a third-party (e.g. insurance company, news publication) refers to the product as powered by AI and/or ML. Additionally, we exclude CPT codes which are also used for billing non-AI devices, as this dilutes the number of AI-specific occurrences. For example, recently, AI has been integrated into a continuous glucose monitoring device, but other non-AI devices are billed under the same code. Another example includes mammography with computer-aided detection (CAD), which is largely dominated by traditional CAD and should be differentiated from modern CAD products26. Next, several CPT codes exist for machine learning-based proprietary lab tests (identified with the letter ’U’), but are excluded from this study since they are designed for specific laboratories and are typically not reimbursed by third-party payers. Finally, to focus our analysis on the usage of recently developed AI, we only include CPT codes developed after 2015.

Grouping CPT Codes

Multiple CPT codes may be related to the same underlying medical procedure but describe different aspects of the procedure. For example, both 0648T and 0649T are used to report quantitative magnetic resonance analysis of tissue composition, but 0649T is used when a diagnostic MRI is also completed, while 0648T is used when it is not. In our analysis, we organize codes that refer to the same underlying medical AI procedure into a CPT code group. To this end, we compute the sum total of all codes in that code group when reporting the number of claims for each procedure.

Claims Data

IQVIA PharMetrics® Plus

We use the IQVIA PharMetrics® Plus for MedTech dataset, a longitudinal health plan database of medical and pharmacy claims27. The dataset consists of more than 210 million unique U.S. enrollees and is comprised largely of commercial health plans. The data are compliant with the Health Insurance Portability and Accountability Act (HIPPA) and are representative of the commercially insured US national population for patients under 65 years of age28. The IQVIA dataset is commonly used for analyses of medical trends in areas like infectious diseases29–33, cardiology34, dermatology35, pulmonology36, oncology37, and neurology38–40. The unit measurement we use in our analysis is a medical claim that uses a CPT code associated with a medical AI procedure. We analyze usage in all 50 U.S. states from 01/01/2015 to 06/01/2023; the dataset consists of 16 billion claims in total. As a point of reference, CMS reports that there are a total of 5 billion claims processed in the US per year 41, which suggests that our dataset has around 40% coverage of all US claims.

Finding Associated Device Names and FDA Approvals

To provide the commercial context for each CPT code, we also locate specific device names associated with each AI CPT code by searching through insurance policies as well as company websites. While we were able to locate at least one product for each procedure, the list may not be comprehensive if not indicated by the company or a third-party source. For the top products we found, we also located their corresponding FDA approval (if applicable) to provide a timeline context for the overall translational pipeline for each product.

Geographic analysis

For each medical AI procedure, we aggregate all unique zip codes that contain an occurrence of at least one code. First, we search for each zip code’s median income and classify it as high-income if it exceeds $100,000, consistent with the IRS’s classification42. Next, we determine whether it is in a metropolitan area by referencing the U.S. Department of Agriculture’s Rural-Urban Commuting Area Codes (RUCA)43. Finally, we compute the percent of all unique zip codes that (1) have a high median income and (2) are metropolitan. We compare these rates to the rates found for all US zip codes, as well as unique zip codes found in a random sample of one million claims (across all CPT codes).

Insurance Pricing

In addition to CPT code billing frequencies, we collect public and private pricing information. First, when available, we look up Medicare pricing for each CPT code that has been made publicly available each year44. Second, we gather negotiated pricing rate data from Anthem Healthcare in California and New York, focusing specifically on in-network rates as of November 2022. This data is made available as part of the Transparency in Coverage (TIC) regulation, which was introduced by the Tri-Agencies (U.S. Departments of Health and Human Services, Labor, and Treasury) on 11/12/202045. The regulation requires health plans to publish their negotiated rates for all items and services for commercial coverage, including in-network files, in machine-readable formats, with monthly updates starting from July 1, 2022. We utilize the November 2022 version of the in-network rate files, which are provided in the Centers for Medicare & Medicaid Services (CMS) defined JSON format.

Results

Billable Medical AI

Given our methodology, we find a total of 16 medical AI procedures billable under CPT codes. Several procedures can be reimbursed through multiple codes, comprising a total of 32 unique CPT codes that are associated with AI. These procedures are detailed in Table 1, alongside the total number of claims containing the codes, product name, and effective date of the codes. The procedures fall within a wide range of healthcare areas, such as cardiology, radiology, and ophthalmology, and have been created very recently, with 15 of 16 medical AI procedures created since 2021 (Figure 1). We find that only 4 out of 16 have more than 1,000 total claims. This is partially because the median age of a medical AI procedure is only about a year (374 days) old.

Figure 1:
Figure 1: Capturing Medical AI Usage in Claims Data.

(A) Medical device usage is captured through billing records and aggregated in our claims database. Each service/procedure that uses a medical device is associated with a CPT code that hospitals and medical practices report for billing purposes. On the right, we provide a map of the geographical distribution of zip codes for a random sample of 1 million claims (out of 16B in our dataset from 01/01/2015-06/01/2023) for comparison with AI CPT codes. Each blue dot represents a single unique zip code where the procedure was billed. (B) We provide details on the top four billable medical AI procedures through CPT codes. Under the procedure name, we list the CPT code(s) and a map of the geographical distribution of zip codes where they have been billed (cumulative over time). Finally, in the bottom row, we provide examples of billable products under each code and a product description image taken from marketing materials from the respective companies.

View this table:
Table 1: Summary of AI CPT codes

A total of 16 medical AI procedures are presented alongside their corresponding CPT code(s). Each procedure is associated with an example commercial product that may be reimbursed through the codes. The effective date is the date on which the code is officially recognized by the American Medical Association (AMA) and can be used for billing and reimbursement purposes. The total claims listed are recent as of 06/01/2023.

Growth Patterns of Medical AI

We find that the overall utilization of medical AI products is still limited and focused on a few leading procedures. However, utilization has generally increased exponentially for each medical AI procedure (Figure 2). The procedure with the most AI usage is coronary artery disease (n=67306, effective 01/01/2018). The associated CPT codes can be used to reimburse products like HeartFlow FFRCT, a medical device that uses computed tomography (CT) scans to create a 3D model of the coronary arteries. The model is then used to calculate the fractional flow reserve (FFR), which is a measure of how well blood flows through the arteries. Among other functions, FFRct can be used to diagnose coronary artery disease and assess the severity of the disease46. HeartFlow FFRCT was given its first FDA approval in 2019 and has had two subsequent updates since47. In November of 2021, CMS set the national payment rate of the device at $930.34 for an office-based setting48.

Figure 2:
Figure 2: Growth of Medical AI in CPT Codes.

(Top) The number of claims per month of each medical AI procedure between the dates of 01/01/2018 to 06/01/2023. The top four procedures by total claims are presented in colors, while the remaining 12 are grouped and added together into an "Other" group in gray. On the right-hand side, we provide a legend for each of the medical AI procedures. These procedures are further grouped by their usage tiers (0-100, 100-1000, and 1000+ total claims). All procedures in the "Other" category are contained in the callout box on the bottom right. (Bottom) We present the cumulative number of CPT AI medical procedures available each year from 2018-2023.

Diabetic retinopathy medical AI has also grown exponentially in usage (n=15097, effective 01/01/2021). The first FDA approval in this category was given on 1/12/2018 to LumineticsCore49, an AI diagnostic system that autonomously diagnoses patients for diabetic retinopathy (including macular edema)50. It is indicated for use by healthcare providers to automatically detect more than mild diabetic retinopathy (mtmDR) in adults diagnosed with diabetes who have not been previously diagnosed with diabetic retinopathy49. The product takes images of the back of the eye, analyzes them, and provides a diagnosis. If more than a mild case is detected, the patient is referred to a specialist51. In 2021, the national payment rate set by CMS for CPT code 92229 was $45.3652.

We also find exponential growth at a smaller scale occurring in medical AI for coronary atherosclerosis and liver MR. Cleery’s Coronary Computer Tomography Angiography (CCTA) algorithm (n=4459, effective 01/01/2021) received its first FDA approval on 10/9/2019 and aims to identify atherosclerosis, the plaque buildup in the arteries of the heart, as well as vascular morphology features for all identified arteries in the CCTA data53. While pricing for this code is not given through CMS, we find it has a median negotiated rate of $371.55 in Anthem’s CA and NY pricing data. Perspectum’s LiverMultiScan (n=2428, effective 01/01/2021) is a non-invasive diagnostic technology for evaluating liver diseases present in multiparametric MRI by quantifying liver tissue54. Receiving its FDA approval on 09/06/2017, it provides a number of quantification tools, such as Region of Interest (ROI) placements, to be used for the assessment of regions of an image to aid in the diagnosis of liver disorders47. The associated CPT code, 0648T, does not have a national payment rate through CMS, given its status as a Category III code. However, in Anthem’s CA and NY pricing data, we find that it reimburses for a median negotiated rate of $692.91.

Finally, we also observe that several procedures have had only nominal or zero usage. CT Vertebral Fracture Assessment and Noninvasive Arterial Plaque Analysis only had a single occurrence in our CPT database since 01/01/2022, and procedures (Facial Phenotype Analysis and X-Ray Bone Density) did not have any occurrences in our database.

Characteristics of Deployed Zip Codes

Next, we analyze the unique zip codes where medical AI procedures are deployed. We find that 32% of zip codes where AI are deployed are high-income (>$100k median household income), which is significantly higher than non-AI claims (17%, p<0.001), as well as the US general population average (10%, p<0.001). An average of 89% of the zip codes for AI are metropolitan, which is much higher than the US average (41%, p<0.001), and marginally higher compared to the random sample of non-AI claims (87%, p=0.002) (Figure 3). Additionally, we provide a map of the geographical distribution of claims for the top 4 medical AI procedures for each year of their availability and find that usage is generally well distributed across coasts and regions of the US (Supp. Figure 1). Additionally, we list out the top 10 zip codes and city/state for each of the top 4 medical AI procedures in our analysis (Supp. Table 1).

Figure 3:
Figure 3: Representation of Zip Codes in AI Procedures

All unique zip codes are categorized by income level and metropolitan area. The top row shows the percentage of all unique zip codes that have a median income above $100K for AI devices, all medical devices (including non-AI), and the general US population. The bottom row shows the percentage of metropolitan zip codes across the same three categories. The black bars represent 95% confidence intervals based on bootstrapped samples.

Discussion

Our study finds that the commercialization of FDA-approved AI products is still nascent but growing, with over 50% of CPT codes effective since 2022. However, only a handful of these devices have reached substantial market adoption, suggesting that the medical AI landscape is still in its early stages. Such usage patterns underscore key themes regarding the deployment of AI in medicine, including clinical implementation challenges, payment, and equal access.

Successful clinical adoption of medical AI involves overcoming key implementation barriers. First, the addition of AI may require significant changes to the clinical workflow. For example, studies have detailed how the success of a diabetic retinopathy detection algorithm is mediated by deployment factors like patient consent, internet speed and connectivity, and poor lighting conditions64,65. Another study finds that the added benefit of an AI algorithm in pathology depends on the pathologist’s interaction with the algorithm’s outputs66. Moreover, the value of an AI algorithm is to clinical practices67. For instance, researchers have argued that clinics that use diabetic retinopathy algorithms may operate at a deficit for every patient evaluated and propose modifications to the existing payment structure to encourage adoption15. On the other hand, patients may be incentivized to visit practices that provide state-of-the-art technologies. Medical AI devices need to have a clear value proposition to healthcare providers to achieve widespread adoption, but the value of AI is multifaceted and context-dependent68,69.

In particular, Medicare pricing for medical AI can provide insight into how AI is currently valued. The reimbursement amounts for CPT codes are determined based on three factors: physician work, practice expense, and malpractice cost70. For a given code, each factor is associated with a relative value unit (RVU) that is adjusted to account for differences between procedures. For example, a higher RVU for physician work means the procedure involves more physician time and/or expertise. A key value proposition of medical AI devices is their ability to reduce or remove the work burden of physicians. We find this reflected in the pricing for CPT code 92229 (diabetic retinopathy) in the CMS fee schedule. Despite having a relative value of 0 for physician work, the practice expense relative value (peRVU) for this code is 1.34, which is higher than its non-AI counterpart (CPT code 92228, peRVU=0.53)71. This difference illustrates how the pricing of AI devices shifts some of the value typically assigned to physicians toward the costs of purchasing and operating the device itself.

While we find an increasing rate of utilization of medical AI for several market leaders, we also note disparities in the adoption of these technologies across different healthcare settings. Our findings indicate that, on average, AI devices are deployed in wealthier zip codes at a higher rate than a random sample of claims for all CPT codes, especially in the initial stages of deployment. Such disparities could be due to various factors, such as the cost of the devices and the financial resources required to implement and maintain these technologies. Additionally, healthcare providers in wealthier areas may be more inclined to invest in cutting-edge technologies to attract patients and maintain a competitive edge. Policymakers and stakeholders should implement strategies that promote equitable access to medical AI technologies to prevent further widening of existing healthcare disparities.

Our analysis of medical AI usage has several limitations. First, while our dataset of 16 billion claims (IQVIA PharMetrics® Plus) is representative of the US patient population under 65, it does not capture all medical claims. As such, the number of claims reported in our work only represents a fraction of total usage and should mainly be interpreted through its relative magnitude over time. Second, our analysis focuses specifically on CPT codes, which do not capture all potential types of AI usage. For example, products such as Viz.ai’s LVO detection algorithm are reimbursable under Medicare’s NTAP program, but we do not capture such usage in our study. Additionally, medical AI usage in clinical pilot studies that are not reimbursed will not appear in large national databases. Our analysis also does not capture the usage of medical AI devices that are billed under non-AI-specific CPT codes. For example, CPT code 77066 is used for computer-assisted detection (CAD) for mammograms but does not differentiate the usage of current deep learning-based approaches from older traditional models from the 1990s. As such, while new models for mammography are developed and approved by the FDA, their usage can not be cleanly identified in claims data.

The usage and adoption of medical AI are the product of a complex ecosystem involving AI developers, healthcare providers, payers, and patients. While the last few years have seen rapid growth in the capabilities of AI, careful consideration of forces beyond algorithmic development is required for AI models to have a meaningful clinical impact. As such, monitoring the usage and clinical adoption of medical AI is key to ensuring that these new technologies fulfill the promise of improving the quality of healthcare for broad patient populations.

Data Availability

The IQVIA Pharmetrics Plus claims dataset utilized in this research is available for licensing or for research through IQVIA. The insurance pricing data employed in this study is publicly accessible as part of the Transparency in Coverage (TIC) regulation and can be obtained directly through CMS at https://www.cms.gov/healthplan-price-transparency.

Data Availability

All data produced in the present study are available upon reasonable request to the authors

Supplementary Table and Figures

View this table:
Supp Table 1:

Top 10 sites by the total number of claims across all years for the top 4 medical AI procedures. Each cell represents the major city and state for a unique zip code associated with a medical AI procedure.

Supp Figure 1:
Supp Figure 1: Geographic Distribution of Claims Over Time

The top four most used medical AI procedures are shown geographically across time. Plots for 2023 are recent to 6/1/2023. Each blue dot represents a unique zip code where the AI device is used in a given year. Plots include years from when a medical AI procedure for a device was reported until 2023.

Acknowledgement

This research is supported by funding from the Chan-Zuckerberg Initiative.

References

  1. 1.↵
    Center for Devices & Radiological Health. Artificial intelligence and machine learning (AI/ML)-enabled medical devices. U.S. Food and Drug Administration https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices.
  2. 2.↵
    Wu, E. et al. How medical AI devices are evaluated: limitations and recommendations from an analysis of FDA approvals. Nat. Med. 27, 582–584 (2021).
    OpenUrlCrossRefPubMed
  3. 3.↵
    Wong, A. et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. JAMA Intern. Med. 181, 1065–1070 (2021).
    OpenUrl
  4. 4.↵
    Wu, E., Wu, K. & Zou, J. Explaining medical AI performance disparities across sites with confounder Shapley value analysis. arXiv [cs.LG] (2021).
  5. 5.↵
    Lehman, C. D. et al. Diagnostic Accuracy of Digital Screening Mammography With and Without Computer-Aided Detection. JAMA Intern. Med. 175, 1828–1837 (2015).
    OpenUrl
  6. 6.↵
    Fenton, J. J. Is It Time to Stop Paying for Computer-Aided Mammography? JAMA internal medicine vol. 175 1837–1838 (2015).
    OpenUrl
  7. 7.↵
    Parikh, R. B. & Helmchen, L. A. Paying for artificial intelligence in medicine. NPJ Digit Med 5, 63 (2022).
  8. 8.↵
    Abràmoff, M. D. et al. A reimbursement framework for artificial intelligence in healthcare. NPJ Digit Med 5, 72 (2022).
  9. 9.↵
    American Medical Association. CPT® overview and code approval. American Medical Association https://www.ama-assn.org/practice-management/cpt/cpt-overview-and-code-approval (2019).
  10. 10.↵
    Hirsch, J. A. et al. Current procedural terminology; a primer. J. Neurointerv. Surg. 7, 309– 312 (2015).
    OpenUrlAbstract/FREE Full Text
  11. 11.↵
    Hassan, A. E. New Technology Add-On Payment (NTAP) for Viz LVO: a win for stroke care. J. Neurointerv. Surg. 13, 406–408 (2021).
    OpenUrlFREE Full Text
  12. 12.↵
    Wiens, J. et al. Do no harm: a roadmap for responsible machine learning for health care. Nat. Med. 25, 1337–1340 (2019).
    OpenUrlCrossRefPubMed
  13. 13.↵
    Kanagasingam, Y. et al. Evaluation of Artificial Intelligence-Based Grading of Diabetic Retinopathy in Primary Care. JAMA Netw Open 1, e182665 (2018).
    OpenUrl
  14. 14.↵
    Widely used sepsis prediction tool is less effective than Michigan doctors thought. NHLBI, NIH https://www.nhlbi.nih.gov/news/2021/widely-used-sepsis-prediction-tool-less-effective-michigan-doctors-thought.
  15. 15.↵
    Chen, E. M., Chen, D., Chilakamarri, P., Lopez, R. & Parikh, R. Economic Challenges of Artificial Intelligence Adoption for Diabetic Retinopathy. Ophthalmology 128, 475–477 (2021).
    OpenUrl
  16. 16.↵
    Kaushal, A., Altman, R. & Langlotz, C. Geographic Distribution of US Cohorts Used to Train Deep Learning Algorithms. JAMA 324, 1212–1213 (2020).
    OpenUrlCrossRef
  17. 17.↵
    Wen, D. et al. Characteristics of publicly available skin cancer image datasets: a systematic review. Lancet Digit Health 4, e64–e74 (2022).
    OpenUrlCrossRefPubMed
  18. 18.↵
    Kvedar, J. C., Mittermaier, M. & Pritzker, J. The industry impact of the American Medical Association’s Digital Medicine Payment Advisory Group (DMPAG). NPJ Digit Med 5, 193 (2022).
  19. 19.↵
    American Medical Association. CPT Appendix S: AI taxonomy for medical services & procedures. American Medical Association https://www.ama-assn.org/practice-management/cpt/cpt-appendix-s-ai-taxonomy-medical-services-procedures (2021).
  20. 20.↵
    American Medical Association. Category III Codes. American Medical Association https://www.ama-assn.org/practice-management/cpt/category-iii-codes (2023).
  21. 21.↵
    Thorwarth, W. T., Jr.. CPT®: an open system that describes all that you do. J. Am. Coll. Radiol. (2008).
  22. 22.↵
  23. 23.↵
  24. 24.
  25. 25.↵
  26. 26.↵
    Gao, Y., Geras, K. J., Lewin, A. A. & Moy, L. New Frontiers: An Update on Computer-Aided Diagnosis for Breast Imaging in the Age of Artificial Intelligence. AJR Am. J. Roentgenol. 212, 300–307 (2019).
    OpenUrl
  27. 27.↵
  28. 28.↵
    Snowflake. https://app.snowflake.com/marketplace/listing/GZSOZ5C6AR/iqvia-iqvia-pharmetrics%C2%AE-plus-claims-data.
  29. 29.↵
    Feuerstadt, P. et al. Healthcare resource utilization and direct medical costs associated with index and recurrent Clostridioides difficile infection: a real-world data analysis. J. Med. Econ. 23, 603–609 (2020).
    OpenUrlCrossRef
  30. 30.
    Chen, J. et al. Changes and geographic variation in rates of preterm birth and stillbirth during the prepandemic period and COVID-19 pandemic, according to health insurance claims in the United States, April-June 2019 and April-June 2020. Am J Obstet Gynecol MFM 4, 100508 (2022).
    OpenUrl
  31. 31.
    Chua, K.-P., Conti, R. M. & Becker, N. V. US Insurer Spending on Ivermectin Prescriptions for COVID-19. JAMA 327, 584–587 (2022).
    OpenUrl
  32. 32.
    Barrett, C. E. et al. Risk for Newly Diagnosed Diabetes >30 Days After SARS-CoV-2 Infection Among Persons Aged <18 Years - United States, March 1, 2020-June 28, 2021. MMWR Morb. Mortal. Wkly. Rep. 71, 59–65 (2022).
    OpenUrlCrossRef
  33. 33.↵
    Jolles, S. et al. Risk factors for severe infections in secondary immunodeficiency: a retrospective US administrative claims study in patients with hematological malignancies. Leuk. Lymphoma 63, 64–73 (2022).
    OpenUrl
  34. 34.↵
    Berger, J. S. et al. Real-world effectiveness and safety of rivaroxaban versus warfarin among non-valvular atrial fibrillation patients with obesity in a US population. Curr. Med. Res. Opin. 37, 881–890 (2021).
    OpenUrl
  35. 35.↵
    Murage, M. J. et al. Treatment patterns, adherence, and persistence among psoriasis patients treated with biologics in a real-world setting, overall and by disease severity. J. Dermatolog. Treat. 30, 141–149 (2019).
    OpenUrlPubMed
  36. 36.↵
    Most, J. F. et al. Real-World Assessment of Asthma Specialist Visits Among U.S. Patients with Severe Asthma. J. Allergy Clin. Immunol. Pract. 9, 3662–3671.e1 (2021).
    OpenUrl
  37. 37.↵
    Cohen, J. T. et al. Are National Comprehensive Cancer Network evidence block affordability ratings representative of real-world costs? An evaluation of advanced non-small-cell lung cancer. J. Oncol. Pract. 15, e948–e956 (2019).
    OpenUrl
  38. 38.↵
    dosReis, S., Saini, J., Hong, K., Reeves, G. & Spence, O. M. Trends in antipsychotic use for youth with attention-deficit/hyperactivity disorder and disruptive behavior disorders. Pharmacoepidemiol. Drug Saf. 31, 810–814 (2022).
    OpenUrl
  39. 39.
    Kalilani, L. et al. Assessment and effect of a gap between new-onset epilepsy diagnosis and treatment in the US. Neurology 92, e2197–e2208 (2019).
    OpenUrl
  40. 40.↵
    Royston, M. et al. Neuromyelitis Optica Spectrum Disorder: Clinical Burden and Cost of Relapses and Disease-Related Care in US Clinical Practice. Neurol Ther 10, 767–783 (2021).
    OpenUrl
  41. 41.↵
  42. 42.↵
    IRS increases visits to high-income taxpayers who haven’t filed tax returns. https://www.irs.gov/newsroom/irs-increases-visits-to-high-income-taxpayers-who-havent-filed-tax-returns.
  43. 43.↵
    Cromartie, J. Rural-urban commuting area codes. https://www.ers.usda.gov/data-products/rural-urban-commuting-area-codes/.
  44. 44.↵
  45. 45.↵
  46. 46.↵
  47. 47.↵
    Brindza, L. J. What is a premarket notification 510(k)? Clin. Microbiol. Newsl. 2, 4–5 (1980).
    OpenUrlCrossRef
  48. 48.↵
    CMS sets payment rate for HeartFlow’s FFR-CT solution in 2022 Medicare Physician Fee Schedule. Cardiovascular Business https://cardiovascularbusiness.com/topics/cardiac-imaging/cms-heartflow-2022-medicare-physician-fee-schedule.
  49. 49.↵
    Office of the Commissioner. FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems. U.S. Food and Drug Administration https://www.fda.gov/news-events/press-announcements/fda-permits-marketing-artificial-intelligence-based-device-detect-certain-diabetes-related-eye.
  50. 50.↵
  51. 51.↵
    Savoy, M. IDx-DR for Diabetic Retinopathy Screening. Am. Fam. Physician 101, 307–308 (2020).
    OpenUrl
  52. 52.↵
    Larson, P. M. Medicare: What’s New For 2022. https://www.reviewofophthalmology.com/article/medicare-whats-new-for-2022 (2022).
  53. 53.↵
    Medicare carrier underprices new AI screening code, breaking from peers. American Academy of Ophthalmology https://www.aao.org/eye-on-advocacy-article/medicare-carrier-underprices-new-ai-screening-code (2021).
  54. 54.↵
  55. 55.
    Heartflow. Category III CPT® Codes for FFRCT. Preprint at https://cdn-corpweb.heartflow.com/assets/docs/Category_III_CPT_Codes_for_FFRCT.pdf (2023).
  56. 56.
    Perspectum. Perspectum announces American medical association issues unique category III CPT® codes for LiverMultiScan quantitative multiparametric MR. https://www.perspectum.com/our-company/news/perspectum-announces-american-medical-association-issues-unique-category-iii-cpt-codes-for-livermultiscan-quantitative-multiparametric-mr/.
  57. 57.
    Perspectum. FDA grants clearance to perspectum’s CoverScan – A new platform-based tool to assess multiple organs in one MRI scan. https://www.perspectum.com/our-company/news/fda-grants-clearance-to-perspectum-s-coverscan-a-new-platform-based-tool-to-assess-multiple-organs-in-one-mri-scan/.
  58. 58.
    Koios touts CPT codes for ultrasound AI decision-support software. AuntMinnie.com https://www.auntminnie.com/index.aspx?sec=log&itemID=138483 (2022).
  59. 59.
  60. 60.
    Unknown. American Medical Association Issues Technology-Specific CPT® Code for Milestone Scientific’s CompuFlo Epidural System. The Globe and Mail (2022).
  61. 61.
    Lassiter, R. PRESS RELEASE: CMS assigns new technology payment classification for Optellum’s Lung Cancer Prediction score. Optellum https://optellum.com/2022/06/press-release-cms-assigns-new-technology-payment-classification-for-optellums-lung-cancer-prediction-score/ (2022).
  62. 62.
  63. 63.
    Cross, P. B. MEDICAL POLICY – 10.01.533 Non-covered Experimental/Investigational Services. Preprint at https://www.premera.com/medicalpolicies/10.01.533.pdf (2023).
  64. 64.↵
    Beede, E. et al. A Human-Centered Evaluation of a Deep Learning System Deployed in Clinics for the Detection of Diabetic Retinopathy. in Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems 1–12 (Association for Computing Machinery, 2020).
  65. 65.↵
    Widner, K. et al. Lessons learned from translating AI from development to deployment in healthcare. Nat. Med. (2023) doi:10.1038/s41591-023-02293-9.
    OpenUrlCrossRef
  66. 66.↵
    Kiani, A. et al. Impact of a deep learning assistant on the histopathologic classification of liver cancer. NPJ Digit Med 3, 23 (2020).
  67. 67.↵
    Hendrix, N., Veenstra, D. L., Cheng, M., Anderson, N. C. & Verguet, S. Assessing the Economic Value of Clinical Artificial Intelligence: Challenges and Opportunities. Value Health 25, 331–339 (2022).
    OpenUrl
  68. 68.↵
    Ruamviboonsuk, P., Chantra, S., Seresirikachorn, K., Ruamviboonsuk, V. & Sangroongruangsri, S. Economic Evaluations of Artificial Intelligence in Ophthalmology. Asia Pac J Ophthalmol (Phila) 10, 307–316 (2021).
    OpenUrl
  69. 69.↵
    Pietris, J. et al. Health Economic Implications of Artificial Intelligence Implementation for Ophthalmology in Australia: A Systematic Review. Asia Pac J Ophthalmol (Phila) 11, 554– 562 (2022).
    OpenUrl
  70. 70.↵
    Seidenwurm, D. J. & Burleson, J. H. The medicare conversion factor. AJNR Am. J. Neuroradiol. 35, 242–243 (2014).
    OpenUrlFREE Full Text
  71. 71.↵
Back to top
Posted August 28, 2023.
Download PDF
Email
Characterizing the clinical adoption of medical AI through U.S. insurance claims
Kevin Wu, Eric Wu, Brandon Theodorou, Weixin Liang, Christina Mack, Lucas Glass, Jimeng Sun, James Zou
medRxiv 2023.08.26.23294629; doi: https://doi.org/10.1101/2023.08.26.23294629
Twitter logo Facebook logo LinkedIn logo Mendeley logo
Citation Tools

Subject Area