Healthcare is a large contributor to greenhouse gas (GHG) emissions around the world, given current power generation mix. Telemedicine, with its reduced travel for providers and patients, has been proposed to reduce emissions. Artificial intelligence (AI), and especially autonomous AI, where the medical decision is made without human oversight, has the potential to further reduce healthcare GHG emissions, but concerns have also been expressed about GHG emissions from digital technology, and AI training and inference. In a real-world example, we compared the marginal GHG contribution of an encounter performed by an autonomous AI to that of an in-person specialist encounter. Results show that an 80% reduction may be achievable, and we conclude that autonomous AI has the potential to reduce healthcare GHG emissions.
Introduction
There are efforts worldwide to reduce greenhouse gas (GHG) emissions, as they adversely affect the environment and human health1. Based on present, real world, mix of energy production methods (i.e., lacking renewables, nuclear, or other low or zero-carbon energy production), up to 10% of all annual GHG emissions may be due to healthcare processes and services2. The combined emissions of healthcare in the US, Canada, Australia, and the UK have been estimated to be ~7.5 × 1014 grams (g) of carbon dioxide equivalents (CO2-eq) annually3. These include energy use at the healthcare facility, including heating, ventilation and air conditioning, as well as its construction, provider, personnel and patient travel, training of physicians and other healthcare workers, healthcare technology and systems usage, pharmaceuticals and medical devices, and cleaning and disposables3.
Digital health technology, and specifically telemedicine or telehealth, has the potential to improve patient care, population health, reduce the per capita cost of healthcare, and improve the experience of providing care4. During the recent pandemic, these advantages became more apparent, as shown recently by Cortez et al.5. Importantly in this context, by reducing patient, physician and personnel travel, telemedicine also has the potential to substantially lower GHG emissions from travel energy use6,7,8,9,10,11.
Another digital health technology, artificial intelligence (AI), also has the potential to improve the quality, access, and experience of healthcare12,13. Such AI systems perform tasks intended to mimic human cognitive capabilities. These anthropomorphic AI systems are not explicitly programmed, and instead learn from data to perform highly cognitive tasks, such as those typically performed by trained healthcare professionals. As such, these AI systems have the potential to equalize access to healthcare for underserved populations, while improving care quality, at both the level of the individual patient and the population, at reduced cost for patient, payor, and society14,15. Initially, most healthcare AI systems were assistive, in other words, they assisted the physician or other provider in making medical decisions, but the ultimate authority for the clinical decision remained with the physician16. While such assistive AI may save time or effort on the part of the provider, it may also be additive to GHG emissions.
On the other hand, autonomous AI, where the medical decision is made by the AI system without human oversight, avoids the additive GHG emissions, and is infinitely scalable, with the potential to address healthcare disparities, improve patient outcomes, and patient and physician satisfaction17,18,19,20,21. The introduction of autonomous AI in the healthcare system has the potential to limit GHG emissions, and may even exceed the carbon emission reductions that resulted from telemedicine.
The goal of this study is to estimate the potential GHG emission reduction, if any, that can be achieved by the deployment of autonomous AI for point of care diagnostics.
Healthcare and digital technology GHG emission estimation
Multiple studies have analyzed GHG emissions, typically quantified as carbon dioxide release, resulting from specific healthcare processes, such as cataract surgery, anesthesia, and critical care as a means to determine where emissions can be reduced22,23,24. Generally, the largest fraction of emissions was due to energy use, including transportation of patients, providers, and staff to and from the facility10,11. A study measuring transportation associated GHG emissions at a large U.S. regional health system demonstrated significant reductions in GHG over all outpatient visits, as the use of telemedicine increased: from 8 × 103 g CO2-eq per visit to 4 × 103 g CO2-eq per visit25. The authors went further in accounting for the potential increase in emissions related to use of computer equipment and internet for telemedicine visits, estimating that there would be a net increase of 5 × 107 g CO2-eq for telemedicine encounters, compared to 8.7 × 109 g CO2-eq reduction from patient travel avoided25. Similarly, a telemedicine program in Catalan (Spain) showed a significant reduction in emissions over the course of a year by avoiding more than 9000 in-person visits, and reducing travel time, fuel use, and carbon emissions related to patient travel10. Modeling this out across the US, expansion of telemedicine programs nationwide has the potential to reduce GHG emissions by 2 × 1011 g CO2-eq annually11.
As mentioned above, while most outpatient GHG analyses, including the aforementioned studies, primarily consider patient travel-related emissions, they do not typically account for provider and personnel travel emissions, which is likely to be reduced9.
On the other hand, the use of digital technology itself, contributes to GHG emissions26. For example, a widely cited study based on data from 2015, estimated that digital technology is responsible for ~5 × 1017 g CO2-eq emissions27. The training of a single AI system may create up to 5 × 108 g CO2-eq emissions28, though inference—the calculation of a single output—may be only a fraction thereof29, and newer AI algorithms have shown improved GHG emission efficiency30.
Example of an autonomous AI: IDx-DR
The management of a person with diabetes requires regular diabetic eye exams to prevent blindness and vision loss31. The patient is typically referred by their primary care physician to an eye care provider for an in-person diabetic eye exam, generally requiring a separate clinic appointment. With implementation of autonomous AI at the point-of-care for the diabetic eye exam, screening for diabetic retinopathy can be performed during the same primary care encounter. Autonomous AI for the diabetic eye exam has been validated against patient outcome, with high sensitivity and specificity for diagnosing diabetic retinopathy at the point of care, and has been shown to be cost-effective from the patient perspective19,32. Developed under a rigorous ethical framework16, with FDA clearance of the first autonomous AI in 201833, specifically for the diabetic eye exam (IDx-DR, Digital Diagnostics, Coralville, IA)34, the Centers for Medicare and Medicaid Services (CMS) recently confirmed national coverage and reimbursement for such autonomous AI34,35,34.
In our analysis, we considered that the autonomous AI in this case may potentially reduce GHG emissions, by avoiding both those related to patient travel to the eye care provider, as well as GHG emissions generated by the healthcare facility where the in-person exam would take place.
Methodology
In our analysis, we compare the marginal contribution, i.e., the additional GHG emissions of a single additional diabetic eye exam, either through the traditional eye care provider, or the autonomous AI, at the margin. Thus, our comparison does not consider emissions related to the care episode that leads to the diabetic eye exam, emissions arising from the design and development (including training) of the AI system, or the emissions resulting from the education and training of the ophthalmologist. We compare the marginal contributions, to CO2-eq emissions only, of a diabetic eye exam for a single patient, performed at the point of care by the autonomous AI for the diabetic eye exam, to the estimated marginal emissions if this diabetic eye exam is instead performed by an ophthalmologist after referral. We use the Environmental Protection Agency (EPA) in the U.S., conversion table, based on the present electricity generation mix, to convert a kWh of energy results in carbon emissions of 7.09 × 10−4 metric tons CO2-eq, or 0.709 g CO2-eq per Wh of energy36.
We set the marginal GHG emissions for a diabetic eye exam by an ophthalmologist, after referral, to that estimated by Dacones et al. for an in-person outpatient visit, i.e., ~8 × 103 g CO2-eq per patient encounter25.
In our comparison, when calculating the marginal contribution of the autonomous AI we account for the fact that a ‘disease present’ output will lead to an in-person referral with an eye care provider—in our example about 20% of all cases will lead to a referral, with a concomitant increase in GHG emissions34.
Results
This autonomous AI systems’ diagnostic computations, from interactive image capture, through AI inference—the AI is completely locked and deterministic after FDA authorization—and diagnostic output to the electronic health record and provider, occupies about 144 × 109 CPU cycles. As an example, computations run for 10 s (0.003 h) on 4 parallelized Nvidia Tesla K80 processors37. AI power consumption estimates vary in the literature. For machine learning inference heavy applications, estimates range, for a single CPU, between 2.1 and 25 Wh. Using Sommer et al.’s estimate of 2.1 Wh38, and Bergman et al.’s estimate of 3.3 KWh/24 h39, respectively, we estimate power usage for a single autonomous AI diagnostic exam to be between 0.03 and 0.3 Wh.
Under these assumptions, each autonomous AI diagnostic exam generates 0.02–0.2 g CO2-eq emissions per exam. As mentioned above, if we account for 20% of patients receiving a diagnostic output requiring an additional in-person diabetic eye exam34, the total emissions are estimated to be 1.6 × 103 g CO2-eq leading to an 80% reduction if patients get their initial diabetic eye exam using autonomous AI as opposed to in-person diabetic eye exams.
Discussion
The use of a specific autonomous AI can lead to an 80% reduction in GHG emissions in the present state. We only compared the marginal GHG emissions generated by one additional diabetic eye exam, and did not consider training of either the autonomous AI system or the eye care provider, and other personnel involved in the in-person visit. If we were to do so, the comparison would be more stark: a retina specialist performing the in-person diabetic eye exam typically undergoes 8–10 years of training. Annual CO2-eq emissions per year of higher education per student for the US are estimated at 5 × 106 g CO2-eq, or a total of 5 × 107 g CO2-eq for the entire training40. Such a specialist may perform 100,000 diabetic eye exams in their career16, so that training may contribute 0.5 × 103 g CO2-eq to GHG emissions per exam. Training the AI may take 5 × 108 g CO2-eq, as mentioned above. Such an autonomous AI system has almost unlimited scalability, so that it may perform diabetic eye exams for 30 × 106 people with diabetes. Assuming such an autonomous AI functions without retraining for 5 years, the GHG emissions for the AI system training amounts to only 3 g CO2-eq/exam, in this worst-case scenario, <1% of the training for the human specialist. However, because of the many additional assumptions necessary in accounting for training, we only considered any marginal reduction in greenhouse emissions—and assume GHG emissions for eye care provider education and training and autonomous AI training are accounted for.
Limitations to this study include that we used 2021 estimates for all variables. These are subject to change, for example the production mix of power generation is expected to change over the coming years due to various initiatives, and thereby the GHG emissions per kWh will also change. The carbon efficiency of computer hardware is also changing, and thus the power consumption per inference, all else being equal, will change as well.
Based on the above assumptions and limitations, and extrapolation of currently available data estimates from 2021, autonomous AI has the potential to substantially lower healthcare GHG emissions, and thereby compensate for increasing carbon emissions attributed to information technology41. As use of autonomous AI systems expand in the healthcare industry, measurement of real-world carbon emissions attributed to these systems in comparison to usual care will help elucidate the potential contributions of autonomous AI in reducing healthcare emissions.
Data availability
Data utilized in this manuscript are available without restriction from their respective references.
References
Milner, J. et al. Health benefits of policies to reduce carbon emissions. BMJ 368, l6758 (2020).
Eckelman, M. J. & Sherman, J. Environmental impacts of the U.S. health care system and effects on public health. PLoS ONE 11, e0157014 (2016).
Sherman, J. D., MacNeill, A. & Thiel, C. Reducing pollution from the health care industry. JAMA 322, 1043–1044 (2019).
Tuckson, R. V., Edmunds, M. & Hodgkins, M. L. Telehealth. N. Engl. J. Med. 377, 1585–1592 (2017).
Cortez, C., Mansour, O., Qato, D. M., Stafford, R. S. & Alexander, G. C. Changes in short-term, long-term, and preventive care delivery in US office-based and telemedicine visits during the COVID-19 pandemic. JAMA Health Forum 2, e211529–e211529 (2021).
Yellowlees, P. M., Chorba, K., Burke Parish, M., Wynn-Jones, H. & Nafiz, N. Telemedicine can make healthcare greener. Telemed. J. E Health 16, 229–232 (2010).
Bozoudis, V. & Sebos, I. The carbon footprint of transport activities of the 401 military general hospital of Athens. Environ. Modeling Assess. 26, 155–162 (2021).
Croghan, S. M. et al. Time, cost and carbon-efficiency: a silver lining of COVID era virtual urology clinics? Ann. R. Coll. Surg. Engl. 103, 599–603 (2021).
Wootton, R., Bahaadinbeigy, K. & Hailey, D. Estimating travel reduction associated with the use of telemedicine by patients and healthcare professionals: proposal for quantitative synthesis in a systematic review. BMC Health Serv. Res. 11, 185 (2011).
Vidal-Alaball, J., Franch-Parella, J., Lopez Segui, F., Garcia Cuyas, F. & Mendioroz Pena, J. Impact of a telemedicine program on the reduction in the emission of atmospheric pollutants and journeys by road. Int. J. Environ. Res. Public Health. 16, https://doi.org/10.3390/ijerph16224366 (2019).
Whetten, J., Montoya, J. & Yonas, H. ACCESS to better health and clear skies: telemedicine and greenhouse gas reduction. Telemed. J. E Health 25, 960–965 (2019).
Matheny, M. E., Whicher, D. & Thadaney Israni, S. Artificial intelligence in health care: a report from the National Academy of Medicine. JAMA 323, 509–510 (2020).
Rajkomar, A., Dean, J. & Kohane, I. Machine learning in medicine. N. Engl. J. Med. 380, 1347–1358 (2019).
Emanuel, E. J. & Wachter, R. M. Artificial intelligence in health care: will the value match the hype? JAMA 321, 2281–2282 (2019).
Preston, R. Autonomous AI in Action. https://www.forbes.com/sites/oraclecloud/2020/01/16/autonomous-in-action-self-driving-cars-get-all-the-publicity-but-other-industries-are-already-getting-exceptional-value-from-ai-based-systems/#1ecc65d86e94 (2020).
Abramoff, M. D. et al. Foundational considerations for artificial intelligence using ophthalmic images. Ophthalmology https://doi.org/10.1016/j.ophtha.2021.08.023 (2021).
Thomas, C. G. et al. Racial/ethnic disparities and barriers to diabetic retinopathy screening in youths. JAMA Ophthalmol. https://doi.org/10.1001/jamaophthalmol.2021.1551 (2021).
Wolf, R. M. et al. The SEE study: safety, efficacy, and equity of implementing autonomous artificial intelligence for diagnosing diabetic retinopathy in youth. Diabetes Care. 44, 781–787 (2021).
Wolf, R. M., Channa, R., Abramoff, M. D. & Lehmann, H. P. Cost-effectiveness of autonomous point-of-care diabetic retinopathy screening for pediatric patients with diabetes. JAMA Ophthalmol. 138, 1063–1069 (2020).
Rajkomar, A., Hardt, M., Howell, M. D., Corrado, G. & Chin, M. H. Ensuring fairness in machine learning to advance health equity. Ann. Intern Med. 169, 866–872 (2018).
Goodman, S. N., Goel, S. & Cullen, M. R. Machine learning, health disparities, and causal reasoning. Ann. Intern Med. 169, 883–884 (2018).
Goel, H. et al. Improving productivity, costs and environmental impact in International Eye Health Services: using the ‘Eyefficiency’ cataract surgical services auditing tool to assess the value of cataract surgical services. BMJ Open Ophthalmol. 6, e000642 (2021).
McGain, F. et al. Carbon footprint of general, regional, and combined anesthesia for total knee replacements. Anesthesiology https://doi.org/10.1097/ALN.0000000000003967 (2021).
Purohit, A., Smith, J. & Hibble, A. Does telemedicine reduce the carbon footprint of healthcare? A systematic review. Future Health. J. 8, e85–e91 (2021).
Dacones, I., Cave, C., Furie, G. L., Ogden, C. A. & Slutzman, J. E. Patient transport greenhouse gas emissions from outpatient care at an integrated health care system in the Northwestern United States, 2015–2020. J. Clim. Change Health 3, 100024 (2021).
Dhar, P. The carbon impact of artificial intelligence. Nat. Mach. Intell. 2, 423–425 (2020).
Malmodin, J. & Lundén, D. The energy and carbon footprint of the global ICT and E&M sectors 2010–2015. Sustainability 10, 3027 (2018).
Cowls, J., Tsamados, A., Taddeo, M. & Floridi, L. The AI gambit: leveraging artificial intelligence to combat climate change—opportunities, challenges, and recommendations. AI & SOCIETY https://doi.org/10.1007/s00146-021-01294-x (2021).
Thompson, N. C., Greenewald, K., Lee, K. & Manso, G. F. The computational limits of deep learning. Preprint at arXiv https://arxiv.org/abs/2007.05558 (2020).
García-Martín, E., Rodrigues, C. F., Riley, G. & Grahn, H. Estimation of energy consumption in machine learning. J. Parallel Distrib. Comput. 134, 75–88 (2019).
American Diabetes A. 11. Microvascular complications and foot care: standards of medical care in diabetes-2020. Diabetes Care. 43, S135–S151 (2020).
Scotland, G. S. et al. Cost-effectiveness of implementing automated grading within the national screening programme for diabetic retinopathy in Scotland. Br. J. Ophthalmol. 91, 1518–1523 (2007).
U.S. Food & Drug Administration (FDA). FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems. https://www.fda.gov/newsevents/newsroom/pressannouncements/ucm604357.htm (2018).
Abràmoff, M. D., Lavin, P. T., Birch, M., Shah, N. & Folk, J. C. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. Nat. Digital Med. 1, 39 (2018).
Centers for Medicare & Medicaid Services. e. Proposal to Establish Values for Remote Retinal Imaging (CPT code 92229) (Pages 56ff). https://public-inspection.federalregister.gov/2021-14973.pdf (2021).
U.S. Environmental Protection Agency (EPA). AVERT, U.S. national weighted average CO2 marginal emission rate, year 2019 data. https://www.epa.gov/energy/greenhouse-gases-equivalencies-calculator-calculations-and-references (2020).
NVIDIA. TESLA K80 GPU ACCELERATOR Board Specification. https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-product-literature/Tesla-K80-BoardSpec-07317-001-v05.pdf (2015).
Sommer, E. et al. Cloud jewels: estimating kWh in the cloud. https://codeascraft.com/2020/04/23/cloud-jewels-estimating-kwh-in-the-cloud/ (2020).
Bergman, S. How can i calculate CO2eq emissions for my Azure VM. https://devblogs.microsoft.com/sustainable-software/how-can-i-calculate-co2eq-emissions-for-my-azure-vm/ (2021).
Sinha, P., Schew, W. A., Sawant, A., Kolwaite, K. J. & Strode, S. A. Greenhouse gas emissions from U.S. institutions of higher education. J. Air Waste Manag Assoc. 60, 568–573 (2010).
Freitag, C. et al. The real climate and transformative impact of ICT: a critique of estimates, trends, and regulations. Patterns 2, 100340 (2021).
Author information
Authors and Affiliations
Contributions
Drs. Abramoff, Wolf, Channa, and Lehmann conceptualized the manuscript and designed the work. Dr. Abramoff, Mr. Tava, and Mr. Clarida acquired and analyzed the data. Dr. Abramoff and Dr. Wolf drafted the manuscript. All authors critically reviewed the manuscript and made edits to reach the final submitted and approved version.
Corresponding author
Ethics declarations
Competing interests
Dr. Abramoff reports the following conflicts of interest: Employee, Investor, Director of Digital Diagnostics. Dr. Abramoff has patents and patent applications assigned to the University of Iowa and Digital Diagnostics that are relevant to the subject matter of this manuscript. Mr. Tava and Mr. Clarida are employees of Digital Diagnostics. Dr. Wolf, Dr. Channa, and Dr. Lehmann have no disclosures relevant to this manuscript.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Wolf, R.M., Abramoff, M.D., Channa, R. et al. Potential reduction in healthcare carbon footprint by autonomous artificial intelligence. npj Digit. Med. 5, 62 (2022). https://doi.org/10.1038/s41746-022-00605-w
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41746-022-00605-w
This article is cited by
-
Considerations for integrating wearables into the everyday healthcare practice
npj Digital Medicine (2023)