Can machine learning take the pain -and cost- out of how we pay for healthcare?

Can machine learning transform the onerous and costly process of medical billing into one that reduces healthcare costs and improves healthcare quality?

Despite leading the world in medical innovation, it’s no secret that Americans are paying more for less when it comes to healthcare [[1]]. In 2013, 17.1 percent of U.S. GDP was spent on a healthcare, far greater than any other high-income country, many of whom offer universal healthcare [[2]]. While there are many reasons for why healthcare is so expensive in the United States, one major driver of spending is the high administrative cost associated with the processing and payment of medical claims submitted by physicians to insurers [[3]].

Long the bane of many a physician’s existence, the claims process begins with the submission of an insurer-mandated prior authorization form via phone or fax in order to obtain pre-approval for tests, medicines, and other necessary clinical services. The complexity of this labor-intensive process poses a tremendous burden on physicians, delays the delivery of necessary care for patients, and drives up what are already sky-high administrative costs [[4]]. While insurers believe this gate-keeping function reduces inappropriate healthcare spending and enhances patient safety by reducing unnecessary procedures, a 2012 study by the American Medical Association called into question the utility of spending an estimated $728 million dollars per year on prior authorizations since nearly all claims are approved [[5]]. As others sectors in the healthcare industry reap the benefits of era of “big data,” there is growing support amongst both payers and providers for a tool than can ease the prior authorization process with the hope of reducing the administrative burden involved while also improving the quality of care patients receive.

In order to fill this gap, physicians and insurers alike have turned to machine learning, the science of building algorithms that can complete tasks without being explicitly taught, in order to automate the process of prior authorization. A key example of how machine learning can be harnessed to reduce the hassle of prior authorization can be seen in Evicore’s “intellipath” system. Recently acquired by the nation’s largest pharmacy benefits manager for 3.6 billion dollars[6], Evicore’s system integrates directly into the electronic health record system where it automatically submits pre-populated prior authorization forms to any insurer the physician works with, thereby greatly reducing the need for manual data entry [[7]]. According to a recent study by Accenture, streamlining routine manual prior authorizations and other similar core administrative processes has the potential to save U.S. health insurers up to 7 billion dollars and benefit all stake holders in the healthcare system [[8]]. From the billing departments of major academic medical centers to the smallest physician practices and every payer they interact with, these tools can have an immediate impact on operating income and reduce the cost associated with the processing and payment of medical claims. For physicians, it means less time filling out paperwork and more time focusing on what matters, namely caring for patients. Taken together, these captured efficiencies can align financial and altruistic incentives by both reducing healthcare spending associated with administrative overhead while also improving the quality of care.

As more payers and providers move to adopt machine learning as a tool to automate the prior authorization process, new challenges to keeping the cost of healthcare down and ensuring that patients are receiving appropriate care will emerge. For instance, it is plausible that removing the gatekeeper function and making it easier to submit prior authorizations will actually lead to an increase in utilization and healthcare spending. Moreover, the success of an automated system depends on the integrity of the data set the algorithm is learning from. Due to limitations of the current state electronic health records, many physicians view these systems not as accurate repositories of patient information but merely as vehicles for billing. As a result, critical information that may be necessary for determining the clinical appropriateness of a test or procedure but not necessary for billing is often left out of the record. Without this information, machine learning may end up being another system where garbage in produces garbage out. That being said, further investment in machine learning as a tool to automate core administrative functions associated with medical billing can create significant value for our healthcare system by enabling physicians to provide better care to more people at a reduced cost.

[[1]] Tyler Cowen, “Poor U.S. Scores in Healthcare Don’t Measure Nobels and Innovation,” The New York Times, October 5, 2006, https://www.nytimes.com/2006/10/05/business/05scene.html?module=inline, accessed November 2018.

[[2]] D. Squires and C. Anderson, “U.S. Health Care from a Global Perspective: Spending, Use of Services, Prices, and Health in 13 Countries,” The Commonwealth Fund, October 2015.

[[3]] Thomas Sullivan, “AMA’s National Health Insurer Report Card – $12 Billion Could be Saved Through Increased Claims Automation,” Policy and Medicine, July 18, 2013, https://www.policymed.com/2013/07/amas-national-health-insurer-report-card-12-billion-could-be-saved-through-increased-claims-automation.html, accessed November 2018.

[[4]] Austin Frakt, “The Astonishingly High Administrative Costs of U.S. Health Care,” The New York Times, July 16, 2018, https://www.nytimes.com/2018/07/16/upshot/costs-health-care-us.html, accessed November 2018.

[[5]] “Putting a price on the hassle of preauthorization,” American Medical News, January 21, 2013,  https://amednews.com/article/20130121/business/130129986/6/, accessed November 2018.

[[6]] Stephanie Baum, “As Express Scripts pays $3.6B for eviCore Healthcare, did Amazon make the PBM blink?,” MedCity News, October 10, 2017, https://medcitynews.com/2017/10/express-scripts-pays-3-6b-evicore-healthcare-amazon-make-pbm-blink/?rf=1, accessed November 2018.

[[7]] Evicore, “Provider Solutions: Overview,” https://www.evicore.com/solution/pages/provider.aspx, accessed November 2018.

[[8]] Rebecca Pifer, “AI can save US insurers $7B in admin costs, Accenture says,” Healthcare Dive, August 9, 2018, https://www.healthcaredive.com/news/ai-can-save-us-insurers-7b-in-admin-costs-accenture-says/529578/, accessed November 2018.

Previous:

Farmers Business Network: Democratizing agriculture using machine learning

Next:

MACHINE LEARNING IN RETAIL: AN EDITED APPROACH

7 thoughts on “Can machine learning take the pain -and cost- out of how we pay for healthcare?

  1. Excellent article on a very serious issue. This technology is certainly impressive and its need apparent. I worry, though, about the challenge of automating only one link in a long chain of processes across the whole system. Specifically, if the automated system on the provider side is still sending its output to a manual system on the payer side, could this just result in more confusion and back-and-forth between the two parties, perhaps ultimately requiring escalation to human intervention on the provider side to resolve? That is, while I think it’s clear that an automated->automated system would be strictly better than a manual->manual system, I wonder where the automated->manual system falls on this spectrum. It could be the case that true efficiency gains are only achieved when both sides of the process are automated.

  2. Very interesting topic – thank you for posting. Healthcare indeed appears to be an industry where technology and machine learning can have outsized effects given both the lack of technology implemented across stakeholders and the various problems facing the industry (e.g., manual PA processes, lack of interoperability). This post made me think about our TOM discussions around resource allocation, in addition to the FRC Springfield Hospital case, and the importance of allocating resources (extrapolated as applying to both labor and machines) efficiently, and not having high-cost labor (e.g., physicians) completing PA work that could be more efficiently completed using other resources (e.g., Evicore). I agree the comment above – it will be important for companies to assess the cost/time savings of implementing this tool, and ensure that doing so will not create unanticipated work (e.g., due to lack of interoperability, or additional manual labor needed to “fix” a manual -> automated flow).

  3. Good article on the benefits of machine learning in the healthcare sector, in terms of cost savings. Healthcare costs have escalated in the United States to extent that is unreasonable. I agree in principle that automating a step in the communication chain between providers and insurers can reduce the burden of administrative tasks, and consequently total costs. However, a major issue with machine learning in healthcare is its practicality and adaptability given the volume and unique nature of patient presentations. There some much room to mistakes with patient filing, which is why good case managers are in such high demand. I can only imagine the how costly and time-consuming it would be to course correct a mistake as a result of machine learning system such as Evicore. As Matt B pointed out above, a provider would need to see data around the efficiency and error rate between a automated –> automated chain before making the decision to implement a system like Evicore.

  4. Really interesting article – thank you for writing! One question that came to mind as I read this article was whether or not big data can truly bring down cost to a level that is comparable to other countries which have a single payer (typically government-run) health insurance. I’d be very interested to see if machine-learning and big-data algorithms see greater efficiency gains in countries that have less layers of bureaucracy and fewer diffuse actors in healthcare administration.

  5. Very interesting application of machine learning. For sure it could have a big impact in all of our lives. Besides the application that you mention for the United States and its potential to reduces costs I believe it could be easily applied to fraud detection. If you are able to process all the information that is going around you can monitor drug consumption and evaluate if patients or doctors are taking advantage of the different insurances and for example doing arbitrage with prices. If you get all the information together you can detect this and penalize whoever is doing so.

  6. Really interesting read!
    One question that comes to my mind regarding the general application of machine learning in healthcare systems is the rate of adoption among physicians as I have consistently seen that to be the biggest challenge. Do you have any thoughts on how can we speed up the rate of adoption, some of the ideas that come to my mind include – Collaborating with medical schools so that we can make the future generation much more open to the ideas of EHR and active involvement with data to make decisions.

  7. Very interesting article, thanks for sharing. I laud both sides in the medical community for chasing technology like this. To the point of generalizability, I think this is a step in the right direction. Although it appears this technology is limited to the optimization of the prior authorization process, which is no small feat, I think it can be applied to other costly parts of the healthcare system. Examples include the process of prescribing medicine to patients, the hurdles pharmacists’ have to jump through to contact and obtain approval from insurance companies for drugs, and the entry that goes into updating patient records. For most of the time that a patient spends in the doctor’s office, the healthcare provider is typing into a computer, thereby reducing the quality and quantity of care given to the patient themselves – perhaps similar technology could automate that process through natural language processing and even help providers surmise/analyze the data. The possibilities are endless and, more importantly, the need for innovation is greater than ever.

Leave a comment