©2018. Published in Landslide, Vol. 10, No. 3, Janusry/February 2018, by the American Bar Association. Reproduced with permission. All rights reserved. This information or any portion thereof may not be copied or disseminated in any form or by any means or stored in an electronic database or retrieval system without the express written consent of the American Bar Association or the copyright holder.
Feature
Patenting the Future of Medicine: The Intersection of Patent Law and Artificial Intelligence in Medicine
Susan Y. Tull
Artificial intelligence (AI) is rapidly transforming the world of medicine, and the intellectual property directed to these inventions must keep pace. AI computers are diagnosing medical conditions and disorders at a rate equal to or better than their human peers, all while developing their own software code and algorithms to do so. These recent advances raise issues of patentability, inventorship, and ownership as machine-based learning evolves.
Artificial Intelligence
In the realm of AI, what was once science fiction is becoming fact. AI is “a field of science and engineering concerned with the computational understanding of what is commonly called intelligent behavior, and with the creation of artifacts that exhibit such behavior.”1 While the concept of AI is not new,2 the last three decades have marked a surge in the development of medical AI.3 AI techniques utilized in medicine include artificial neural networks, fuzzy expert systems, evolutionary computation, and hybrid intelligent systems.4
Artificial neural networks are formed of interconnected computer processors that perform parallel processing.5 These networks can learn from historical examples and known patterns and are thus used extensively in clinical diagnosis and image analysis.6 Neural networks have been used for diagnosing prostates as benign or malignant, cervical screening, and imaging analysis (including radiographs, ultrasounds, CTs, and MRIs), as well as for analyzing heart waveforms to diagnose conditions such as atrial fibrillation and ventricular arrhythmias.7
For example, researchers at Stanford University trained a neural network to classify skin lesions into either benign or malignant groupings based on known images and outcomes.8 The researchers started with an algorithm developed by Google to perform image recognition9 and then trained their neural network to recognize skin cancer using 129,450 clinical images of 2,032 different diseases.10 The neural network was then tested against board-certified dermatologists on clinical images that had been confirmed through biopsy.11 The AI performed on par with the certified dermatologists, demonstrating that the AI was capable of classifying skin cancer with the same level of competence as the trained dermatologists.12
As yet another example, digital health start-up Cognoa utilized machine learning AI to diagnose developmental delays in children based on information and videos sent in by parents.13 The machine learning model was trained using 891 autism cases and 75 nonautism controls to recognize autism behaviors in order to make a diagnosis.14 The AI was found to have a sensitivity of 89.9 percent in detecting autism, showing promise as a screening tool.15
Fuzzy logic AI rests on the premise that everything is a matter of degree, utilizing a continuous set membership from 0 to 1 instead of the traditional Boolean logic that uses sharp distinctions between 0 and 1 to define the algorithm.16 Because diseases, symptoms, and diagnoses are described in imprecise terms, fuzzy logic can recognize “partial truth logics,” beyond just the true and false values applied in traditional programming.17 Fuzzy logic AI has been applied to cancer diagnosis for lung cancer, acute leukemia, breast cancer, and pancreatic cancer.18
“Evolutionary computation is the general term for several computational techniques based on natural evolution process that imitates the mechanism of natural selection and survival of the fittest in solving real-world problems.”19 Genetic algorithms are the most widely used form of evolutionary computation in medicine, creating numerous solutions to a single problem, and then evolving those solutions from one generation to the next to arrive at the best solution.20 Evolutionary computation has been used in diagnosis, prognosis, imaging, and signal processing.21
Questions Surrounding Patenting AI in Medicine
As the use of AI in medicine becomes even more prevalent, the patent system must answer increasingly difficult questions regarding the protection afforded these technologies. The first, and perhaps most obvious, question is that of subject matter eligibility. With the Supreme Court decisions in Alice and Mayo, the hurdle to meet subject matter eligibility has grown ever higher.22
In Mayo, the Supreme Court found that claims directed to the relationships between concentrations of certain metabolites in the blood and the likelihood that a drug dosage will prove ineffective or cause harm were not subject matter eligible under § 101.23 According to the Court, “[b]eyond picking out the relevant audience, namely those who administer doses of thiopurine drugs, the claim simply tells doctors to: (1) measure (somehow) the current level of the relevant metabolite, (2) use particular (unpatentable) laws of nature (which the claim sets forth) to calculate the current toxicity/inefficacy limits, and (3) reconsider the drug dosage in light of the law.”24 Finding the claims lacked patent-eligible subject matter, the Court concluded that the claims provide “instructions [that] add nothing specific to the laws of nature other than what is well-understood, routine, conventional activity, previously engaged in by those in the field.”25
Alice addressed the holding of Mayo, enunciating a two-step test for subject matter patent eligibility: (1) determine whether the claims are directed to a patent-ineligible concept (laws of nature, abstract ideas, and natural phenomena); and (2) determine whether the claim’s elements, considered both individually and as an ordered combination, transform the nature of the claims into a patent-eligible application.26
These two recent Supreme Court cases present a hurdle that medical AI inventions will have to overcome in order to receive patent protection. Current AI medical device/system patents are directed to both the methods and apparatuses that perform the above-described analyses. Many AI medical patents are directed to the AI algorithms and the machines used to generate those algorithms.27
For example, Cognoa, described above, has filed a patent application directed to “methods and apparatus to determine developmental progress with artificial intelligence and user input.”28 The claims of the Cognoa application generally protect an apparatus “for evaluating a subject for risk of having a developmental disorder”29 Dependent claims further define the hardware used and processing steps followed in performing that evaluation.30 The broadest claims, however, do not require that level of detail. Claim 1, for example, is as follows:
An apparatus for evaluating a subject for risk of having a developmental disorder among two or more related developmental disorders, the apparatus comprising:
a processor comprising a tangible medium configured with instructions to,
present a question to the subject, the question configured to assess a clinical characteristic related to the two or more related developmental disorders,
receive an answer corresponding to the clinical characteristic of the subject related to the two or more related developmental disorders; and
determine, in response to the answer, whether the subject is at greater risk of a first developmental disorder or a second developmental disorder of the two or more related developmental disorders, with a sensitivity and specificity of at least 80%.31
These AI inventions in medicine are directed to diagnosis and prognosis, relating known images to new cases and extrapolating based on the similarities or differences between the two. In some instances, this is the same process followed by a doctor or medical expert, just with greater efficiency or accuracy. The steps for diagnosis struck down in Mayo echo the steps taken in many medical AI algorithms.
Indeed, the Federal Circuit has already found revolutionary diagnostic technology to be patent-ineligible subject matter under the Mayo/Alice framework. In Ariosa Diagnostics, Inc. v. Sequenom, Inc., the court concluded that a novel method of prenatal diagnosis of fetal DNA was not directed to patent-eligible subject matter, despite agreeing that the claimed method “reflects a significant human contribution . . . that revolutionized prenatal care.”32 The patent claims were generally directed to detecting the presence of cell-free fetal DNA in maternal plasma. Because the presence of cell-free fetal DNA was a natural phenomenon, the court turned to the second step in the Mayo/Alice framework—whether the claim contained an inventive concept sufficient to transform the naturally occurring phenomenon into patent-eligible subject matter.33 The court found that the second step was not met because the method steps “were well-understood, conventional, and routine,” despite acknowledging their breakthrough nature.34
Although the Supreme Court cautioned against construing the exclusionary principle of § 101 overbroadly, “lest it swallow all of patent law,”35 many believe it has done just that in the life sciences and medical spaces.36 The concurring opinion in Ariosa echoed these concerns, stating that “[b]ut for the sweeping language in the Supreme Court’s Mayo opinion, [there was] no reason, in policy or statute, why this breakthrough invention should be deemed patent ineligible.”37 The same reasoning could well curtail the patent protections afforded medical AI absent a change in Supreme Court precedent or statute. Until such a change occurs, AI inventors and owners must draft their patents with an eye to this two-step test, including features related to the AI in the claims, such as detailing the computing or mathematical techniques applied by the system or describing how the computer interacts with other components to drive the AI processing.
In addition to § 101 concerns, AI in medicine raises questions of inventorship and ownership in patent law. The US patent system only recognizes individuals as inventors,38 not companies39 or machines.40 But with AI, it may be the machine that is taking the inventive leap, not the human programmer. Recently, both Google and Facebook have seen AI develop its own language to perform the assigned tasks, eschewing known languages in favor of a more efficient means of communication.41 As the use of AI grows in medicine and the life sciences, it is more and more likely that the AI will be the entity taking the inventive step, drawing new conclusions between the observed and the unknown. Indeed, current AI systems develop their own code as a result of the system’s training.42 If that is the case, the United States Patent and Trademark Office (USPTO) and the courts will have to decide whether the current Patent Act encompasses computer-based inventors, and if not, who among the humans responsible for the AI should be considered an inventor.43 The list of possible human inventors includes the AI software and hardware developers, the medical professionals or experts who provided the data set with known values or otherwise provided input into the development of the AI, and/or those who reviewed the AI results and recognized that an invention had been made.
Similarly, AI may confuse the question of ownership for medical inventions generated by the AI itself. Patent ownership often turns on the question of inventorship44 (followed by assignment), and thus will be equally complicated when AI develops its own code and conceives its own inventions. Given that AI continues to advance after its initial programming, the question of inventorship and ownership may have to be answered years after the initial system programming. Development, assignment, and employment contracts will have to account for this possibility of continued and ongoing AI invention.
Finally, AI may impact one of the most basic tenets of patentability because the “person of ordinary skill in the art” may no longer be a person. Claim construction, novelty, and obviousness are all determined from the perspective of a person of ordinary skill in the art. With the rise of AI, the question must be answered, who (or what) is that “person”? Is it the AI itself that developed its own code and network? Is it the medical professional or experts against whom the AI is compared and who prepared the initial data that was used to program the AI? Is it the software and hardware developers who designed the original AI system? The answer to these questions could well determine the patentability of an invention as well as its scope.
Conclusion
As the use of AI in medicine grows ever more prevalent and sophisticated, these questions will have to be answered by Congress, the USPTO, and the courts. The most pressing question to be resolved is that of subject matter eligibility, so that innovation in this burgeoning field is not stifled. The long shadow cast by Mayo and Alice has already prevented the patenting of “revolutionary” diagnostic technology; patent drafters must take care to prevent the same from occurring with AI inventions until this question is answered. From there, the definitions of “person,” “inventor,” and “individual” will have to be revisited, so that our understanding of inventorship and ownership evolves with this rapidly advancing technology. Early recognition and resolution of these issues will allow patent law to keep pace with this new “rise of the machines.”
Endnotes
1. 1 Encyclopedia of Artificial Intelligence 54–57 (Stuart C. Shapiro ed., 2d ed. 1992).
2. Alan Turing is credited as the founder of modern computer science and AI, coining the “Turing test” to answer the question whether machines can think. A.M. Turing, Computing Machinery and Intelligence, 59 Mind 433, 433–34 (1950).
3. A.N. Ramesh et al., Artificial Intelligence in Medicine, 86 Annals Royal C. Surgeons Eng. 334 (2004).
4. Id. at 335.
6. Id.
7. Id. at 335–36.
8. Andre Esteva et al., Dermatologist-Level Classification of Skin Cancer with Deep Neural Networks, 542 Nature 115 (Feb. 2, 2017).
9. Taylor Kubota, Deep Learning Algorithm Does as Well as Dermatologists in Identifying Skin Cancer, Stan. News (Jan. 25, 2017), http://news.stanford.edu/2017/01/25/artificial-intelligence-used-identify-skin-cancer/.
10. Esteva et al., supra note 8.
11. Id.
12. Id.
13. Artificial Intelligence Is on the Precipice of Revolutionizing Medical Diagnosis, 13D Res. (Apr. 14, 2017), https://latest.13d.com/artificial-intelligence-is-on-the-precipice-of-revolutionizing-medical-diagnosis-be6427239f58; see also Cognoa, https://www.cognoa.com/ (last visited Nov. 4, 2017); Marlena Duda et al., Clinical Evaluation of a Novel and Mobile Autism Risk Assessment, 46 J. Autism & Developmental Disorders 1953, 1953 (2016).
14. Duda et al., supra note 13, at 1956.
15. Id. at 1953.
16. Ramesh et al., supra note 3, at 336.
17. Angela Torres & Juan J. Nieto, Fuzzy Logic in Medicine and Bioinformatics, J. Biomedicine & Biotechnology (2006).
18. Id.
19. Ramesh et al., supra note 3, at 336.
20. Id.
21. Id.
22. Alice Corp. Pty. Ltd. v. CLS Bank Int’l, 134 S. Ct. 2347 (2014); Mayo Collaborative Servs. v. Prometheus Labs., Inc., 566 U.S. 66 (2012).
23. Mayo, 566 U.S. at 69.
24. Id. at 82.
25. Id.
26. Alice, 134 S. Ct. at 2355.
27. See, e.g., Artificial Intelligence Sys. for Genetic Analysis, U.S. Patent No. 8,693,751 (filed Jan. 12, 2012); Clinical Decision-Making Artificial Intelligence Object Oriented Sys. & Method, U.S. Patent App. No. 14/324,396 (filed July 7, 2014); Integrated Med. Platform, U.S. Patent App. No. 15/039,419 (filed July 7, 2015); Local Diagnostic & Remote Learning Neural Networks for Med. Diagnosis, U.S. Patent App. No. 00/027,220 (filed Oct. 3, 2000).
28. U.S. Patent. App. No. 15/234,814 (filed Aug. 11, 2016).
29. Id.
30. Id.
31. Id.
32. 788 F.3d 1371, 1376, 1379 (Fed. Cir. 2015).
33. Id.
34. Id. at 1377.
35. Alice Corp. Pty. Ltd. v. CLS Bank Int’l, 134 S. Ct. 2347, 2354 (2014) (citing Mayo Collaborative Servs. v. Prometheus Labs., Inc., 566 U.S. 66, 71–72 (2012)).
36. Alexa Johnson, A Crisis of Patent Law and Medical Innovation: The Category of Diagnostic Claims in the Wake of Ariosa v. Sequenom, 27 Health Matrix 435 (2017); Patent Publius, Federal Circuit Threatens Innovation: Dissecting the Ariosa v. Sequenom Opinion, Ctr. for Protection Intell. Prop. (June 23, 2015), https://cpip.gmu.edu/2015/06/23/federal-circuit-threatens-innovation-dissecting-the-sequenom-v-ariosa-opinion/; Gene Quinn, Supreme Court Denies Cert. in Sequenom v. Ariosa Diagnostics, IPWatchdog (June 27, 2016), http://www.ipwatchdog.com/2016/06/27/70409/id=70409/.
37. Ariosa, 788 F.3d at 1381 (Linn, J., concurring).
38. 35 U.S.C. § 100(f).
39. New Idea Farm Equip. Corp. v. Sperry Corp., 916 F.2d 1561, 1566 n.4 (Fed. Cir. 1990).
40. Ben Hattenback & Joshua Glucoft, Patents in an Era of Infinite Monkeys and Artificial Intelligence, 19 Stan. Tech. L. Rev. 32, 46 (2015).
41. Tony Bradley, Facebook AI Creates Its Own Language in Creepy Preview of Our Potential Future, Forbes (July 31, 2017), https://www.forbes.com/sites/tonybradley/2017/07/31/facebook-ai-creates-its-own-language-in-creepy-preview-of-our-potential-future/#671deb97292c; Sam Wong, Google Translate AI Invents Its Own Language to Translate with, New Scientist (Nov. 30, 2016), https://www.newscientist.com/article/2114748-google-translate-ai-invents-its-own-language-to-translate-with/.
42. Toby Bond, How Artificial Intelligence Is Set to Disrupt Our Legal Framework for Intellectual Property Rights, IPWatchdog (June 18, 2017), http://www.ipwatchdog.com/2017/06/18/artificial-intelligence-disrupt-legal-framework-intellectual-property-rights/id=84319/.
43. Hattenback & Glucoft, supra note 40, at 46.
44. Id.