PATENTS Mapping the Patent Landscape of Medical Machine Learning Mateo Aboy, W. Nicholson Price II, Seth Raker Patent office data show robust and rising patenting of AI inventions in the medical field, contrary to fears that medical machine learning patents might be largely unavailable due to post-Alice/Mayo challenges to their subject-matter eligibility and over a decade of limited MML patenting activity in the US and Europe. Artificial intelligence (AI) is rapidly entering the field of medicine, but the role of patents in this process remains relatively opaque. Regulators repor t hundreds of machine learning (ML) medical devices that have passed regulatory oversight1, including systems involved in r a d i o l o g y , c a r d i o l o g y , ophthalmology, and many other fields. Major hospitals and academic m e d i c a l s y s t e m s h ave b o t h developed and deployed AI/ML systems, and some AI tools have been embedded in electronic health records used by health systems covering millions of patients. Nevertheless, despite this wave of innovation in medical machine learning (MML), the influence of patents on that process has only b e e n s k e t c h e d r a t h e r t h a n interrogated in detail.   Absent a detailed picture of patenting in the field of medical AI, legislators, patent offices, scholars, and other stakeholders are operating in the dark when determining how to use various policy levers to shape the innovative landscape for medical AI. How common are patents on medical AI inventions, and at what speed are they being issued? Who is obtaining patents on medical AI—is it a diverse group of inventors, or is it dominated by large entities in either the medical device space or the software space?  What kind of claims are used to protect medical AI inventions? Answers to these more basic questions are essential before tackling more complex policy issues, such as whether patents provide adequate or necessary incentives to overcome regulatory hurdles, the extent to which patents are fulfilling their disclosure function and enabling cumulative innovation, and how pa ten t s shape the availability of inventions to a broad set of potential users.   Many scholars, including one of us, have been ske p t i ca l o f the availability and strength of patents for a r t i f i c i a l in te l l i g ence in medicine2,3. In the United States, p a t e n t a b l e s u b j e c t m a t t e r jurisprudence after the 2012 Mayo v. Prometheus, 2013 Myriad v. AMP and 2014 Alice v. CLS Bank cases raised doctrinal hurdles to obtaining patents for AI inventions that seem like they could easily rely on abstract ideas or laws of nature4. Adequate disclosure of AI systems also creates potential challenges for patentability, since AI system complexity or opacity could theoretically lead to inadequately descriptive functional d i s c l o s u r e s o r i n s u f f i c i e n t enablement.   Empirical observations of patenting in medical AI can begin to clarify the role of patents in the field. Although the optimal level of patenting in any f ie ld is largely unknowable5, significant rates of patent issuance on medical AI inventions would suggest that patentability concerns are at least not blocking patenting activity. Similarly, if patenting has remained stable or at a low level while the field has seen increasing development and use, that would suggest that patents are not a particularly important incentive compared to, for instance, secrecy, proprietary training data, first-mover advantage, or forms of technological lock-in. If, on the other hand, patents are increasingly issued in the field to a range of inventors, they are likely to be playing some broader role in creating incentives for medical AI invention.   Here, we begin to empirically detail the role of patents in medical AI using global patent data but focusing principally on data from the United States Patent and Trademark Office (PTO) and the European Patent Office (EPO), which grants patents Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY1 _______________________________________ Mateo Aboy is with Centre for Law, Medicine, and Life Sciences (LML), Faculty of Law, University of Cambridge, Cambridge, UK; Nicholson Price and Seth Raker are with U n i v e r s i t y o f M i c h i g a n L a w S c h o o l . C o r r e s p o n d i n g A u t h o r E - m a i l : ma608@cam.ac.uk mailto:ma608@cam.ac.uk cove r ing 38 Member S t a t e s including the UK. We describe general patterns of patenting, the entities receiving most patents, what sorts of inventions are receiving patents, and how these inventions are claimed. Research Questions In this paper we examine the emerging patent landscape of MML technologies. Specifically, we address the following research questions: 1) What have been the patenting trends over the last 20 years for AI/ MML?; How many of these MML patents are granted per year?; What has been their growth rate?, and What is their allowance rate?; 2) Which organizations are leading the patenting activity for MML?; and What is the patent office of choice?; 3) What types of claim strategies and formulations are being used to protect these inventions and what is their relative prevalence?; and 4) What medical applications and input signals are MML patents focusing on? Search Strategy & Landscaping We developed a search strategy designed to answer the above questions. The strategy follows the r ecommenda t ions on pa ten t landscaping for l i fe sciences innovation6, as well as the checklist o f i n f o r m a t i o n f o r p a t e n t landscapes recommended to ensure quality and transparency,7 but narrowed to answer the specific research questions of this study as opposed to providing a general p a t e n t l a n d s c a p e . S i m i l a r methodologies have been used to analyze the patent landscape of quantum technologies8, gene patents9,10, drug repurposing11 and other medical-related inventions12. Table 1 provides a summary of the search strategy and results. This strategy is designed to identify AI/ ML patents from 2001 to 2021 focused on medical applications. The search strategy ranges from high sensitivity (Table 1 Search ID: S1) to high specificity to minimize false positives (Table 1 Search ID: S7). In particular, the search strategy i d en t i f i e s ( a ) g l oba l p a t en t documents related to AI/ML published in the last 20 years (S1); (b) the subset of US and EPO patent applications and granted patents focused on AI/ML (S2); and (c) the subset of these patents related to medical applications (S3). While searches S1-S3 identify patents focused on AI/ML by searching for the AI terms in the title, abstract, or claims (TAC), searches S4-S6 are designed to identify patents containing claims directed to medical AI/ML. The S1 search s t ra teg y i s optimized for high sensitivity (but low specificity) and provides an estimated upper bound of the total number of core AI/ML patents over the last 20 years. Specificity is substantially improved by restricting the search to classes focused on medical patents (Cooperative Patent Class-CPC: A61), resulting in 11,891 patent applications and 3,755 granted patents (S3). Since we are particularly interested in investigating claim formulations that are being used to protect MML inventions, searches S4-S7 identify these MML patents with increasing degrees of specificity by requiring the claims to contain limitations directed to core AI/ML keywords (i.e., “artificial intelligence,” “AI,” “mach ine l ea r n ing ,” “neura l network,” or “deep learning”) and fu r the r na r row ing the CPC subclasses. For S4-S5, the AI/ML limitation can be in any of the claims, but patent searches S6-S7 require the AI/ML limitation to be present in the independent claims of the patent. Thus, in S5-S7, the claims with the broadest scope— presumably the core of the claimed invention—contain the limitation directed to AI/ML for a medical application. The search achieves a higher specificity by further narrowing the results to specific CPC sub-classes. T h i s l e v e r a g e s t h e m a n u a l classification conducted by USPTO and EPO patent exper ts to categorize each patent application and granted patent in the relevant CPC class, effectively combining the resu l t s o f au tomat ic sea rch algorithms with manual expert reviews. The CPC is an extension of t h e I n t e r n a t i o n a l P a t e n t C l a s s i f i c a t i o n ( I P C ) s y s t e m a d m i n i s t e r e d by t h e Wo r l d Intellectual Property Organization (WIPO). The CPC is jointly managed by the USPTO and EPO to achieve harmonization across patent offices and improve patent searching. Our search strategies leverage the CPC classes to identify the AI/ML patents in medical science (S3-S4; CPC A61) or with medical applications, including diagnosis, surgery, and identification (S5-S6; CPC A61B). Finally, for our expert review, S7 focuses on patents within CPC A61B5/7264-7. This class contains patents directed to medical applications “using neural networks, statistical classifiers, expert systems or fuzzy systems” based on physiological signals or patient data and “involving training the class- action device,” and thus represents a set of patents that USPTO or EPO examiners have manually classified as AI/ML medical applications. Expert Review & Claim Analysis The granted patents containing independent claims directed to machine learning for diagnostic or identification purposes (S7) were reviewed by one of the authors (SR) to further analyze and classify the patents. The expert review was used to (a) determine the specificity of the search algorithm and (b) manually classify the S7 patents b a s e d o n t h e i r c l a i m s ( s e e Supplementary Information). Each patent was reviewed to confirm that the AI/ML limitation was contained within independent claims. The independent claims were Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY2 analyzed to classify further the patents based on the nature of the medical application and the type of input data used by the AI/ML method. Specifically, based on the expert review of the broadest claims, the S7 MML patents were c l a s s i f i ed a s d i r ec ted to 1 ) technology improvements, 2) measurement, 3) analysis, 4) decision s u p p o r t , 5 ) d e t e c t i o n , 6 ) classification, 6) diagnosis, 7) prognosis, 8) monitoring, and 9) treatment (see Supplementary Information). The patents were further classified according to the input signal needed b y t h e A I / M L m e t h o d : 1 ) physiologic signals or patient data, 2) images, and 3) video. Within these three broad classes, patents were classified according to the specific input signal used (e.g., ECG, EEG, E M G , S p O 2 , M R I ) . T h e Supplementar y Infor mation includes the patent claim details and the corresponding results of the classification by input signal, general application, and medical application. The expert review indicates that the search algorithm has an estimated specificity greater than 99.76%, as no false positives were found in the sample of 421 patents reviewed. In all instances, the independent claims had AI/ML limitations. Landscape Results & Discussion The search strategy output was further analyzed using patent analytics to answer the research questions. 1)What have been the patenting trends over the last 20 years for AI/MML?; How many of these MML patents are granted per year?; What has been their growth rate?, and What is their allowance rate?; Our search strategy (S4) yielded 10 ,942 USPTO/EPO pa ten t applications with claims including AI/ML limitations and classified as having a medical application (CPC: A61). Of these, 3,479 have been granted and published as issued patents over the last 20 years. Figure 1 shows the annual pa ten t ing ac t iv i t y fo r MML technologies at the USPTO and EPO and the corresponding legal status of the patent documents published in a given year (S3). The figure shows 1) the MML patents granted in the particular year, 2) the r e j e c t e d / a b a n d o n e d p a t e n t applications, 3) the previously granted patents that expired that given year, and 4) the pending patent applications. Our S3 results shown in Figure 1 indicate that there has been a substantial upward trend in MML patents since 2014. The number of MML patent documents (S3) increased from 264 in 2013 to 2,661 in 2021, corresponding to a compound annual growth rate (CAGR) of 33.48% in the last 8 years. Similarly, the number of granted patents with MML claims (S4) increased from 55 in 2013 to 745 in 2021 (CAGR=38.51%). The exponential nature of this growth can be seen in Figure 1 and later graphs of patent grants. That said, this growth should not be taken for granted. As seen in Figure 1, our results also show 10 years of limited growth from 2004 to 2013. The relative proportion of granted applications to the total number of applications (for years with no pending applications) indicates that the patent allowance rate has increased from 48% in 2007 to 64% in 2012. As seen in Figure 1, there is a significant number of MML pending applications (blue) for patent applications filed after 2012. Consequently, allowance rates can only be estimated. That said, for those applications for which there is a final disposition, the trend of allowance rates higher than the EPO overall grant rate of 49% continues. While AI/ML applicants are often impacted by 35 USC 101 subject- matter eligibility rejections based on Alice, we estimate that the post-Alice (since 2014) allowance rates for MML inventions will range between 55% and 70%. The legal status of the granted patents from S3 reveals that the majority (81.73%) are active (in force). The other 18.27% (n=686) Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY3 have already expired and are now in the public domain and freely available for society to use. Figure 1 shows that the expired patents a primarily for granted patents from 2001-2005 which have priority dates before 2001. 2) Which organizations are leading the patent activity for medical uses of known products? What is the patent office of choice? The results of S5 were analyzed to determine which organizations are leading the patent activity in AI/ML with medical applications. Figure 2 lists the top owners of patents with MML claims. The figure shows the current top assignees (patent owners) and their corresponding number of patents granted with AI/ ML claims (S5) classified within Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY4 Fig. 1 Annual patenting activity for medical machine learning technologies and legal status. US and EPO patent applications and granted patent publications with claims directed to medical machine learning (MML) from 2001 to 2021 by publication date of the patent application (S3). The relative proportion of granted applications (red) and abandoned/rejected applications (grey) to the total number of applications for years with minimal number of pending applications (light blue) indicates the patent allowance rate (ranging 48% in 2007 to 64% in 2012). N o. P at en ts 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 20 10 20 11 20 12 20 13 20 14 20 15 20 16 20 17 20 18 20 19 20 20 20 21 0 250 500 750 1000 1250 1500 1750 2000 2250 2500 2750 Granted patent applications Rejected/Abandoned patent applications Pending patent applications Expired patents Fig. 2 Organizations with the highest count of MML pending patent applications, patent grants, and expired patents (S5). The graph shows the current patent owners (assignee) and their corresponding number of AI patents in CPC code A61B (diagnosis, surgery, identification, medical devices and methods). Ultimate Assignee N o. P at en ts SI EM EN S AG SA M SU NG E LE CT RO G EN ER AL E LE CT RI G E PR EC IS IO N HE UN IV C AL IF O RN IA SO NY C O RP FU JI FI LM C O RP CA NO N KK UN IV C AS E W ES TE G O VE RN M EN T O F T ST AN FO RD U NI V DE XC O M IN C PH IL IP S EL EC TR O M ED TR O NI C PU BL I IB M M ER AT IV E US L P HE AR TF LO W IN C CA NO N M ED IC AL S AL PH AB ET IN C JO HN SO N & JO HN S VE RI LY L IF E SC I M IC RO SO FT C O RP SI EM EN S M ED IC AL 0 50 100 150 200 250 300 350 400 Granted patent applications Pending patent applications Expired patents CPC A61B (diagnosis, surgery, identification, medical devices, methods). Among the top MML patent owners, we find large corporations active in the medical d e v i c e s / t e ch n o l o g i e s f i e l d , inc lud ing S iemens, Ph i l l ips, Samsung, Medtronic, GE and IBM. Universities are also among the top MML owners; the University of California, Case Western, and Stanford are among the top 20 assignees. Finally, our results reveal that large technology companies such as Google (Alphabet/Verily) and Microsoft are also active in MML and among the top patent owners. Figure 3 shows the number of granted MML patents (S5) segregated by the patent office (USPTO vs. EPO). Our results indicate that the USPTO has been the patent office of choice over the last 20 years. The vast majority of granted MML patents have been granted by the USPTO (87% in 2 0 2 1 ) . T h i s o v e r w h e l m i n g preference exists even though the US landscape for AI patent eligibility has been unsettled, compared to the EPO’s relative stability with regard to subject- matter eligibility and the specific “EPO Guidelines for Examination of Artificial Intelligence and Machine Learning (G.II.3.3.1).” This f inding may surpr ise European policymakers, especially since the EPO Guidelines make it clear that MML inventions make a “technical contribution” and therefore are subject matter eligible under EPC Art. 52. Notably, the EPO Guidelines provide specific examples of AI/ML applied to medical applications (MML): “For example, the use of a neural network in a heart monitoring apparatus for the purpose of identifying irregular heartbeats makes a technical contribution. The classification of digital images, videos, audio or speech signals based on low-level features (e.g.  edges or pixel attributes for images) are further typical technical applications of classification algorithms” (EPO Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY5 Fig. 3 Choice of patent office (USPTO vs. EPO) for MML granted patents (S5). N o. P at en ts 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 20 10 20 11 20 12 20 13 20 14 20 15 20 16 20 17 20 18 20 19 20 20 20 21 0 100 200 300 400 500 600 700 800 USPTO EPO Fig. 4 Granted patents with MML claims per year. The blue bar shows the count of granted patents (S5) with MML claims (independent or dependent) and the green bar shows the number of patents (S6) with independent MML claims (broadest claims the patent). The results indicate that in 2021 MML patents contained the AI limitations as part of their independent (broadest scope) claims in approximately 58% of the patents while ten years earlier it was approximately 20% (2011). N o. P at en ts 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 20 10 20 11 20 12 20 13 20 14 20 15 20 16 20 17 20 18 20 19 20 20 20 21 0 100 200 300 400 500 600 700 800 S5-AI/ML in claims S6-AI/ML in independent claims G.II.3.3.1). Other specific MML examples l is ted in the EPO G.II.3.3.1 AI/ML Guidel ines include “providing a medical diagnosis by an automated system p r o c e s s i n g p h y s i o l o g i c a l measurements”, and “providing a genotype estimate based on an analysis of DNA samples, as well as providing a confidence interval for this estimate so as to quantify its reliability.” 3) What types of claim strategies and formulations are being used to protect these inventions (MML) and what is their relative prevalence? Figure 4 shows the number of granted patents with MML claims per year and the proportion of these patents that conta in AI/ML l i m i t a t i o n s a s p a r t o f t h e independent claims (S4). In 2021, the overwhelming majority (94%) of the patents that included an AI/ML keyword in the title or abstract of the patent also contained claims directed to AI/ML. The broadest (independent) claims of MML patents are increasingly focused on AI/ML claims. Our S5 search results revea l AI/ML l imitat ions in independent claims for 58% of granted patents, as shown in Figure 4. This is a substantial increase compared to the practice 10 years earlier, where only 20% of granted patents included AI/ML limitations as part of the independent claims. We further analyzed the claim formulations used to protect MML inventions and their relat ive prevalence. Figure 5 shows the relative prevalence of MML patents (S6) with claim limitations directed to (a) machine learning, (b) neural networks, (c) artificial intelligence/ AI, and (d) deep learning. From 2001 to 2016, “neural network” was the most prevalent limitation used in the independent claims. Since 2017, “machine learning” is the most used limitation, followed by “neural n e t w o r k ” a n d “ a r t i f i c i a l intelligence”/“AI.” Notably, “deep learning” started to be present as a limitation as part of the independent claims of granted medical of related patents (classified under CPC A61B) in the last 4 years. Assuming the recent trend continues, we expect “deep learning” will be more prevalent than “AI” by 2023. 4) What medical applications and input signals are MML patents focusing on? Determining the type of medical applications and input signals requires expert review of the patents. All the MML-granted patents in S7 were analyzed and classified based on an expert review of their independent claims. The goal was to determine the nature of the MML contribution considering the level of support provided by the AI/ML. This ranged in the clinical task spectrum from a) measurement (e.g., the AI/ML is used to obtain a non-invasive measurement of a physiologic signal) or analysis (e.g., relevant physiologic parameters are calculated using the AI/ML) and b) decision support, detection, and classification (e.g., the AI/ML is used to analyze the signals to detect or classify a possible malignant feature) to c) diagnosis, prognosis, m o n i t o r i n g , a n d t r e a t m e n t . Accordingly, the MML patents (S7) were classified as directed to 1) technology improvements (e.g., an improvement to a generic medical device such as an MRI by using AI/ ML), 2) measurement, 3) analysis, 4) decision support, 5) detection, 6) classification, 6) diagnosis, 7) prognosis, 8) monitoring, and 9) treatment (see Supplementary Information). Figure 6 shows the classification of granted MML patents (S7) by type of medical application and input signal dimensions. Contrary to the AI/MML hype, our results indicate that there are not so many patents directed to fully automatic AI-based diagnosis. Even broadly defined, the diagnosis category (i.e., Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY6 Fig. 5 Relative prevalence of MML patents (S6) with claims limitations directed to (a) machine learning, (b) neural networks, (c) artificial intelligence/ AI, and (d) deep learning. N o. P at en ts 20 12 20 13 20 14 20 15 20 16 20 17 20 18 20 19 20 20 20 21 0 100 200 25 50 75 125 150 175 225 S6-L1: Machine Learning S6-L2: Neural Network S6-L3: Artificial Intelligence S6-L4: Deep Learning including also semi-automatic and AI-assisted diagnosis), only accounts for 5% of the MML patents. Similarly, AI-based prognosis was only claimed in 3% of the cases and treatment in 2%. Patents do not require an “actual reduction to practice” (i.e., creating a working prototype) as long as the written description, enablement, and best mode requirements (e.g., 35 USC 1 1 2 ) a r e s a t i s f i e d ( i . e . , a “constructive reduction to practice” to enable the claims with sufficient disclosure for those with ordinary skill in the art). Nevertheless, our results indicate that applicants are largely conservative when claiming MML inventions. Most of the MML claims are focused toward the lower- end of the clinical task spectrum including measurement and analysis (27%) and mid-level support tasks such as detection and classification (36%). Furthermore, 26% of the patents with MML claims were d i r e c t e d t o t e c h n o l o g y improvements such as platform tools and medical devices. Diagnosis, p r o g n o s i s , m o n i t o r i n g , a n d treatment together comprised only 12% of MML patents. Granted MML patents also differed based on the input signals used. One-dimensional (1D) input (i.e., 1xn vectors) includes health data and physiologic signals such as ECG, EEG, EMG, SpO2; two-dimensional (2D) input (nxn matrices) includes images such as MRI, sonography, images of eye, skin, tissue; and three-dimensional input (3D) requires video inputs to the AI/ML system claimed in the patent. The broadest independent claims of granted MML patents (S7) were d i rected to 1D hea l th data , physiologic signals and time series data in 59% of the cases, whereas images accounted for 36% of the cases and video for 5% of the cases. This distribution is particularly noteworthy since there are many more 1D signals/time series and medical devices that operate from these 1D signals—the 41% of MML patents focused on images or video (36% on images) is very high relative to that baseline. Images are the most prevalent input signal in claims directed to a) medical technology improvements and b) medical diagnosis (Figure 6). AI/ML methods are known to work well with images and be useful in a large va r i e t y o f imag e -process ing applications. Our results show that rea l -wor ld patent ing act iv i ty supports this and suggest that AI/ ML techniques are particularly well suited for image-based medical diagnosis. MML Medical Application Examples The Supplementary Information provides examples from the expert review with patents classified by t h e i r i n d e p e n d e n t c l a i m s . Additionally, it contains detailed infor mat ion for each patent i n c l u d i n g i n d e p e n d e n t a n d dependent claims, current assignees, original assignees, inventors, priority dates, and citations. Table 2 provides an example for each application classification. The examples for measurement and Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY7 Fig. 6 Classification of granted MML patents (S7) by type of medical application and input signal dimensions. One dimensional (1D) signals include physiologic signals such as ECG, EEG, EMG, SpO2; two dimensional (2D) are images such as MRI, sonography, images of eye, skin, tissue; and three dimensional (3D) are video. 3D input (nxnxn) - Video 2D input (nxn) - Images (e.g, MRI) 1D input (1xn) - Signals (e.g, ECG, EEG, EMG) Diagnosis TreatmentMonitoring 0 10 20 30 40 50 60 Technology Improvement Measurement Analysis D.Support Detection Classification Prognosis a n a l y s i s a r e i l l u s t r a t i ve o f conservative claims that handle 1D input. For instance, one claim produces analysis of a signal by estimating “psychological conditions based on frequency variations of the physiological fluctuation signal” but does not go further to use that analysis (ID 29, Table 2). Mid-level support tasks, including decision s u p p o r t , d e t e c t i o n , a n d classification, are less conservative. For example, another claim analyzes ECG signals and then uses that analysis to detect anomalies and “assign an anomaly label to the ECG signal for a detected anomaly” (ID 33, Table 2). Similarly, in another example, the independent claim was directed to “detecting and grading carcinoma” from MRI (ID, 10). For higher-level clinical tasks, such as diagnosis, prognosis, monitoring, and treatment, one exemplary claim is directed to detecting disordered breathing using a neural network and then treating that condition by “controlling a breathing gas pressure in response” to the neural network output (ID 121, Table 2). However, the majority of claims do not reach that far on the clinical task spectrum. A substantial number MML patents are claimed with a high level of specificity. For example, a claim directed towards improving MRI imaging technology includes the steps of under-sampling an image, inputting the image into a neural network, and then performing a recursive process that decomposes the image, applies a thresholding function, and reconstructs the image (ID 3, Table 2). Similarly, the broadest claim for the patent focused on “detecting and grading carcinoma” (ID 10) from MRI requires specific pre-processing and feature extraction based on a convolution neural network (CNN), a support vector machine (SVM), and a Gaussian radial basis function (RBF) kernel SVM classification of combined SVM decision values and statistical features to produce a final decision. This level of detail can render the claim difficult to infringe and runs counter to a scholarly narrative that patents directed to AI are likely to be overly broad. Conclusions The landscape of patents on MML have several implications for existing scholarship and for policy going forward. First, fears that patents would be unavailable for MML inventions seem to be largely unwarranted. Our results show that after a decade of relatively low growth, since 2013 MML patent grants have been rapidly growing with a CAGR of 38.51%. In 2021, over 700 patents were granted containing claims directed to AI/ML for medical applications. The allowance rate for these MML patents s tead i l y increased from 48% in 2007 to 64% in 2012 and looks to be staying relatively high. The majority of the MML patents granted over the last 20 years are in force (81.73%), but 18.27% are now expired and are already in the public domain available for society to use freely. Second, the doctrinal instability around patentability of AI in the United States also seems not to have dissuaded MML patenting at the USPTO. Despite the relative greater stability regarding subject-matter eligibility of the EPO compared with the US jurisprudence, our results indicate that the USPTO has been the patent office of choice over the last 20 years (87% USPTO vs. 13% EPO in 2021). And even though AI/ML applications are often impacted by 35 USC 101 subject-matter eligibility rejections based on Alice, our results indicate that the post-Alice allowance rates for MML inventions are likely to range between 55% and 70%. Inventors’ preference for the USPTO exists notwithstanding the fact that European-headquartered corporations are among the top MML patent owners. Over the last 20 years, the MML patenting activity has been led by large corporations active in the medical technologies field, such as Siemens, Phillips, Medtronic, Samsung, and GE. More recently, we also find technology companies such as Google and Microsoft among the top MML patent owners, as well as universities such as University of California, Case Western and Stanford. The two clear leaders, Siemens and Philips, are among the three largest manufacturers of MRI scanners, (GE, the third, is fifth among MML patentees). This area concentration is congruent with the high rate of image-input patents and with the dominance of radiology among FDA-cleared or classified1 MML- based products (representing three- quarters of all FDA-listed MML products13). Nevertheless, despite this moderate imaging focus, MML patenting more generally is not especially concentrated; the top-20 S5 patentees together represent only about 28% of all MML patents. Third, patents referencing MML are actually about MML inventions. Given the hype over AI and machine learning, one might expect that many patents reference these techniques in the abstract, but that Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY8 the references are buzzwords without real importance to the claimed invention. While this might once have been the case, it is not today. In 58% of MML patents, AI/ ML limitations appear in the broadest (independent) claims. This represents a sharp increase; merely a decade ago, only 20% of MML patents had such AI/ML limitations as part of the independent claims. Claim practice is also evolving as the technology does, with “machine learning” emerging as the most prevalent limitation since 2017, followed by “neural network,” “AI”, and “deep learning”. Fourth, despite the MML hype, our results show that claims are relatively conservative. There are relatively few patents directed at fully automatic AI-based diagnosis. Most of the MML claims analyzed focused instead on measurement and analysis (27%) and detection and classification (36%). In the case of diagnosis, images are the most prevalent input signal to the AI/ML system. We suspect this pattern may change, and the current focus on ear l ier c l inical tasks ref lects capabilities and the current state of knowledge—but for now, most MML patents are farther from clinical decisions, even as policy debates focus on MML systems that focus on the point of care itself14,15. More work remains to be done. Further research is required to investigate the scope of protection for these MML patent claims, their validity in light of current subject- matter eligibility requirements and AI/ML prior art, as well as the sufficiency of disclosure of the patent specifications3. But an early look suggests that patenting remains an important player in innovation incentives for AI/ML in medicine, with a range of inventors actively seeking patent protect ion in s u r p r i s i n g l y c o n s e r v a t i v e applications. Acknowledgements The research was supported, in part, by a Novo Nordisk Foundation Grant for a scientifically independent Collaborative Research Programme in Biomedical I n n o v a t i o n L a w ( g r a n t n o . NNF17SA027784). Competing interests The authors declare no competing interests. Additional Information Supplementar y infor mation i s available for this paper at https:// doi.org/TBC References 1. Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices. (FDA, https:// www.fda.gov/medical-devices/ software-medical-device-samd/ artificial-intelligence-and-machine- learning-aiml-enabled-medical- devices, 2022). 2. Price, N. BIG DATA, PATENTS, AND THE FUTURE OF MEDICINE. CARDOZO LAW REVIEW 37, 1401-1453 (2017). 3. Tull, S. & Miller, P. Patenting Artificial Intelligence: Issues of Obviousness, Inventorship, and Patent Eligibility. Robotics, Artificial Intelligence & Law 1, 315-325 (2018). 4. Aboy, M., et al. How does emerging patent case law in the US and Europe affect precision medicine? Nat Biotechnol 37, 1118-1125 (2019). 5. Williams, H. How Do Patents Affect Research Investments? Annual Review of Economics 9, 441-469 (2017). 6. Bubela, T., et al. Patent landscaping for life sciences innovation: toward consistent and transparent practices. Nat Biotechnol 31, 202-206 (2013). 7. Smith, J.A., et al. The Reporting Items for Patent Landscapes statement. Nat Biotechnol 36, 1043-1047 (2018). 8. Aboy, M., Minssen, T. & Kop, M. Mapping the Patent Landscape of Quantum Technologies: Patenting Trends, Innovation and Policy Implications. IIC-International Review of Intellectual Property and Competition Law 53, 853-882. 9. Aboy, M., Liddell, K., Liddicoat, J. & Crespo, C. Myriad's impact on gene patents. Nat Biotechnol 34, 1119-1123 (2016). 10. Aboy, M., Liddicoat, J., Liddell, K., Jordan, M. & Crespo, C. After Myriad, what makes a gene patent claim 'markedly different' from nature? Nat Biotechnol 35, 820-825 (2017). 11. Aboy, M., Liddell, K., Jordan, M., Crespo, C. & Liddicoat, J. European patent protection for medical uses of known products and drug repurposing. Nat Biotechnol 40, 465-471 (2022). 12. Aboy, M., Crespo, C., Liddell, K., Minssen, T. & Liddicoat, J. Mayo's impact on patent applications related to biotechnology, diagnostics and personalized medicine. Nat Biotechnol 37, 513-518 (2019). 13. Reuter, E. 5 takeaways from the FDA’s list of AI- enabled medical devices. in MEDTECHDIVE (2022). 14. Minssen, T., Gerke, S., Aboy, M., Price, N. & Cohen, G. Regulatory responses to medical machine learning. J Law Biosci 7, lsaa002 (2020). 15. Clinical Decision Support Software: Guidance for Industry and Food and Drug Administration Staff. 1-26 (FDA, 2022). Patents Article - Author’s Manuscript Version - © Nov 2022 - LML - University of Cambridge, UK - Submitted to: NATURE BIOTECHNOLOGY9 https://doi.org/TBC https://doi.org/TBC