We downloaded approved US patents between 1995 and 2017 from the US Patent and Trademark Office website [14]. The whole dataset included more than 5 million patent documents, whose formats changed three times: Green Book format (1995-2001) [15], Red Book Standard Generalized Markup Language (SGML; 2002-2004) [16], and Red Book XML (2005-2017) [17]. These patent documents were classified into different innovative domains (eg, agriculture, sports, and foodstuffs) based on the United States Patent Classification (USPC) system [18] and the Cooperative Patent Classification (CPC) system [19]. We set the following inclusion criteria to extract patent documents on biomedicine: Publication date should be between January 1, 1995, and December 31, 2017, and the patent document should contain at least one USPC or CPC classification code listed in Multimedia Appendix 1.

Any documents that did not meet the abovementioned criteria were not considered to be related to biomedicine. We developed a python parser (Multimedia Appendix 2) to parse patent documents in all three formats in order to retrieve information such as patent identification, issue date, title, abstract, claims, and USPC and CPC classification codes. We then filtered for biomedicine-related patents using the compiled lists of USPC and CPC classification codes (Multimedia Appendix 1).

For patent documents related to biomedicine, we used MetaMap, an application developed at the National Library of Medicine [20] to extract and map biomedical concepts (key words or phrases) of 74 semantic types in the sections of title, abstract, and claims to the Unified Medical Language System (UMLS) metathesaurus. The complete list of 74 UMLS semantic types is shown in Multimedia Appendix 1.

UMLS assigns a Concept Unique Identifier (CUI) to each concept and links it to the source thesauri such as the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) [21]. We leveraged ICD-9-CM to identify concepts of diseases and medical conditions in patent documents and mapped ICD-9-CM to PheCode, which represents clinically meaningful phenotypes used by clinicians [22]. To address the issue of concept granularity, we only used the 704 root PheCodes such as diabetes mellitus, influenza, and pain. Thus, patent documents were grouped by diseases and medical conditions before ROI/PHI and topic modeling analysis.

In the second phase, we started by calculating the patent coverage of each disease or medical condition by dividing the number of patent documents mentioning the disease or medical condition by the total number of patent documents mentioning all diseases and medical conditions in each year. This is a preparation step for computing the ROI for each disease or medical condition and PHI for all diseases and medical conditions in each year.

To prepare for topic modeling, we encoded multiword biomedical concepts using their corresponding CUIs before tokenizing the patent documents, removing stop words, and lemmatizing words. Multiword concept encoding helps preserve compound concepts such as “type 2 diabetes.” Without multiword concept encoding, such a compound name would be broken down to individual words (unigram) during topic modeling. Tokenization breaks text into smaller meaningful elements such as words, numbers, or punctuation marks. Stop words like “the,” “is,” and “are” are usually filtered out from natural language processing. Lemmatization aims to reduce the morphological variations of words by returning to the base or dictionary form of a word (eg, “walked,” “walking,” and “walks” have the same base form “walk”).

The data on disease burden, publications, and clinical trials were collected in the same way as that in our previous work [12]. However, in this study, we included more data from recent years to obtain an updated longitudinal analysis for 17 years (2000-2016). Disease burden was estimated using the total treatment cost in million population each year from OptumLabs Data Warehouse [23]. OptumLabs Data Warehouse is a comprehensive, de-identified administrative claims database for commercially insured and Medicare Advantage enrollees in a large and private US health plan. The diagnosis records were coded by ICD-9-CM before October 2015 and International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) [24] after those claims.

For publications, we downloaded all Medical Subject Heading (MeSH)-indexed abstracts in English from the MEDLINE database (via the PubMed query interface), the biomedical publication database maintained by the United States National Library of Medicine from 2000 to 2016. Subsequently, we used the number of publications annotated by each MeSH disease or medical condition term to approximate the attention it received from the biomedical research community. We also downloaded the aggregated, MeSH-indexed clinical trial database [25]. Similarly, we used the number of clinical trials related to each specific disease or medical condition to approximate feasibility and popularity of carrying out clinical research in each disease or medical condition area. We converted ICD-9-CM, ICD-10-CM, and MeSH codes to the root PheCode before further analysis.

We previously proposed ROI and PHI for quantitatively measuring resource allocation for a particular disease and all the medical conditions as a whole [12]. The calculations of ROI and PHI are flexible and can incorporate many quantitative factors that impact research prioritization and resource allocation in biomedicine. In this work, we updated the ROI and PHI calculations by including more data from recent years and included patent data in addition to disease burden, publications, and clinical trials, since patents are an important form of intellectual property on technological innovations.

ROI and PHI are defined as follows:

where Y_nd is the raw measurement n for a disease d. In our model, we used treatment cost per million people as indicators of burden of disease (Y_bd ) and the number of research publications (Y_rd ), the number of clinical trials (Y_cd ), and the number patent documents (Y_pd ) as an approximation of resources spent in biomedical research and development. Since the raw measurement Y_nd is subjective to inflation (eg, increasing treatment cost and number of publications over time) and cannot be compared across different units (eg, treatment cost in dollars vs number of research publications by count), we used the normalized measure X_nd instead of Y_nd in ROI and PHI calculation. ROI quantifies the imbalance between needs and resource investment on multiple dimensions for each disease or medical condition. PHI describes the overall resource allocation efficiency for all diseases and medical conditions.

Topic modeling automatically uncovers the topics or themes in a large collection of documents, in terms of a set of keywords occurring together and most frequently [26-29]. We applied a dynamic topic model (DTM) [13] to learn the topics in patent documents related to specific diseases and medical conditions and the evolution of these topics over years.

DTM is an extension of the static Latent Dirichlet Allocation method [26] for analyzing the temporal changes of topics of a large collection of documents. Static Latent Dirichlet Allocation does not consider the input order of documents in the large collection. It assumes the Dirichlet prior distributions for topic distributions in a document and word distributions over a topic. DTM recognizes the importance of temporality in a large collection of documents and studies the dynamics of topics from time interval t-1 to t. It assumes a Gaussian distribution for the prior parameters at t, given their value at t-1.

We used the open-source DTM C++ package [30] wrapped in Gensim library [31] to learn the temporality of the topics in patent documents related to a specific disease or medical condition. After multiword concept encoding, tokenization, stop word removal, and lemmatization, each patent document was converted into a vocabulary vector, where the elements were frequency of each lemma (including CUIs) without considering the order of lemma. As such, all the patent documents on the same disease or medical condition were converted into a D × V matrix, where D stands for count of patent documents and V denotes the size of the entire vocabulary in those patent documents. The D × V matrix was then chunked into time intervals in the calendar year for dynamic topic modelling.

We calculated topic coherence [32] quantitatively and asked domain experts to qualitatively evaluate the learned topics. More specifically, we evaluated the topic coherence at different topic numbers (ie, 2, 4, 6, 8, 10, 12, 14, 16, 18, and 20) to determine the optimal topic number for three selected diseases: diabetes mellitus, breast cancer, and epilepsy. We found that the optimal topic number was 6 for breast cancer–related patents, 8 for diabetes mellitus–related patents, and 16 for epilepsy-related patents (Multimedia Appendix 1). The average number of topics for the three diseases was 10. Thus, we set the topic number to 10 empirically for the remaining 639 identified diseases and medical conditions.

We also investigated the hyperparameters of α and σ in DTM: α controls the topic distributions over a document, and a smaller α results in fewer topics statistically associated with a document, whereas σ determines how fast the topics evolve over time, and a smaller σ leads to more similar word distributions over a topic over time. In our experiment, we used the default values of 0.01 and 0.005 for α and σ, respectively, as previously suggested [33], as those authors reported that both α and σ did not affect topic distributions and word distributions significantly over time and did not have an effect on topic interpretation by domain experts.

We collected 5,010,329 patent documents from 1995 to 2017, of which 550,961 (about 11%) were related to biomedicine. Figure 2 shows the percentage of patent documents related to biomedicine and the number of diseases and medical conditions covered in those patent documents in each year. It seemed that the approved US patents on biomedicine fluctuated in the range of 9.6%-14.6% during 1995 and 2017. However, the number of diseases and medical conditions covered in US patents expanded from 502 to 596, suggesting that technology innovations had been focusing on more diseases and medical conditions during the same time.

Figure 3 shows the coverage percentages of 24 most mentioned diseases and medical conditions in US patent documents between 1995 and 2017. We found that 16 of them were chronic conditions such as dysrhythmia, “other cancer,” diabetes mellitus, and obesity. “Other cancer” refers to malignant neoplasm of head, face, neck, abdomen, pelvis, or limb; disseminated malignant neoplasm; Merkel cell carcinoma; and malignant neoplasm, malignant carcinoid tumors, neuroendocrine tumor, and carcinoma in situ of an unspecified site, according to PheCode [22]. The percentages of patented inventions on “other cancer” are shown to be steadily increasing (from 2.29% in 1995 to 6.56% in 2017). This time period witnessed a steady increase of innovative technologies from functional magnetic resonance imaging to immunotherapy, which improve cancer diagnosis and treatment [34]. With increasing awareness and expenditure on cancer diagnosis and treatment [35], we believe that technological innovations associated with cancer are likely to grow continuously. Similarly, obesity ranked 21st in the most mentioned 24 diseases and medical conditions, and its patent coverage ranged from 0.87% to 1.65% during the study period. Such an increase in patented inventions for obesity aligns well with its high prevalence in the US, the severe implications on people’s life and the society, and national strategic plans to deal with the obesity epidemic [36]. Influenza, an acute disease, received decreasing attention in terms of patents before 2000 and increasing attention after 2010, reflecting its high prevalence and substantial burden (eg, large morbidity and mortality) on the health care system and society [37]. The Centers for Disease Control and Prevention reported that influenza caused 9.3-49.0 million illnesses, 140,000-960,000 hospitalizations, and 12,000-79,000 deaths in the United States each year since 2010 [38].

Before ROI and PHI analysis, we examined the multicollinearity between the relative number of patent documents and other factors, namely, the relative treatment costs, the relative number of scientific publications, and the relative number of clinic trials used in our previous model [12]. The low variance inflation factors of 1.0-2.6 (Multimedia Appendix 1) indicated a low multicollinearity between the relative number of patent documents and the other factors [39].

We then calculated the ROIs to measure the misalignment between the resources allocated to a disease or medical condition and the burden it imposed between 2000 and 2016, as shown in Figure 4 (see Multimedia Appendix 3 for the raw data). For example, neuroendocrine tumors were overstudied, indicated by negative ROIs from 2000 to 2016. However, their ROI increased from –17.93 (in 2000) to –1.87 (in 2016) over 17 years, implying that resources allocated to it were more aligned with its burden. Further, digging into the dependent variables of ROI, we found that the driving factor for such improvement was that the relative treatment cost (indicator of disease burden) increased much faster than relative publications, relative clinical trials, and relative technological innovations. Similar patterns were observed for most of the overstudied diseases and medical conditions. There were only a few diseases and medical conditions that were more overstudied over time. Injury to nerves not elsewhere classified was such an example. Its ROI scores declined from –6.73 to –9.43 from 2000 to 2016, primarily because its relative treatment cost decreased while it received steady attention in biomedical research (in terms of the relative number of publications) and increasing attention in development (in terms of the relative numbers of clinical trials and patents).

Figure 4 B highlights the understudied diseases and medical conditions with positive ROIs. For example, ankle and foot fracture had steadily declining ROIs (from 12.52 to 6.41) from 2000 to 2016, mostly because the resources allocated to it increased, but its burden slightly decreased over time. More specifically, the relative number of publications increased more than 180 times, the relative number of clinical trials increased over 2 times, and the relative treatment cost decreased by about 20%. The relative numbers of publications, clinical trials, and patents were all disproportionally small, compared to the burden of ankle and foot fracture. In contrast, intracranial hemorrhage (injury) had gradually increasing ROIs (from 6.31 in 2000 to 10.19 in 2016) and was becoming more understudied due to increase in the relative treatment cost and decline in the relative number of publications. There were also diseases and medical conditions that showed fluctuating ROIs over time. For instance, the ROIs for contact dermatitis increased from 6.90 to 8.20 during 2000-2014 and decreased to 5.63 in 2016. Our calculation showed that its relative treatment cost declined steadily by a factor of 1.6 from 2000 to 2016, but its relative number of publications and patents fluctuated dramatically during the same period.

The overall alignment between disease burden and allocated resources for all the diseases and medical conditions measured by PHI from 2000 to 2016 is shown in Figure 5. It demonstrated a clear shrinking pattern over the 17-year period, with small increases in 2015 and 2016. As the smaller PHI indicates better alignment between research and development and the distributions of needs across all diseases and medical conditions, the results suggest that the resource allocation for all the diseases and medical conditions as a whole had improved significantly over time.

Using a dynamic topic modeling technique, we identified the latent topics and their changing patterns over time for patent documents related to various diseases and medical conditions. In Figure 6, we highlight three meaningful topics identified from patent documents related to diabetes mellitus, breast cancer, and epilepsy, together with their changing patterns from 1995 to 2017. The results for the other diseases and medical conditions are provided in Multimedia Appendix 3.