We conducted a literature search in PubMed using an expanded version of an existing query [20] containing search terms related to offenders and prisons. These were combined with either the Medical Subject Headings term “epidemiology” to capture all types of epidemiological studies or with all the available (in PubMed) publication types (eg, meta-analysis and clinical trial) to ensure the results will return clinical trials and secondary research (eg, review). We also added terms related to randomization/natural experiments and synthetic control. These choices prevented articles that made only passing references to prisoner and offender studies from entering the dataset, resulting in a high-quality corpus for our analysis. The search was restricted to English-language articles that have an abstract and involved only human participants (ie, veterinary research was excluded). The full query (Textbox 1) was run on July 20, 2023.

The system’s performance was evaluated at the abstract level and used the standard definitions of the precision, recall, and F₁-score metrics [47]. True positive (TP) was defined as the identification of either all the correct mentions of a population characteristic or the extraction of a number of mentions for one population characteristic, even if the system failed to pick up some mentions in an abstract. For example, if an abstract had 2 mentions of the female sex (eg, “females” and “women”), then the detection of either one or both mentions would be considered a TP at the abstract level with “female” being the standardized sex in this abstract. The same process was applied in cases where there can be more than one different mention of a population characteristic (eg, mentions of 2 different nationalities for the investigated population). A false positive (FP) is an identification of an incorrect mention of a population characteristic while a false negative (FN) is an incorrectly ignored mention of a population characteristic. Precision measures the accuracy of TP predictions, recall measures the completeness of identifying TPs, and the F₁-score balances both by combining them in a single metric.

Overall, at the abstract level, the mean precision and recall were 95.2% and 90.9%, respectively, whereas the F₁-score was 93%. However, since the number of mentions between the 4 characteristics varied drastically in the evaluation set, we reported on the micro values of precision, recall, and F₁-score to offer a more weighted approach to the system’s performance. Microprecision was 94.3%, with microrecall at 85.9% and micro–F₁-score at 89.9%. The largest recall drop was observed in age (6.2%), and it was the only recall with a value below 80% (78.8%), while nationality had the highest recall (95.2%). Sex had the highest precision (100%), followed by age (97.6%; Table 3). The highest F₁-score was observed for sex (96.4%), followed by nationality (93%). Age and offense type had similar F₁-scores with 87.1% and 87.3%, respectively (Table 3).

Our query returned 34,481 justice health study abstracts with the earliest recorded in 1963 (Multimedia Appendix 5). Half of the abstracts (17,039/34,481, 49.4%) did not have any characteristic mentions. Most abstracts either mentioned only age and nationality (1676/34,481, 4.9%) or age and offender type (1082/34,481, 3.1%). Only 1.3% (443/34,481) of abstracts reported all 4 characteristics (sex, age, offender type, and nationality).

A total of 5170 (14.9%) abstracts out of 34,481 reported the population’s age; of the 5170 abstracts, 3581 (69.3%) mentioned minors, 3037 (58.7%) mentioned adult populations, and 181 (3.5%) not specifying the age. The largest adult group was that of 18-24 years old (33.4%) followed by 35-44 years old (10.4%). Studies involving 55-64-year-olds had the lowest number of mentions (7.4%; Table 4).

A total of 8169 (23.7%) out of 34,481 abstracts reported the sex of the investigated population in the abstract. Around 39.7% (3241/8169) of the abstracts reported only male populations, 42.9% (3501/8169) reported only female populations, and 17.4% (1418/8169) reported both male and female populations. Less than 1% (n=21) of abstracts reported transgender populations. Although there has been a gradual increase in study rates involving only female populations since 1990 (see the trend line in Figure 1), from 2014 onward, a decrease was noted with regard to abstracts reporting only male populations, surpassed by abstracts reporting only female populations. In 2023, a total of 48.6% (105/216) of abstracts reported only female populations versus 33.3% (72/216) reporting only male populations (Figure 1; Multimedia Appendix 6 shows the rates per year in detail).

A total of 9525 (27.6%) out of 34,481 abstracts reported the nationality of the investigated population. The most common nationality was the United States (ie, American; 2992/9525, 31.4%), followed by the United Kingdom (786/9525, 9.2%) and Australia (730/9525, 7.6%; Table 5). However, to account for the size of the country’s population, which we assumed to be broadly linked to the size of its prisoner population (Pearson r=0.73), and this in turn being a potential driver of the volume of research reflected by the number of publications, we derived a publication rate based on the average prisoner population size over the period of 2000 to 2020 [48] and calculated a rate per 1000 prisoner population. In this case, the Nordic countries were in the top 4 in terms of publication rate followed by Australia (Table 5). Only 4 countries from Asia (China, India, Japan, and South Korea) and 1 country from Africa (South Africa) were in the top 20 of both crude and rate ranks.

A total of 4814 (13.9%) out of 34,481 abstracts mentioned the offender type. Offenders with mental illness were reported in 17.4% (840/4814) of the PubMed abstracts, followed by sex offenders (620/4814, 12.9%). Child crime–related offenders (eg, child abusers) had the lowest number of mentions with 1.7% (84/4814; Table 6).

From 1990 to 2023, the overall number of PubMed abstracts with an offender type increased (Multimedia Appendix 7). However, the rate (ie, the number of PubMed abstracts with a specific offender type divided by the total number of PubMed abstracts that had a mention of an offender type) revealed a general increase for offenders with mental illness. Mentions for sex, drug-related, nonviolent, and violent offenders had an overall decrease, with the biggest noted for sex offenders (10%; Figure 2).

When the sex offenders are broken down by male and female, the rates of study abstracts through time remain roughly the same since 1990, with an average rate of 33% for male offenders and 15.5% for female offenders. A minimal increase in male sex offenders and a minimal decrease in female sex offenders were observed (Figure 3).

This text mining study demonstrated that key population characteristics (age, sex, offending type, and nationality) can be derived by applying text mining to a large corpus of study abstracts available in PubMed. Our findings enable researchers to investigate the presence of potential research and knowledge gaps over time that arise from examining certain offending groups within an entire discipline. Half of the abstracts (17,039/34,481, 49.4%) did not report any characteristics, while the number of abstracts that mentioned 1 characteristic ranged from 13.9% (n=4814; offender type) to 27.6% (n=9525; nationality). This highlights a larger problem regarding the reporting of the necessary information in abstracts for the description of populations within the justice health area as only 1.3% (n=443) of our sample abstracts reported sex, age, offender type, and nationality.

Previous research has showcased that despite several reporting guidelines covering observational, experimental, and secondary study reporting (eg, STROBE, CONSORT, SPIRIT, and PRISMA) [9-12], justice health abstracts do not appear to adequately detail their study designs and examined variables [31,34] and, based on our results, nor do they adequately describe the population under investigation. The importance of population description in research is needed to not only understand predictors for recidivism but also to enable the conduction of meta-analyses and other future studies [49].

We initially standardized the extracted age into 2 groups (younger than 18 years and older than 18 years) to examine age-related trends in offending populations. However, age was mentioned in 14.9% (5170/34,481) of our PubMed abstract sample; thus, this finding should be taken with caution. Although it has been suggested that the justice health area should focus more on young people [14], our results showed that 69.3% (3581/5170) of abstracts mentioned a population aged <18 years. This finding reflects the rise in youth involvement in the youth justice system [50]: the United States saw a 30% increase in juvenile cases between 1985 and 2009 [51], and Australia noted a recent increase of 6% from 2021-2022 to 2022-2023 [52].

The high number of abstracts reporting minors in the justice health area could also be explained through reporting practices. It is possible that the majority of researchers who investigate minor populations are more likely to specify their age. Most abstracts referred to the investigated population in generic terms such as “offenders” or “incarcerated individuals,” which could imply adult populations. This would separate them from younger people who are described by more specific terms such as “adolescent,” “juvenile,” and “delinquent.” This, however, was not taken into account for this research. Considering this and along with the inspection of full-text studies that might describe in detail the age of the participant population, our finding could be different.

Although male individuals make up the overwhelming majority of incarcerated populations (10.9 million worldwide vs <1 million for female individual), there was an overrepresentation of studies involving female offenders, suggesting that male individuals in prison are an underresearched group [53]. Increased research into female populations since 2000 aligns with an increase in the number of incarcerated female individuals worldwide since 2000 [54], with female offender studies rising from 29.2% (19/65) in 1990 to 48.6% (105/216) in 2023. Conversely, male-focused research decreased from 50.8% (33/65) in 1990 to 33.3% (72/216) in 2023. This disparity evokes a consideration of equity in justice health research. Equity and not equality should be prioritized in health [55,56]. Therefore, it is possible that although inequality is shown through rates of research between male and female offending populations, an equity approach can contribute to our understanding of why there is disproportionately more research on female populations.

However, since only 23.7% (8169/34,481) of our data sample reported sex, it is possible that the remaining studies that did not detail the population’s sex in the abstract focused on male populations. Given that most prisons hold male prisoners only, investigators focusing on female populations in the justice health area might be better at reporting female sex. Nevertheless, this highlights the need for more detailed reporting in PubMed abstracts, to allow other researchers to accurately synthesize information more effectively and accurately.

Nationality was the population characteristic with the highest prevalence in our sample (9525/34,481, 27.6%). Using the crude rank, the United States was the most common nationality. However, by implementing the publication rate, the United States dropped to the 19th place (ie, among the most common nationalities), with the Nordic countries (Norway, Finland, Sweden, Denmark) occupying the top 4 spots and Australia the fifth one. Previous research analyzing PubMed justice health abstracts showed similar rankings for the Nordic countries in terms of their total published outputs [20]. Nordic countries are regarded as having a progressive approach to offender rehabilitation, with proportionally lower numbers of incarcerated individuals and recidivism rates compared with many other countries [57,58]. However, these results are based only on 27.6% of PubMed abstracts with a reported nationality; so, it is possible that in full-text studies, the actual nationality of the examined population is described, which could in hindsight reveal a different picture in the rankings.

Offenders with mental illness were the most common group identified from the abstracts (840/4814, 17.4%). In the United States, it has been estimated that 24% of the inmate population have a mental illness [59], with approximately 50% to 75% of the 2 million young people meeting the criteria for a mental health disorder [60,61]. In the last 10 years, reliance on the juvenile justice system to meet its population’s mental health needs has increased and so has the research to examine the effectiveness of intervention and treatment programs [62]. Offenders with mental illness have higher rates of recidivism, exhibiting rehabilitation needs and prison adjustment difficulties that differentiate them from the general offender population [63,64]. Our results highlight the depth of this problem with researchers examining a total of 58 unique mental illness concepts (Multimedia Appendix 8) in the last 70 years, ranging from behavioral disorders (eg, attention-deficit/hyperactivity disorder) to mood disorders (eg, depression and bipolar disorder) and anxiety disorders (eg, posttraumatic stress disorder), with substance use disorders and intellectual disability receiving the most focus.

Most research on the longitudinal pattern of criminal careers has focused on generally violent offenders [65], which could explain the relatively low number of abstracts mentioning nonviolent offenders (96/4814, 2%) involved, for example, with theft and shoplifting. This indicates the need to investigate a more diverse range of offender groups [66,67]. Crimes such as theft, stalking, and driving under the influence may cause significant harm toward others, and yet, there is a lack of related work focusing on cases of, for example, fraud and sextortion, which can have significant effects on survivors [68,69].

Considering that sex offenders are regarded as one of the more serious offender groups [70,71], it is not surprising they were the second most commonly researched group in our sample (620/4814, 12.9%). European surveys have suggested that up to 10% of male offenders commit sexual violence against adult women [72,73], with Australia noting an average of 36.4% of all offenses recorded have been related to sexual assault in the last 15 years [74]. US statistics also put the prevalence of sexual assault at half a million incidents per year [75]. Despite male individuals comprising the majority of the sex offender population (eg, in Australia, 97% of sex offenders are men) [76], research suggests that the proportion of female sex offenders is higher than thought [77]. A recent meta-analysis with data from 12 countries reported that victimization surveys indicated that prevalence rates of female sex offenders were 6 times higher than official data (11.6%). This disparity is similar to our findings that saw female sex offenders comprising 13.1% (n=81) versus male sex offenders with 34% (n=211) from 620 abstract studies.

Despite an estimate of 1 billion children aged 12-17 years experiencing child abuse and maltreatment [78], our findings suggest that individuals responsible for committing such offenses are underresearched, with only 1.7% (84/4814) of the abstracts reporting child-related crimes. Since such offenses are hard to detect due to the involvement of minors and adolescents, with only 1.7% of our sample older than 70 years involving those convicted of child sex offenses, highlighting a research gap in justice health. To design and implement effective prevention and intervention programs for child-related crimes, it is necessary to conduct more evidence-based research on individuals committing this type of offense.

Our study comes with several limitations. First, PubMed articles might not be sufficient to portray a complete picture of offending and incarcerated populations since government articles and reports can remain unpublished and so fall outside the scope of this study. Second, research with a sociological and criminological focus is unlikely to appear in journals covered by PubMed. Thus, our dataset could potentially underestimate the total number of research outputs in this area. Third, we focused only on English-language abstracts, which carried a risk of “English-language” bias. However, the incorporation of non–English-language abstracts in our PubMed sample could ensure greater research transparency and findings and reduce bias.

Fourth, using only abstract text almost certainly does not give a full picture with regard to the investigated population. As noted in our findings, only a fraction of abstracts reported any of the 4 characteristics we examined: offender type (4814/34,481, 13.9%), nationality (9525/34,481, 27.6%), age (5170/34,481, 14.9%), and sex (8169/34,481, 23.7%). It is likely that full-text articles, especially those that might adhere to official reporting guidelines (eg, PRISMA and STROBE), detail their population reporting in the body text of the article, which would elicit different findings than those presented here. However, this was not feasible and would have involved permissions from numerous publishing houses and be extremely costly. In addition, changes over time in local and global research priorities including publication practices are likely to have influenced the results. Thus, our result interpretation should be taken with caution and mindfully of these macro-level influences.

Finally, despite a reliable performance from our methodology, the number of identified characteristics could be underrepresented (especially for age and offense type). Using specific rules might not have been enough to identify all mentions of age while more descriptive cases for offender types could have resulted in FNs. It is possible that a hybrid approach that uses both machine learning and rules will limit the number of FNs and thus enhance the accuracy of identifying these types of population characteristic mentions.

Our study demonstrated that it is feasible and efficient to extract key information from populations within a large sample of justice health study abstracts over time. Our findings align with existing research that has highlighted a focus on female offender studies and has revealed an emphasis on offenders with mental illness and minors with rising rates in the last 30 years. Interestingly, research involving child crime–related offenders was not common. Despite the United States having the largest incarcerated population in the world, adjusting its publication rate by the prisoner population demonstrates that Nordic countries with progressive approaches to offender rehabilitation have published proportionately more research. Our findings offer new insights into the whole area of justice health, with clear implications to promote diversity in cohort selection and limitation of bias and research gaps.