This study was a retrospective observational study of criminal court decisions involving UMP-related convictions in South Korea between 2005 and 2023. We identified closed criminal cases in which at least one defendant was convicted of an offense related to UMPs. The dataset was filtered from a comprehensive corpus of over 3 million court precedents collected by LBox, a legal technology company in South Korea [21]. We selected 2005 as the starting year because systemic indexing in the source corpus began in that year, allowing for consistent retrieval and longitudinal comparison across the study period [13,15]. This dataset serves as a proxy for the detected portion of UMPs, offering a rare window into the types of practices and defendants capture through regulatory oversight. This manuscript was reported with reference to the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement [22], and the completed checklist is provided in Multimedia Appendix 1.

Processing raw legal data posed significant challenges because court decisions are often stored as unstructured images or PDFs. To address this, we used a specialized data processing pipeline [15]. First, a rule-based layout classifier distinguished text-only pages from those containing tables or figures. Then, texts were extracted using LBox’s custom rule-based parser for PDFs and an optical character recognition engine for image files [23]. Third, to extract variables from the unstructured text bundles, a custom-made parser based on bidirectional encoder representations from transformers [24] segmented the texts into functional components, such as facts, legal rulings, reasons for sentencing, and conclusions. This methodology aligns with emerging standards in legal text mining for public health, enabling the conversion of complex narrative data into structured epidemiological variables.

We used an enhanced proprietary version of the end-to-end Information Extractor for Statistical Legal Analysis [25], an end-to-end neural system based on a domain-adapted mT5-Large architecture [26], for automated information extraction. The foundational model has been validated, achieving an average F₁-score of 93.1 across Korean legal information extraction tasks [25].

Given the high-stakes nature of legal and health data, we addressed the “black box” problem of artificial intelligence by implementing a rigorous human-in-the-loop validation process. Initially, 181 cases were labeled by the researchers to create the gold-standard set, which provided the few-shot examples for GPT-4 (OpenAI) [27] to generate labels for an additional 250 cases [28]. Second, these model-generated labels were thoroughly reviewed and corrected by 3 professional labelers, yielding a verified dataset of 631 cases. Third, this corrected dataset was used to fine-tune the Information Extractor for Statistical Legal Analysis–based system, which subsequently generated variable values autoregressively for the remaining cases. In total, 920 cases were labeled by the fine-tuned system and verified by the labelers. Cases involving complex compound charges (eg, UMPs combined with drug-related charges) were excluded (n=19) to ensure reliability. Finally, 1532 unique labeled cases were included in the final analysis (Figure 1).

As a single judicial decision could involve multiple charges, defendants, and legal rulings, counts across analytic levels were not directly comparable. Therefore, all percentages are reported with their explicit denominators. Analyses were conducted at 4 levels: cases (n=1532), charges (n=1718), defendants (n=1716), and legal rulings (n=2004).

Variables were defined through close collaboration between experts in computer science and medical law. For practitioner classification, we used 3 operational categories: health care professionals, health care providers, and ordinary persons. Health care professionals were defined as physicians, dentists, doctors of Korean medicine, midwives, and registered nurses under the Medical Service Act of South Korea [5]. Health care providers were defined as other licensed or certified health care–related personnel, such as nurse aides, medical technologists, and physical therapists. Finally, we defined ordinary persons as defendants without a health care–related professional license. Legal rulings were classified as imprisonment without suspension, fines, suspended sentences, suspension of imposition of sentence, acquittal, or other. Sentencing factors (eg, settlement with victims and self-reflection) were categorized according to the guidelines of the Sentencing Commission of South Korea [29].

The variables were labeled and classified into legal and medical sections for each case. Before labeling the variables, ontologies were built for the individual sections, and a labeling page was set up using Label Studio (HumanSignal) [30]. To ensure the validity of the results, the researchers independently labeled the variables for each case (181 samples) and then cross-checked them. The section on legal facts and rulings was labeled by a PhD in Engineering at LBox (MC), which provides a cloud-based service to support the digitization of legal practices by processing court documents and records in South Korea [21]. A PhD in Medical Law and Ethics (MJK) verified the medical sections to ensure accuracy.

Statistical analyses were performed using SPSS (version 27.0; IBM Corp) and SAS (version 9.4; SAS Institute Inc). We first conducted descriptive analyses at the case, charge, defendant category, and ruling entry levels. We then examined annual case counts from 2005 to 2023 to visualize the temporal distribution of UMP cases prosecuted. Because the dataset was conditioned on prosecuted cases with at least one UMP-related conviction, inferential analyses were framed as exploratory models of sentencing patterns among detected, prosecuted, and adjudicated cases rather than as estimates of underlying UMP prevalence, probability of offending, probability of conviction, or guilt. For the primary inferential analysis, we fit a multinomial logistic regression using the principal legal ruling recorded in each case as the dependent variable. Suspended sentence was used as the reference, and suspension of imposition of sentence, acquittal, and other rare outcomes were collapsed into a single “others” category due to small cell counts. Predictors included case composition (ordinary person only, licensed personnel only, mixed ordinary person and licensed personnel, or uncoded), involvement in the Act on Special Measures for the Control of Public Health Crimes charges, profit motive, and practice type. As secondary analyses, we fit logistic models of professionals’ involvement among category-coded cases and of health care providers’ coinvolvement among licensed personnel–involved cases as an exploratory proxy for delegation- or boundary-related scenarios.

The joint institutional review board designated by the Ministry of Health and Welfare of South Korea confirmed that the study protocol met the criteria for exemption from regular review because this research was based on literature that was publicly accessible by any individual and owing to the absence of direct or indirect involvement of human materials, animal experiments, or pathogens (P01-202406-01-025).

A total of 1532 criminal cases were analyzed. The annual distribution of these cases from 2005 to 2023 is illustrated in Figure 2, showing the longitudinal trends in UMP prosecutions. Annual case counts remained low until 2011, increased sharply from 2012, peaked at 203 cases in 2013, remained high until 2019, and steadily declined to 32 cases by 2023.

All cases contained 1718 charge entries, of which 987 (57.5%) were related to violations of Article 5 of the Act on Special Measures for the Control of Public Health Crimes, whereas 731 (42.5%) were related to violations of Article 27(1) of the Medical Service Act. Profit motive was coded in 91.6% (1404/1532) of cases, indicating that the vast majority of prosecuted UMP cases were commercially motivated. At the legal ruling level (n=2004), suspended sentences were most common (1261/2004, 62.9%), followed by fines (421/2004, 21%), imprisonment without suspension (209/2004, 10.4%), and acquittal (57/2004, 2.8%). Analysis of sentencing factors revealed that payment or deposit for victims was recorded in 72.4% (63/87) of cases, settlements with victims were recorded in 75.4% (101/134) of cases, and self-reflection was recorded in 95.6% (586/613) of cases. Prior similar-offense counts were coded for 26.9% (412/1532) of cases, with a mean of 0.8 prior similar offenses per coded case (SD 1.2; range 0-12), and prior different-offense counts were coded for 11.3% (173/1532) of cases, with a mean of 0.3 prior different offenses per coded case (SD 0.6; range 0-5; Table 1).

Defendant category coding was available for 97.8% (1499/1532) of cases, yielding 1716 defendant category entries as some cases involved more than one defendant category. Ordinary persons accounted for 75.4% (1294/1716); whereas health care professionals accounted for 15.4% (264/1716); and health care providers accounted for 9.2% (158/1716), including nurse aides, medical technologists, or physical therapists. Among health care professionals, physicians were the most frequent offenders (166/1716, 9.7%), followed by doctors of Korean medicine (41/1716, 2.4%) and dentists (35/1716, 2%). At the case level (n=1532), general UMP was the most common practice type (1056/1532, 68.9%), followed by unlicensed Korean medical practice (302/1532, 19.7%) and unlicensed dental practice (175/1532, 11.4%). Notably, one case was double coded for both general unlicensed medical and Korean medical practice (Table 2).

The distribution of legal rulings varied by defendant category. In the cross-tabulation of legal rulings and defendant categories among 1499 category-coded cases, suspended sentences (1129/1294, 87.2%) were the most frequently reported among ordinary person–involved cases. In contrast, fines were most frequently observed among physician-involved cases (130/166, 78.3%), followed by dentist-involved cases (26/35, 74.3%) and doctor of Korean medicine–involved cases (31/41, 75.6%; Table 3).

To address the determinants of legal rulings, we fit an exploratory multinomial logistic regression model. Compared with ordinary person–only cases, licensed personnel–only cases were more likely to result in fines (adjusted relative risk ratio [aRRR] 2.13, 95% CI 1.36-3.32) and imprisonment without suspension rather than suspended sentences (aRRR 1.79, 95% CI 1.02-3.16). Mixed ordinary person and licensed personnel cases showed even stronger associations with fines (aRRR 3.56, 95% CI 1.62-7.81) and imprisonment without suspension (aRRR 2.75, 95% CI 1.18-6.40). Act on Special Measures for the Control of Public Health Crimes–related cases were comparatively more likely to end in suspended sentences than in fines (aRRR 0.02, 95% CI 0.01-0.03) or imprisonment without suspension (aRRR 0.55, 95% CI 0.37-0.83; Table 4).

Table 5 illustrates the secondary exploratory logistic model among 1499 category-coded cases. Professional involvement was less common in Act on Special Measures for the Control of Public Health Crimes–related cases (adjusted odds ratio [aOR] 0.10, 95% CI 0.06-0.14), profit-driven cases (aOR 0.47, 95% CI 0.28-0.77), and Korean medical practice cases (aOR 0.09, 95% CI 0.05-0.16).

In a further exploratory model restricted to 248 cases involving at least one health care professional, physician involvement, compared with involvement of other health care professionals, was associated with health care provider coinvolvement (aOR 4.67, 95% CI 2.40-9.10), a pattern consistent with possible delegation-related or boundary violation scenarios.

This retrospective study used end-to-end neural systems to analyze 1532 criminal precedents and demonstrate the public health utility of legal infoveillance for prosecuted UMPs in South Korea. Three major findings emerged. First, commercially motivated offending predominated, with 91.6% (1404/1532) of cases coded as profit driven. Second, suspended sentences were the most common judicial outcome. Third, UMPs were not confined to ordinary persons without licenses; licensed health care professionals and providers were also substantially represented in the prosecuted case set. Taken together, these findings suggest that UMPs in South Korea should be understood not only as lay “quackery” but also as a problem of regulatory boundary violation within and around formal health care settings.

Health care professionals, particularly physicians, constituted the largest subgroup among them, and fines were common among professional-involved cases. The exploratory inferential analyses further suggested that licensed personnel–only and mixed ordinary person and licensed personnel cases were more likely than ordinary person–only cases to result in fines or imprisonment without suspension rather than suspended sentences. Although judicial records cannot reveal underlying institutional arrangements with certainty, this pattern is consistent with broader concerns about scope-of-practice breaches and commercially driven delegation or task shifting [4,31,32]. In this context, the coinvolvement of physicians and health care providers should be interpreted cautiously as a signal of possible boundary violation scenarios rather than as direct proof of unlawful delegation.

Our findings also help explain why UMPs may persist even in a setting with universal health coverage. In South Korea, the persistence of prosecuted UMPs is unlikely to be explained solely by lack of access to formal care [3]. Rather, the concentration of profit-driven cases and the involvement of licensed actors suggest that legal ambiguity, commercial incentives, and inadequately deterred noncompliance may all contribute to the continuing presence of UMPs [4]. The predominance of suspended sentences and fines may further indicate that many courts viewed these offenses as serious but still compatible with conditional leniency, which may weaken deterrence when the underlying conduct is economically rewarding.

From a methodological standpoint, this study supports legal infoveillance as a supplementary surveillance approach for hidden health-related misconduct. Because UMPs are clandestine and rarely captured directly in routine public health or administrative datasets [19], digitized court decisions provide a rare observational window into detected, investigated, prosecuted, and adjudicated events [18,33]. The framework is potentially transferable to other jurisdictions, but its usefulness will depend on the degree of judicial digitization [14], public access to court records [13], and availability of natural language processing tools capable of handling jurisdiction-specific legal language [34].

The findings need to be carefully interpreted in light of several limitations. Most importantly, the dataset represents prosecuted cases rather than the true prevalence of UMPs. This creates selection and detection bias [33] and makes the observed patterns sensitive to enforcement practices, prosecutorial priorities, and reporting behavior [35]. Accordingly, the results describe the enforcement-visible portion of UMPs rather than the absolute population burden. Therefore, these models should be read only as exploratory descriptions of sentencing patterns among detected, prosecuted, and adjudicated cases; they do not estimate the underlying prevalence of UMPs, the probability that an individual will offend, or the probability that a reported or prosecuted case will result in conviction. Formal interannotator agreement metrics were not calculated, and some complex judicial features had to be simplified into structured variables. Nevertheless, this study narrows an important empirical gap by showing that large-scale legal text analysis can generate population-relevant evidence from otherwise difficult-to-study judicial corpora [24,36].

This study demonstrates that ML-assisted analysis of criminal court precedents can serve as a useful supplementary method for monitoring hidden UMPs. In South Korea, prosecuted UMPs were predominantly profit driven and involved both ordinary persons and licensed personnel, with exploratory analyses suggesting distinct judicial outcome patterns when health care professionals were involved. These findings support stronger oversight of scope-of-practice and delegation-related violations and reinforce the value of judicial records as a practical source of public health intelligence. Future research should build on this descriptive foundation with richer inferential modeling, improved annotation reliability metrics, and cross-jurisdictional comparisons of how legal design and enforcement shape the detected profile of UMPs.