Semantic Mapping of Public Health Legislation and Emergency Legal Responses Using Natural Language Processing Techniques

Abstract

Background

Public-health legislation is heterogeneous across jurisdictions, creating challenges for reproducible policy surveillance.

Methods

Six publicly available policy-surveillance datasets from the LawAtlas Policy Surveillance Portal were consolidated and analyzed. After excluding records coded as “No bills” and collapsing repeated longitudinal snapshots within source domains, the data comprised 4328 snapshot rows, 2126 jurisdiction-by-bill source-domain records, and 1534 unique bills. Observed legal coverage extended from 6 August 2020 to 20 May 2022. Semantic mapping used TF-IDF, truncated singular-value decomposition, and K-Means clustering. Records containing at least 10 alphabetic tokens formed the semantic-mapping sample (n = 1875). Cluster number was selected using the maximum cosine silhouette coefficient across k = 2–12. Independent validation used the six curated LawAtlas source-domain memberships as multilabel targets for 821 unique bills; unsupervised cluster assignments were not used as supervised ground truth.

Results

The maximum silhouette coefficient was 0.336 at k = 12, with five-seed stability yielding a mean pairwise adjusted Rand index of 0.718 (SD = 0.069). In the held-out validation set (n = 164), Logistic Regression achieved macro-F1 = 0.543 and micro-F1 = 0.752; five-fold cross-validation macro-F1 was 0.551 ± 0.039.

Conclusion

The findings support transparent semantic mapping and source-domain classification of public-health-law surveillance records. Temporal findings are descriptive, and operational use requires human legal review and continuing performance monitoring.

Introduction

Laws, regulations, executive orders, and policy guidance define who may act, when, where, and through which mechanisms, in the prevention, detection, and control of disease, injury, and environmental hazards.^1,2 Between January 2020 and mid-2022, almost every US state introduced bills that would limit, shift, or expand the powers of governors and state health officials to issue emergency orders, and dozens of states enacted such legislation into law.³ For treating law as a measurable determinant and an object of systematic analysis.⁴ The volume, heterogeneity, and jurisdictional fragmentation of public-health legal texts nevertheless create substantial challenges for systematic surveillance.

Within this framework, LawAtlas formalises policy surveillance by systematically collecting and coding public-health laws across jurisdictions.⁵ Computational legal analysis has developed rapidly, while natural language processing has expanded across both legal and public-health applications; domainspecific models and benchmark datasets demonstrate the value of specialized text representation and classification.^6–9 The COVID-19 pandemic intensified the need for scalable legal surveillance as states adopted divergent reforms affecting emergency powers and public-health authority.^3,10,11 Computational analyses of pandemic-policy texts in other settings further show that text-mining approaches can recover interpretable policy structures.^12,13

This study addresses that gap by applying a reproducible sparse-text pipeline to six public-health-law surveillance exports. The analysis uses TF-IDF, truncated singular-value decomposition (SVD), K-Means clustering, cosine-similarity mapping, and an independent multi-label source-domain classification task. Unsupervised clusters are used only for semantic discovery and are not treated as supervised ground truth. The analysis addresses four questions: (i) which semantic structures organise the LawAtlas public-healthlaw surveillance records; (ii) how are the six source domains related in semantic space; (iii) what descriptive coverage patterns are present during the observed 2020–2022 surveillance period; and (iv) to what extent can source-text summaries reproduce independently coded source-domain membership without using unsupervised cluster assignments as supervised targets?

The primary contribution of this study is substantive rather than the novelty of the individual NLP algorithms. It applies semantic surveillance to a fragmented body of US public-health law to identify recurring legal themes, compare jurisdictional policy content, and examine changes in the observed composition of emergency-response provisions. TF-IDF, truncated SVD, K-Means, and classical one-vs-rest classifiers are established methods; their value in the present study lies in enabling a transparent and reproducible semantic analysis across multiple public-health-law surveillance domains. The resulting semantic map is intended to support human-led assessment of emergency preparedness, pre-emption, governance risk, and policy diffusion rather than to replace legal interpretation.

Systematic Search Supporting Study Positioning

Toposition the study within the legal-NLP, legalepidemiology, and computational-policy literature, we conducted a structured search between 1 February and 30 March 2025. Databases queried were PubMed, Web of Science, Scopus, ACL Anthology, arXiv (cs.CL), Google Scholar, and SSRN. Search strings combined the terms (“LawAtlas” OR “legal epidemiology” OR “public-health law” OR “public health emergency”) AND (“topic modelling” OR “topic modeling” OR “BERTopic” OR “LDA” OR “Legal-BERT” OR “NLP” OR “transformer”). Inclusion criteria were: English-language, peer-reviewed or preprint studies published from January 2015 through March 2025 that applied a documented NLP pipeline to a US or international public-health legal corpus. The structured search did not identify a directly comparable study applying a traceable semantic-clustering workflow with independently defined source-domain validation to this combination of US public-health-law surveillance domains.

The remainder of the paper is organised as follows. Literature Review reviews the literature on legal NLP, publichealth legal epidemiology, and computational analysis of pandemic policy. Methodology details the dataset, the eight-stage methodology, and the mathematical formulation of each modelling step. Results presents corpus construction, cluster-number diagnostics, sourcedomain semantic similarityindependent classification performance, and a coverage-aware temporaldescription. Discussion discusses theoretical and practical implications, limitations, and a research agenda. Conclusion concludes.

Literature Review

Natural Language Processing in the Legal Domain

Recenscholarship describes legal natural language processing as a rapidly developing field encompassing document classification, information extraction, summarisation, retrieval, legal question answering, and related tasks.⁶ Domainadapted models such as Legal-BERT and benchmark datasets such as LexGLUE demonstrate.

The value of legal-specific language representations, while LEDGAR established a large multi-label corpus of legal provisions for textclassification research.^8,9,14 Political-text classification studies further show that pretrained language models and transfer-learning approaches can reduce annotation requirements while maintaining useful classification performance.^15,16 Transformer-based similarity analysis has also been applied to legal court documents,¹⁷ and BERT-based policy-text classification has been demonstrated in non-public-health policy corpora.¹⁸ These studies provide a methodological basis for computational legal analysis but do not themselves provide a provenance-aware semantic mapping framework for multi-domain US public-health-law surveillance.

Public Health Law, Legal Epidemiology, and Policy Surveillance

Legal epidemiologyexamines law as a measurable factor inpopulation health and provides a methodological basis for legal mapping, policy surveillance, and impact evaluation.^1,2,5 Recent scholarship has documented substantial changes in US stateemergency powers and public-health authority during and after the COVID-19 pandemic, including legislative restrictions on executive and health-agency action and broader debates over the appropriate balance between emergency authority and individual liberty.^3,10,11,19 Related legal scholarship has emphasized that emergency civil-liberty safeguards.²⁰ The COVID Analysis and Mapping of Policiesdataset further demonstrates the value of large- structured policy data for comparative pandemic-policy research.²¹ Computational studies of pandemic-policy textsoutside the United States likewise show that text-mining methods can recover interpretable policy structures from large policy corpora.^12,13

NLP Methods Relevant to Public Health Legislation

Computational legal surveillance can be separated into three complementary tasks: semantic representation, unsupervised structure discovery, and supervised classification. Sparse TF-IDF representations remain useful when transparency and traceability are priorities, while dimensionalityreduction techniques can provide more compact representations for clustering and similarity analysis. KMeans offers a reproducible partitioning framework whose structural adequacy can be assessed using internal cluster-validation criteria. Legal-NLP research has also demonstrated the value of domain-specific language models and transfer-learning approaches, although reported performance depends strongly on label quality, domain match, and source-corpus composition.^{6,8,9,14–18} For public-health-law surveillance, semantic similarity can support the comparison of related legal records and policy domains.¹⁷ The present study therefore prioritizes transparent TF-IDF/SVD representation and independently defined source-domain validation rather than treating unsupervised semantic groups as ground-truth classes.

Research Gap and Positioning

The literature reviewed reveals three converging gaps. First, although LawAtlas provides systematically coded public-health-law data, the structured search did not identify a directly comparable study applying a reproducible semantic mapping workflow with independently defined source-domain validation to a multi-domain LawAtlas public-health-law corpus. Second, while legal NLP has produced competent, domain-specific models, most applications have been developed around case law, contracts, or non-US corpora rather than US state public-health legislation. Third, the post-COVID legal literature does not provide a reproducible quantitative tool to track the spread and nature of new restrictions on emergency authority. The present study addresses these gaps by combining traceable LawAtlasderived surveillance exports with TF-IDF/SVD semantic representation, K-Means clustering, source-domain similarity analysis, and an independently defined multi-label source-domain validation task. Representative studies and the specific gap addressed by the present analysis are summarized in Table 1.

Table 1.

Comparison of Representative Prior Studies with the Present Work

Study	Domain	Method	Dataset	Gap Addressed Here
Arai et al 2025.6	Legal NLP survey	PRISMA-style review	127 included studies	No application to LawAtlas/ publichealth law
Chalkidis et al 2020.8	Legal text classification	Legal-BERT	LEDGAR/LexGLUE	Not tested on US state public-health statutes
Oliveira & Sperandio Nascimento 2025.17	Legal-document similarity	NLP/transformer similarity analysis	Legal court documents	Not focused on public-health-law surveillance or provenance-aware multi-domain mapping
Liu et al 2023.12	COVID-19 policy	PMC-AE text mining	301 Chinese policies	Limited transformer integration; non-US
Davis et al 2023.3	Public-health emergency law	Narrative legal review	State PHE laws (2020–2022)	No NLP/quantitative semantic analysis
Mello and Gostin 2023.19	Emergency-powers reform	Normative legal analysis	State legislation 2020–2023	No machinereadable semantic map
Present study	US public-health-law surveillance	TF-IDF + truncated SVD + K-Means +	Six LawAtlas policysurveillance datasets;	Reproducible semantic mapping
		Independent multilabel classification	4,328 longitudinal snapshots, 2,126 domain-bill records, and 1,534 unique bills	With noncircular source-domain validation and coverage-aware temporal description

Notes: LawAtlas = Center for Public Health Law Research policy-surveillance platform; PHE = public-health emergency. Reference.⁶ Is a survey rather than an empirical study. Bold text identifies the present study.

Theoretical Framework

The analytical framework integrateslegal epidemiology policy surveillance, anddistributional semantics. Legal epidemiology treats law as a measurable factor in population health, while policy surveillance emphasizes systematic collection and comparison of legal rules across jurisdictions.^1,2,5 The policy-governance perspective recognizes that emergency-law reforms may alter the allocation and limits of public-health authority.^3,19 Distributional semantics represents legal text in vector space, enabling clustering and similarity-based comparison in computational legal analysis.¹⁷ In the present framework, LawAtlas records and jurisdictional metadata enter a provenance-aware analytical pipeline; TFIDF/SVD creates the semantic space; K-Means identifies descriptive groupings; cosine similarity characterizes relationships among source domains; andindependently defined source-domain labels provide a separate validation task. The resulting semantic map supports comparative, policy surveillance and humanled risk assessment rather than causal or autonomous legal inference. The resulting analytical framework is summarized in Figure 1.

Figure 1.

Conceptual framework for reproducible semantic mapping of six LawAtlas public-health-law surveillance datasets. Source records undergo provenance auditing and deterministic deduplication, TFIDF/SVD semantic representation, K-Means semantic mapping, and separate multi-label source-domain validation before human-led policy-surveillance and risk-management use.

Methodology

Research Design

We adopted a quantitative secondary computational design combining provenance auditing, unsupervisedsemantic clustering, source-domain similarity analysis, and independent multilabel classification. The design is appropriate because the research questions concern semantic structure within the LawAtlas surveillance records, relationships among the six source domains, descriptive coverage during the observed 2020–2022 period, and the ability oftext summaries to reproduce independently defined source-domain membership. All analyses use secondary, fully anonymised, publicly available data and therefore do not require institutional review board (IRB) approval under 45 CFR 46.104.

Dataset and Source

The primary data source was the LawAtlas Policy Surveillance Portal, maintained by the Center for Public Health Law Research at Temple University (https://lawatlas.org). Six publicly available policy-surveillance datasets were analyzed: Strengthening Public Health; Regulating Public Health Measures; Public Health Authority Limits; State Limits on Enforcement of Federal Law; Public Health Preemption; and Reallocation of Public Health Authority. The archived data exports carry embedded file-version timestamps of 23–24 January 2023. Public Health Authority Limits is dated 23 January 2023, while the remaining five exports are dated 24 January 2023.

Across the six datasets, 4328 longitudinal snapshot rows were available for 51 jurisdictions. Rows whose series_title was coded as “No bills” were excluded case-insensitively. Repeated observations were collapsed within each source domain using normalized jurisdiction and series_title, yielding 2126 jurisdiction × bill × source-domain records. Cross-domain deduplication identified 1534 unique jurisdiction × bill records. Observed legal coverage extended from 6 August 2020 to 20 May 2022.

The unit of analysis for semantic mapping was one jurisdiction × bill × source-domain record. When the same bill appeared in more than one LawAtlas domain, the separate domain memberships were retained because they represented distinct policy classifications. For independent classification, records were further collapsed to one unique jurisdiction × bill with a six-label source-domain vector, preventing the same bill from appearing independently in both training and testing data. Semantic mapping required at least 10 alphabetic tokens in the constructed text, whereas independent classification required at least five alphabetic tokens in citation/caution prose.

LawAtlas is a curated policy-surveillance resource rather than a complete census of all US public-health legislation. Accordingly, the analytical corpus represents legislation captured within the six selected surveillance domains and observed coverage period. The variables, data-derived representations, and analytical roles used throughout the final workflow are summarized in Table 2. Source-file provenance and corpus-construction details are reported in Supplementary Tables S1 and S2. The archived LawAtlas exports do not retain separate portal dataset identifiers or the original retrieval timestamps; therefore, dataset titles, embedded export-version dates, and file-level cryptographic fingerprints are reported as the traceable provenance record.

Table 2.

Variables and Representations Used in the Final Analytical Workflow

Variable	Type	Source	Analytical role
Jurisdiction	Categorical, 51 jurisdictions	LawAtlas metadata	Record grouping and deduplication
Bill/series_title	String	LawAtlas record	Bill identification and deduplication
Bill introduction date	Date	LawAtlas metadata	Coverage-aware temporal description
Source-domain membership	Multi-label, 6 domains	LawAtlas dataset classification	Independent supervised target
Semantic text	String	Citation/caution prose + coded policy descriptors	Unsupervised semantic input
Validation text	String	Citation/caution prose only	Independent classifier input
TF-IDF vector	Sparse numerical vector	Computed	Text representation
SVD representation	Dense, 50 dimensions	Computed	K-Means and cosine-similarity input
Cluster assignment	Categorical, 12 clusters	K-Means output	Descriptive semantic grouping
Multi-label prediction	Six binary outputs	Logistic Regression/SVM/Complement Naive Bayes	Independent source-domain classification

Notes: Unsupervised K-Means cluster assignments were not used as supervised target labels. Independent classification targets were the six curated LawAtlas source-domain memberships.

Preprocessing

The preprocessing workflow is summarized in Figure 2. Text for semantic mapping was constructed from distinct non-empty citation and caution fields together with the coded policy-descriptor fields PHM_phmtype, PHA_glim, PHA_slim, PHA_llim, PHP_pphm, PHS_gshift, PHS_sshift, and PHS_lshift. URLs and repeated strings were removed, and jurisdiction names and generic legislative boilerplate were removed before vectorization. Records containing fewer than 10 alphabetic tokens were excluded from semantic mapping, leaving 1875 records. Because the LawAtlas data exports contain coded surveillance summaries and policy descriptors rather than a reproducible corpus of complete statutory text, the final analytical workflow does not rely on transformer truncation or a first-512-token assumption.

Figure 2.

Reproducible analytical workflow comprising acquisition of six public LawAtlas surveillance datasets, file-level provenance recording, exclusion of “No bills,” longitudinal snapshot collapse, semantictext construction, minimum-text eligibility, TF-IDF/SVD representation, K-Means cluster-number selection and stability assessment, and independent multi-label source-domain validation.

Semantic Representations

A common sparse-text semantic representation was generated for the final analysis. Term-frequency– inverse-document-frequency vectors were constructed from unigrams and bigrams using min_df = 3, max_df = 0.90, sublinear term-frequency scaling, English stop-word removal, and a maximum vocabulary of 3,000 features. The sparse TF-IDF matrix was reduced to 50 dimensions using truncated singular-value decomposition with random_state = 42, followed by L2 normalization. All TF-IDF, truncated SVD, K-Means, and classification procedures were implemented using scikit-learn.²² The TF-IDF representation follows established information-retrieval principles in which term frequency is weighted by the specificity of a term across the document collection.²³ This common semantic space was used for K-Means clustering and source-domain cosine-similarity analysis.

Mathematical and Analytical Model

TF-IDF weighting (Equation 1) and cosine similarity (Equation 2) provide the sparse baseline:

Equation 1 (TF-IDF): where is the frequency of term t in document d, N is the total number of documents, and df_t is the number of documents containing term t.

Equation 2 (Cosine similarity): Where v_i andv_j are the normalized TF-IDF/SVD representations of records d_i and d_j.

Equation 3 (K-Means objective):

where is thenormalized SVD representation of record , and is thecentroid of cluster .

Equation 4 (Silhouette coefficient):

where is the mean within-cluster distance and is the minimum mean distance from observation to another cluster. The mean cosine silhouette coefficient was used as the prespecified criterion for selecting .

Semantic Clustering and Cluster-Number Selection

K-Means models were fitted for candidate values k = 2–12 using n_init = 20, max_iter = 300, and random_state = 42. The prespecified selection rule was the maximum mean cosine silhouette coefficient; interpretability was not used to override the numerical optimum. The silhouette coefficient jointly evaluates within-cluster cohesion and separation from neighbouring clusters and was therefore used as the formal internal cluster-validation criterion.²⁴ The maximum silhouette coefficient was 0.336 at k = 12. Stability of the selected solution was evaluated by repeating K-Means with seeds 7, 17, 27, 37, and 47. Pairwise adjusted Rand indices across the five solutions had a mean of 0.718 (SD = 0.069), providing a chance-adjusted measure of agreement among alternative cluster solutions.²⁵ Complete clusternumber diagnostics and the selected cluster solution are provided in Supplementary Tables S3 and S4.

Independent Source-Domain Validation

To avoid circular validation, unsupervised cluster assignments were not used as supervised target labels. A separate multi-label classification task used the six curated source-domain memberships as independent targets. The unit of analysis was one unique jurisdiction × bill. Classification input was restricted to citation and caution prose; coded policy-descriptor fields used in constructing source-domain records were excluded to reduce target leakage. Bills containing fewer than five alphabetic tokens in the classification text were excluded, leaving 821 unique bills. A deterministic multi-label-aware 80/20 split produced 657 training bills and 164 held-out bills. Logistic Regression, Linear SVM, and Complement Naive Bayes were fitted in one-vs-rest form using TF-IDF features. Five-fold shuffled cross-validation used random_state = 42. For the model with the highest held-out macro-F1, uncertainty was estimated with 1,000 bootstrap resamples of the held-out predictions.

Exploratory Public Health Relevance Assessment

An exploratory public-health relevance assessment examined whether the descriptive clusters corresponded to recognizable governance domains such as vaccination, quarantine, emergency powers, pre-emption, and public-health authority. This assessment did not use an external validation corpus and did not test causal relationships with epidemiological outcomes. It is therefore interpreted as a substantive interpretability check rather than an independent validation analysis.

Results

Corpus Characteristics

The six LawAtlas policy-surveillance datasets contained 4328 raw longitudinal snapshot rows. After excluding “No bills” entries and collapsing repeated within-source snapshots, 2126 jurisdiction × bill × source-domain records remained. Cross-domain deduplication identified 1534 unique jurisdiction × bill records. A total of 1875 domain-bill records met the minimum-text criterion for semantic mapping. Observed source coverage extended from 6 August 2020 through 20 May 2022. Source-level raw and deduplicated record counts are shown in Figure 3 and summarized in Table 3.

Figure 3.

Raw longitudinal snapshot rows and unique jurisdiction × bill records across the six LawAtlas policy-surveillance datasets. Orange bars show raw snapshot rows and green bars show records after within-source snapshot collapse.

Table 3.

Corpus Construction and Analytical Sample Sizes

Statistic	Value
Raw longitudinal snapshot rows	4328
Jurisdictions represented	51
Jurisdiction × bill × source-domain records	2126
Unique jurisdiction × bill records	1534
Semantic-mapping records meeting text criterion	1875
Unique bills eligible for independent validation	821
Training bills	657
Held-out bills	164
Observed legal coverage	6 Aug 2020–20 May 2022
Source domains	6
Selected K-Means clusters	12
Maximum cosine silhouette coefficient	0.336
Five-seed mean adjusted Rand index	0.718 (SD = 0.069)

Notes: Counts reflect deterministic within-source snapshot collapse and cross-domain deduplication. The semantic-mapping sample includes domain-bill records meeting the minimum-text criterion; the validation sample was independently deduplicated at the unique jurisdiction × bill level.

Semantic Clustering and Cluster-Number Selection

Across candidate solutions k = 2–12, the cosine silhouette coefficient increased from 0.159 at k = 2 to a maximum of 0.336 at k = 12. The prespecified numerical rule therefore selected the 12-cluster solution (Figure 4). Stability across five alternative random seeds yielded a mean pairwise adjusted Rand index of 0.718 (SD = 0.069). The 12 clusters represented recurring semantic structures related to quarantine and isolation, vaccination measures, emergency-order scope restrictions, public-health emergency provisions, business measures, COVID-19 vaccine restrictions, place-of-worship measures, authority and contacttracing provisions, vaccination-status requirements, governor emergency powers, duration and termination limits, and mask/school/testing measures. Cluster sizes and highest-weight terms are reported in Table 4 and Figure 5.

Figure 4.

K-Means cluster-number selection using cosine silhouette coefficients across k = 2–12. The maximum coefficient was 0.336 at k = 12, which was retained according to the prespecified numerical criterion.

Table 4.

K-Means Semantic Clusters in the Final 12-Cluster Solution

Cluster	Descriptive Label	n	Share (%)	Highest-weight TF-IDF Terms
1	Quarantine and isolation	38	2.0	Quarantine; isolation; quarantine measures; isolation measures; business
2	Vaccination measures	285	15.2	Vaccine measures; vaccine; measures; vaccine administration; administration
3	Scope restrictions on emergency orders	175	9.3	Scope emergency; scope; restricted; emergency order; order
4	Public-health emergency provisions	148	7.9	Health; public health; health emergency; emergency; declared
5	Business-related measures	64	3.4	Business measures; business; measures; vaccine; school measures
6	COVID-19 vaccine restrictions	158	8.4	COVID; COVID vaccine; vaccine; vaccine measures; immunization
7	Place-of-worship measures	76	4.1	Worship; worship measures; place of worship; measures; business
8	Authority and contact-tracing provisions	361	19.3	Vaccine; entity; measures; tracing; contact tracing
9	Vaccination-status requirements	121	6.5	Vaccination; COVID vaccination; COVID; vaccine; proof
10	Governor emergency powers	147	7.8	Governor; emergency; powers; emergency powers; order
11	Duration and termination limits	184	9.8	Emergency order; order; emergency; duration; limited
12	Mask, school, and testing measures	118	6.3	Mask; mask measures; school; school measures; testing measures
Total		1875	100.0

Notes: Cluster names are descriptive interpretations of the highest-weight TF-IDF terms. They are not externally established legal categories and were not used as supervised classification targets.

Figure 5.

Distribution of the 1875 semantic-mapping records across the selected 12-cluster K-Means solution. Cluster names are descriptive interpretations and not externally established legal classes.

Semantic Similarity and Inter-Topic Structure

Source-domain centroid similarities were calculated in the normalized TF-IDF/SVD space. The strongest relationship was between Regulating Public Health Measures and Public Health Preemption (cosine similarity = 0.913). Strengthening Public Health was moderately related to Regulating Public Health Measures (0.407), whereas most other domain pairs showed lower semantic overlap. The complete centroid-similarity matrix is shown in Figure 6 and Supplementary Table S5. A two-dimensional multidimensional-scaling projection of the same dissimilarities is provided in Figure 7 for visualization only.

Figure 6.

Pairwise cosine similarity among the six LawAtlas source-domain centroids in the normalized TFIDF/SVD semantic space. Full source-domain names are shown.

Figure 7.

Two-dimensional multidimensional-scaling representation of source-domain centroid dissimilarities, calculated as one minus cosine similarity. The projection is used for visualization only.

Cluster Stability and Structural Interpretation

The selected 12-cluster solution remained reasonably stable across five alternative K-Means initializations, with a mean pairwise adjusted Rand index of 0.718 (SD = 0.069). Cluster sizes ranged from 38 records for the quarantine-and-isolation cluster to 361 records for the authority-and-contact-tracing cluster. These clusters are interpreted as descriptive semantic groupings rather than externally established legal categories.

Independent Source-Domain Classification

On the held-out set of 164 unique bills, Logistic Regression achieved macro-F1 = 0.543, micro-F1 = 0.752, and exact subset accuracy = 0.476. Its 1,000-bootstrap 95% confidence interval was 0.425–0.649 for macro-F1, 0.708–0.794 for micro-F1, and 0.402–0.555 for subset accuracy. Five-fold cross-validation macro-F1 was 0.551 ± 0.039. Linear SVM achieved held-out macro-F1 = 0.540, micro-F1 = 0.766, and subset accuracy = 0.512, while Complement Naive Bayes achieved macro-F1 = 0.406, micro-F1 = 0.738, and subset accuracy = 0.506. Logistic Regression was retained as the primary model because it produced the highest held-out macro-F1, the prespecified class-balanced performance criterion. Model-level classification performance is summarized in Table 5 and Figure 8. Complete model-level and per-domain validation results are reported in Supplementary Tables S6 and S7, with per-domain performance visualized in Figure 9. Full analytical settings are provided in Supplementary Table S8.

Table 5.

Independent Multi-Label Source-Domain Classification Performance

Model	Subset Accuracy	Macro Precision	Macro Recall	Macro-F1	Micro-F1	Hamming Loss	5-fold CV Macro-F1
Logistic Regression	0.476	0.700	0.485	0.543	0.752	0.136	0.551 ± 0.039
Linear SVM	0.512	0.758	0.451	0.540	0.766	0.125	0.527 ± 0.040
Complement Naive Bayes	0.506	0.892	0.317	0.406	0.738	0.131	0.459 ± 0.059

Notes: Classification targets were curated six-domain source memberships rather than unsupervised cluster assignments. Training n = 657 and held-out n = 164. Bold text identifies the best value for the respective metric; macro-F1 was the primary class-balanced selection criterion.

Figure 8.

Held-out performance of Logistic Regression, Linear SVM, and Complement Naive Bayes for independent multi-label source-domain classification. Macro-F1, micro-F1, and exact subset accuracy are shown for 164 held-out bills.

Figure 9.

Per-domain precision, recall, and F1-score for the held-out Logistic Regression model. Held-out positive support is shown with each source-domain label.

Coverage-Aware Temporal Description

Because the six LawAtlas policy-surveillance datasets have heterogeneous surveillance start dates and cover only August 2020 through May 2022, temporal results are reported descriptively rather than as a test of a pandemic-driven structural break. Across the 2126 jurisdiction × bill × source-domain records, 21 bills were introduced in 2020, 1464 in 2021, and 641 in 2022. Figure 10 displays these counts by source domain. Source-domain-specific annual counts are reported in Supplementary Table S9. No Mann– Kendall trend test, changepoint analysis, 2018–2024 trajectory, or early-warning inference is reported because the observed LawAtlas surveillance coverage does not provide a comparable annual denominator for those claims.

Figure 10.

Coverage-aware distribution of jurisdiction × bill × source-domain records by bill-introduction year and LawAtlas source domain. Because source surveillance windows differ, the annual composition is descriptive and is not interpreted as a common-denominator legislative-incidence trend.

Exploratory Public Health Relevance

The 12-cluster solution was substantively consistent with recognizable public-health governance domains. The recovered groups included quarantine and isolation, vaccination measures, public-health emergency provisions, governor emergency powers, duration and termination limits, business-related measures, place-of-worship measures, authority and contact-tracing provisions, vaccination-status requirements, and mask, school, and testing measures. These themes correspond to policy areas commonly examined in legal epidemiology and emergency preparedness.^3,5,10,11,19 The assessment is interpretive rather than causal and should not be understood as independent external validation.

Discussion

This study maps six LawAtlas public-health-law surveillance datasets using a reproducible sparse-text pipeline while separating unsupervised semantic discovery from independently defined source-domain validation. The principal contribution is substantive: the analysis identifies recurring patterns involving public-health authority, pre-emption, emergency orders, vaccination, quarantine, and related governance measures and demonstrates how these records can be organized for human-led legal-policy surveillance.

Theoretical Implications

The results support two substantive interpretations. First, the 12-cluster solution shows that recurring public-health-law themes can be recovered from coded surveillance text using a transparent sparse-text representation. Second, the source-domain similarity analysis indicates that some surveillance domains share substantial semantic content—most notably Regulating Public Health Measures and Public Health Preemption—while others remain more distinct. These findings concern semantic organization within the analyzed 2020–2022 LawAtlas surveillance data; they do not demonstrate a post-2020 causal shift, transformer superiority, or an independently validated legal taxonomy.

Practical Implications

Public-health agencies could use the semantic-mapping and source-domain classification workflow to prioritize incoming records for human review and to identify provisions that warrant closer assessment for possible effects on emergency-response authority. Legislators and advocacy organisations can use cosine-similarity mapping to identify analogous laws in peer jurisdictions, accelerating the diffusion of evidence-based provisions. The use of automated legal-policy surveillance must also account for the lawful scope of emergency powers, quarantine authority, movement restrictions, civil liberties, proportionality, andethical safeguards during public-health emergencies.^26–29 For healthcare risk management, the semantic map should be used as a screening and prioritisation aid rather than as an autonomous legal decision-maker. False negatives may cause agencies to overlook provisions that materially constrain public-health authority, whereas false positives may divert review resources or mischaracterise jurisdictional risk. Operational use therefore requires human legal review, documented decision thresholds, periodic performance checks against newly coded laws, monitoring for model drift as statutory language and policy priorities change, and versioned audit records linking each output to its source record and model version.

Limitations

Several limitations qualify the findings. First, the analysis is restricted to six LawAtlas policy-surveillance datasets and therefore reflects the scope and coding architecture of those domains rather than a complete census of US public-health legislation. Second, the archived exports contain coded record text, citations, cautions, and policy descriptors rather than a reproducible corpus of complete statutory text. Third, surveillance coverage extends from 6 August 2020 to 20 May 2022 and differs across source domains; consequently, the descriptive annual counts cannot establish a 2018–2024 legislative trajectory or a pandemic-driven structural break. Fourth, source-domain sizes are highly imbalanced, and held-out performance for the smallest domains should be interpreted cautiously. Fifth, LawAtlas source-domain membership provides an independent target relative to the unsupervised clusters but is not equivalent to an external expert-coded legal gold standard. Sixth, no independent external expert-labelled corpus was included in the final analysis; therefore, generalizability beyond the LawAtlas policy domains was not directly evaluated. Finally, the study is descriptive and does not estimate causal effects of legal provisions on population-health outcomes. Future work should evaluate the workflow using independently coded full-text legal corpora, broader time windows, and prospective model-drift monitoring.

Ethical Considerations on Dual Use

The pipeline is designed to support legitimate research, public-health surveillance, and democratic accountability. We acknowledge, however, that any tool that automates the detection and comparative mapping of legal text carries a dual-use risk: the same techniques could in principle be deployed by authoritarian or anti-democratic actors to flag and sanction jurisdictions that adopt rights-protective public-health measures, or to monitor civil-society legal advocacy. To reduce this risk, we (a) restrict our published corpus and analyses to openly licensed, government-published US statutes that are already in the public domain; (b) avoid disclosing operational details that would uniquely advantage adversarial deployment (eg, real-time scraping infrastructure or automated alerting endpoints); (c) recommend that downstream users adopt the FAIR4RS and FAIR data-stewardship principles, including transparent change-logs and access controls when working with non-public legal corpora; and (d) encourage replication only in settings governed by independent ethics review and applicable open-government law.

Researchers building on this work should consider analogous dual-use assessments at the design stage.

Future Research

Future research should extend the framework in five directions. First, pre-training a public-health-lawspecific BERT variant on the full LawAtlas corpus combined with WHO International Health Regulations documents and state administrative codes would likely raise classification accuracy and topic coherence. Second, linking semantic-map output to epidemiological outcome data (eg, CDC COVID-19 mortality, vaccination coverage) would enable causal-inference work in legal epidemiology. Third, multilingual extensions using mBERT or XLM-R could map analogous emergency-power laws internationally. Fourth, integrating sentence-level argument-mining methods would permit fine-grained extraction of legal duties, rights, and exceptions.⁶ Fifth, real-time monitoring pipelines could be deployed for sentinel surveillance of restrictive amendments to emergency-power statutes. Future monitoring studies should pre-register temporal hypotheses, for example on the Open Science Framework, before deployment to strengthen the confirmatory character of the findings.

Conclusion

Semantic mapping of six LawAtlas public-health-law surveillance datasets identified reproducible thematic structure across public-health authority limits, regulatory measures, pre-emption, authority reallocation, vaccination, quarantine, and related governance domains. The 12-cluster solution was selected transparently using the maximum cosine silhouette coefficient, while a separate multi-label analysis based on curated LawAtlas source-domain membership avoided circular use of unsupervised clusters as supervised ground truth. The findings support human-assisted legal-policy surveillance and healthcare risk-management screening within the observed 2020–2022 coverage period. They do not establish a 2018–2024 structural break, an autonomous early-warning capability, or independently validated external-corpus performance.

Funding Statement

This research received no specific grant from any funding agency in the public, commercial, or not-forprofit sectors.

Generative AI Disclosure

ChatGPT (GPT-5.6 Sol, OpenAI) was used for grammar correction, copy-editing, methodological wording, and assistance in preparing reproducibility and figure-generation code. The authors independently reviewed the analytical workflow and checked the generated numerical outputs against the underlying.

LawAtlas data exports and analysis scripts. Authors confirm the originality and accuracy of all manuscript content and take full responsibility for the integrity of the whole content, including the accuracy of references. Authors have checked the terms of use for the specific AI tool used and therefore confirm suitability for publication.

Data Sharing Statement

The data analyzed in this study arepublicly available through the LawAtlas Policy Surveillance Portal maintained by the Center for Public Health Law Research at Temple University (https://lawatlas.org. The analysis included six policy-surveillance datasets: Strengthening Public Health; Regulating Public Health Measures; Public Health Authority Limits; State Limits on Enforcement of Federal Law; Public Health Preemption; and Reallocation of Public Health Authority. Archived export-version timestamps are 23 January 2023 for Public Health Authority Limits and 24 January 2023 for the remaining five datasets. The reproducibility materials include the source-data manifest, SHA-256 file fingerprints, preprocessing rules, model parameters, random seeds, train/test assignments, numerical outputs, and scripts used to generate the reported tables and figures.

The link of the reproducibility material were list as follow:https://doi.org/10.5281/zenodo.21992973.

Ethics Approval

This study analyses secondary, fully de-identified, publicly available legal-text data and does not involve human subjects within the meaning of 45 CFR 46.102. Accordingly, institutional review board approval was not required. No individual-level data, personal identifiers, or protected health information were processed at any stage of the analysis.

Informed Consent Statement

Not applicable. The study does not involve human participants; informed consent is therefore not required.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare no competing financial or non-financial interests relevant to the content of this article.