We analysed individual case safety reports from 1 January 2019 to 31 March 2025 in FAERS (U.S. Food and Drug Administration, 2018; U.S. Food and Drug Administration, 2014). This period was chosen to focus on contemporary reporting patterns and to minimise structural changes present in earlier FAERS data. It also encompasses the coronavirus disease 2019 (COVID-19) pandemic, during which substantial changes in healthcare utilisation, prescribing and reporting may have affected both the mix and volume of reports; this is considered in the interpretation of findings.

The quarterly ASCII archives (DEMO, DRUG, REAC, INDI and deleted-case files) were downloaded or accessed locally and parsed using the faers R package and custom scripts. Adverse events are coded using MedDRA, and drug information is recorded in multiple free-text and structured fields.

We first excluded reports flagged as deleted in the FDA deleted-case files (U.S. Food and Drug Administration, 2018; U.S. Food and Drug Administration, 2014). Remaining reports were deduplicated by case identifier (CASEID), retaining the record with the highest CASEVERSION, in accordance with FDA recommendations (U.S. Food and Drug Administration, 2018; U.S. Food and Drug Administration, 2014). We additionally applied rule-based deduplication across CASEIDs using a key-field fingerprint (Figure 1). After identification of insomnia reactions and exclusion of sleep-related indications, the final analysis universe comprised 2,935,560 unique reports (Figure 1).

Characteristics of the final analysis universe were summarized using standard descriptive statistics: continuous variables were reported as median (interquartile range, IQR), and categorical variables were reported as counts and percentages. Summaries were generated for demographics, reporter type, and recorded route of administration, with detailed results presented in the Results section and Supplementary Tables.

Insomnia cases were defined a priori using a narrow, clinically face-valid set of MedDRA Preferred Terms (PTs) in the REAC table: insomnia, initial insomnia, middle insomnia, terminal insomnia, and early morning awakening (Jain and Jain, 2011; Van Gastel, 2018; Doufas et al., 2017). This PT cluster is not a standardized MedDRA grouping (e.g., SMQ); it was selected to prioritize specificity and reduce contamination from loosely related sleep complaints, guided by clinical face validity and prior literature on drug-associated insomnia (Jain and Jain, 2011; Van Gastel, 2018; Doufas et al., 2017).

To minimize confounding by indication (i.e., insomnia recorded as the reason for prescribing rather than an adverse reaction), we excluded reports in which the INDI table contained sleep-related indications, including insomnia/sleeplessness and broader sleep disorder terminology (e.g., sleep disorder/disturbance, dyssomnia), circadian rhythm sleep–wake disorders (including non-24-h sleep–wake disorder), parasomnias, restless legs syndrome, and narcolepsy. This exclusion was implemented using a prespecified regular-expression pattern aligned with these indication categories (Figure 1), and the full operational pattern/term list is provided in Supplementary Tables.

Drug exposures were derived from the DRUG table. We focused on parent systemic drugs administered via oral, intravenous (including unspecified intravenous and intravenous drip), intramuscular, subcutaneous, inhalation/respiratory, transdermal or nasal routes. Topical, ophthalmic, otic and other clearly local formulations were not considered parent systemic drugs for disproportionality analyses. Reports involving only non-systemic routes were excluded from the disproportionality analyses. Transdermal products were retained only when intended for systemic delivery, consistent with the route coding and product context in FAERS (Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Hammad et al., 2025).

To minimise artificial splitting of signals across salt forms and formulations, proprietary names and active ingredient strings were normalised to parent drugs using a rule-based mapping function. This function removed salt/ion terms (e.g., hydrochloride, mesylate), hydrates, release modifiers (extended-release, controlled-release, XR, SR, etc.), species descriptors (e.g., porcine) and non-alphanumeric symbols, followed by lower-casing and whitespace trimming, and when possible was informed by inspection of high-frequency strings and cross-checked against WHO ATC/DDD and FDA Structured Product Labeling ingredient lists for common salts and modified-release formulations (Evans et al., 2001; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; van der Heijden et al., 2002; Hammad et al., 2025). Fixed-dose combinations were not decomposed; such products were mapped to a composite parent label (e.g., “pseudoephedrine + antihistamine”), and signals for these parents should therefore be interpreted as relating to the combination rather than individual components. All disproportionality analyses were conducted at this parent-drug level.

We constructed 2 × 2 contingency tables for each parent systemic drug and insomnia as the event of interest (Evans et al., 2001; Bate and Evans, 2009; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002).

We selected the reporting odds ratio (ROR) as the primary disproportionality measure, with PRR reported as a complementary metric, because both are widely used, transparent, and facilitate comparison with prior pharmacovigilance studies and guidance documents (Evans et al., 2001; Bate and Evans, 2009; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002). Bayesian shrinkage metrics (e.g., Information Component or EBGM) were not included in the primary analysis to preserve interpretability and consistency with the prespecified analytic plan; evaluating robustness with alternative disproportionality frameworks may be addressed in future work. The primary analysis used drugs coded as primary suspect (PS); a secondary analysis considered drug–report pairs in which the drug was coded as a primary or secondary suspect (ANY, i.e., PS + SS). Both PS and ANY analyses were restricted to parent systemic drugs with the routes defined above.

Let a denote insomnia reports with the drug of interest, b insomnia reports without the drug, c non-insomnia reports with the drug and d non-insomnia reports without the drug within the relevant “universe” of report IDs (Evans et al., 2001; Bate and Evans, 2009; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002). To avoid contamination between analyses, event sets and drug-exposed IDs were intersected with the current universe of report identifiers for each analysis.

For inclusion in the PS set, we required a ≥5 and (a + c) ≥ 50, in line with previous FAERS and EudraVigilance work and regulatory guidance (Evans et al., 2001; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Van der Heijden et al., 2002). To reduce instability from zero cells, Haldane–Anscombe 0.5 continuity corrections were applied to all 4 cells (Evans et al., 2001; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Van der Heijden et al., 2002). For each drug we computed:

The reporting odds ratio (ROR) = (a·d)/(b·c) with standard error SE[log(ROR)] = √(1/a +1/b + 1/c + 1/d) and 95% CI = exp[log(ROR) ± 1.96 · SE];

Forest plots for the strongest signals were generated using ggplot2 on a logarithmic scale, with a vertical reference line at ROR = 1 (Figures 2, 3; Supplementary Figures S1–S6).

Sex-stratified PS analyses were conducted by repeating the above approach within female and male report subsets (Watson et al., 2019; Yu et al., 2016; de Vries et al., 2019; Visser et al., 2021; Zucker et al., 2020). For each drug we extracted the raw cell counts (aF, bF, cF, dF and aM, bM, cM, dM) and computed sex-specific RORs (ROR_F, ROR_M) with 0.5 corrections. Drugs with exposure limited to a single sex (e.g., finasteride, estrogens, testosterone) were classified as sex-specific and not subjected to interaction testing; these were tabulated separately (Supplementary Tables S3–S4b).

For drugs with non-zero exposure in both sexes and at least five insomnia reports in each sex stratum (aF ≥ 5 and aM ≥ 5), we evaluated sex-related heterogeneity using two complementary approaches:

Difference in log-ROR: logROR_diff = log(ROR_F) − log(ROR_M) with standard error SE_diff = √(SE_F² + SE_M²). A z statistic (z_diff = logROR_diff/SE_diff) and two-sided p value (p_diff) were computed under the null hypothesis of equal RORs.

For both p_diff and BD_p, multiple testing was addressed using the Benjamini–Hochberg FDR method, controlling the expected proportion of false discoveries at 5% (Benjamini and Hochberg, 1995). To reduce spurious interaction signals, Breslow–Day tests were performed only for drugs meeting this minimum stratum-specific count criterion. Accordingly, sex-stratified disproportionality estimates are interpreted as exploratory indicators of potential heterogeneity rather than as measures of sex-specific causal risk.

All analyses were conducted in R (version 4.3.2; R Foundation for Statistical Computing, Vienna, Austria) using the data.table, dplyr, stringr, ggplot2, readr and DescTools packages. The overall analysis was implemented in a reproducible pipeline, with code and non-identifiable summary outputs provided in the Supplementary Material. Reporting of methods and results was aligned, as far as possible, with guidance from the READUS-PV statement and related recommendations for transparent disproportionality analyses (European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Fusaroli et al., 2024; Hammad et al., 2025).

Figure 1 summarises construction of the analytic cohort. After removal of deleted cases, deduplication by CASEID (retaining the highest CASEVERSION), rule-based deduplication across CASEIDs using a key-field fingerprint, identification of insomnia reactions, and exclusion of sleep-related indications, the final analysis universe comprised 2,935,560 unique reports (Figure 1). Within this universe, 74,444 reports contained insomnia reactions (Figure 1). Characteristics of the final analysis universe are summarized in Table 1, and a detailed breakdown of recorded routes among insomnia reports is provided in Supplementary Table S1.

In the final analysis universe, the median age was 58 years (IQR 40–71), and females accounted for 49.0% of reports (Table 1).Route information was variably recorded; among reports contributing to the systemic-route analytic universe with a recorded route, oral administration was most common, followed by parenteral routes and other/unknown systemic routes (Table 1; Supplementary Table S1).

In the PS analysis, parent systemic drugs meeting the minimum reporting thresholds were retained for signal screening. Figure 2 and Table 2 summarise the top 15 parent systemic drugs with the highest RORs for insomnia in the PS set. The strongest signals were observed for oxybate (ROR 13.56, 95% CI 10.14–18.13) and istradefylline (13.43, 9.78–18.44), followed by daridorexant (11.31, 6.02–21.24), trofinetide (10.97, 8.70–13.82) and niraparib (10.35, 9.31–11.50) (Table 2; Figure 2). Additional high-ranking signals included viloxazine, diphenhydramine + naproxen, loratadine + pseudoephedrine, flibanserin, guaifenesin + pseudoephedrine, pseudoephedrine, pimavanserin, ziprasidone, ropeginterferon alfa-2b-njft, and suvorexant (Table 2; Figure 2). These disproportionality estimates represent signals of disproportionate reporting rather than incidence or causal risk.

In the ANY analysis, which considered drugs coded as primary or secondary suspect, point estimates of insomnia association were generally similar or modestly attenuated compared with the PS analysis. Chlorhexidine showed an unusually high ROR driven by a relatively small cluster of insomnia reports (Supplementary Figure S6). Importantly, disproportionality screening in this study was restricted to the prespecified systemic-route framework; therefore, reports involving only clearly non-systemic routes were not part of the analytic universe for disproportionality analyses. The additional route stratification shown for chlorhexidine is descriptive and based on the recorded route field among included reports, highlighting that this cluster was enriched in potentially systemic, ambiguous, or poorly specified routes, whereas clearly topical-only contexts were uncommon within the included set. Given the small number of contributing reports, potential route misclassification, and atypical exposure context, this observation should be interpreted cautiously as hypothesis-generating (U.S. Food and Drug Administration, 2018; U.S. Food and Drug Administration, 2014; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Fusaroli et al., 2024; Hammad et al., 2025).

Sex-stratified PS analyses yielded estimable RORs for a substantial proportion of drugs in both female and male reports. Forest plots of the strongest sex-specific PS signals are presented in Figure 3A (females) and Figure 3B (males). In females, leading signals included oxybate, trofinetide, niraparib, and several combination products containing pseudoephedrine (Figure 3A). In males, leading signals included dupilumab, montelukast, finasteride, sofosbuvir, es(z)opiclone, vilazodone, and stimulant-related exposures (Figure 3B). Drugs with exposure limited to a single sex (e.g., finasteride, estrogens, testosterone) were classified as sex-specific and not subjected to interaction testing; these are reported separately in Supplementary Tables S3–S4b.

Among drugs with exposure in both sexes and sufficient stratum-specific counts for interaction testing, formal heterogeneity testing did not identify robust sex-related heterogeneity after Benjamini–Hochberg FDR correction. Consistent with this, the sex-heterogeneity volcano plot shows that essentially no drug–insomnia pairs exceed the FDR significance threshold (Supplementary Figure S7). These results should be interpreted as exploratory and hypothesis-generating.

This study has several strengths. First, we used a prespecified and transparent workflow to extract and preprocess FAERS data, including removal of deleted cases, CASEID/CASEVERSION consolidation, and additional rule-based deduplication across CASEIDs, followed by reproducible generation of summary tables and figures (U.S. Food and Drug Administration, 2018; U.S. Food and Drug Administration, 2014; Fusaroli et al., 2024). Second, we implemented parent-drug normalisation to reduce artificial fragmentation of signals across salts, hydrates, proprietary names and modified-release formulations, which is a common challenge in spontaneous reporting systems with heterogeneous drug naming (Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Hammad et al., 2025). Third, the analyses were restricted to systemic routes of administration to improve biological plausibility and interpretability of insomnia SDRs, and we applied consistent inclusion thresholds and continuity corrections to mitigate instability in sparse cells (Evans et al., 2001; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002). Fourth, we used a narrow, clinically focused insomnia event definition and excluded reports with sleep-related indications, aiming to prioritise specificity and reduce conflation of underlying sleep disorders or symptom-directed prescribing with adverse reactions (Jain and Jain, 2011; Van Gastel, 2018; Doufas et al., 2017). Finally, sex-stratified analyses combined stratified ROR estimation with formal heterogeneity testing and multiplicity control, aligning reporting and interpretive principles with the READUS-PV framework (Benjamini and Hochberg, 1995; Fusaroli et al., 2024).

Important limitations are inherent to spontaneous reporting data and to disproportionality analyses. FAERS is subject to under-reporting, stimulated reporting, and temporal changes in reporting behaviour, all of which can affect disproportionality estimates independently of any causal relationship (Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Hammad et al., 2025). Disproportionality metrics such as the ROR and PRR quantify SDRs within the reporting database rather than incidence or absolute risk in treated populations, and they cannot establish causality (Evans et al., 2001; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Fusaroli et al., 2024; Hammad et al., 2025). Accordingly, all results should be interpreted as hypothesis-generating signals rather than as definitive risk estimates.

Confounding by indication and comorbidity is a major concern for insomnia. Insomnia is influenced by multiple clinical and contextual factors (e.g., pain, anxiety/depression, cardiometabolic disease, substance use, and concurrent medications), many of which are incompletely captured and not fully adjustable in spontaneous reporting data (Zhu and Lv, 2025b). Insomnia reports may therefore reflect underlying disease activity, symptom burden, or care pathways rather than a drug effect per se (Jain and Jain, 2011; Van Gastel, 2018; Doufas et al., 2017; Natter et al., 2021). Although we excluded reports with sleep-related indications to reduce symptom-directed prescribing bias, indication fields in FAERS may be incomplete, inconsistently coded, or missing, and residual confounding is likely (Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Hammad et al., 2025). In addition, concomitant medications were not adjusted for analytically; because many reports include multiple suspect and concomitant drugs and role assignment may vary across reporters, disproportionality methods alone cannot reliably disentangle independent contributions of co-administered therapies (Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Hammad et al., 2025).

Limited availability and quality of covariates further constrain inference. Key demographic and clinical factors—most notably age, comorbidity burden, disease severity and duration—cannot be comprehensively adjusted for in routine disproportionality analyses, and age is particularly relevant because both insomnia prevalence and prescribing patterns vary substantially across the life course (Jain and Jain, 2011; Van Gastel, 2018; Doufas et al., 2017; Natter et al., 2021; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002). Missingness and misclassification in drug identity, route, reaction coding and reporter fields may also contribute to bias (U.S. Food and Drug Administration, 2018; U.S. Food and Drug Administration, 2014; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; van der Heijden et al., 2002). Although we applied systemic-route restrictions and parent-drug normalisation, route misclassification remains possible, and fixed-dose combinations were treated as composite entities, so signals for these parents should be interpreted as relating to the combination rather than individual components (Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Hammad et al., 2025). Finally, our study period includes the COVID-19 pandemic, during which changes in healthcare utilisation, prescribing and reporting could have influenced both the composition and volume of FAERS reports; these temporal factors may affect the comparability of signals across time and should be considered when interpreting results (U.S. Food and Drug Administration, 2018; U.S. Food and Drug Administration, 2014; Hammad et al., 2025).

Sex-specific analyses warrant additional caution. Sex-stratified RORs and interaction tests may be influenced by sex differences in exposure prevalence, prescribing indications and reporting behaviour, and therefore do not directly quantify biological susceptibility (Watson et al., 2019; Yu et al., 2016; De Vries et al., 2019; Visser et al., 2021; Zucker et al., 2020). Formal heterogeneity testing also requires sufficient stratum-specific counts; for less frequently reported drugs, power may be limited and estimates may be unstable despite continuity corrections (Evans et al., 2001; Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002). Consequently, both positive and null findings for sex-related heterogeneity should be interpreted conservatively, and any drug–sex combinations of interest should be evaluated in independent longitudinal datasets with validated sleep outcomes and appropriate confounding control (Bate and Evans, 2009; European Medicines Agency, 2006; Wisniewski et al., 2016; Cutroneo et al., 2023; Van der Heijden et al., 2002; Fusaroli et al., 2024; Hammad et al., 2025).