Characterization of Demographics, Drug Latency, and Mortality of Severe Cutaneous Adverse Reactions in an FDA Pharmacovigilance Database

Eric Milan Mukherjee, Dodie Park, Michelle Martin-Pozo, Elizabeth Phillips
doi: https://doi.org/10.1101/2025.03.05.25323441

Abstract

Background Severe cutaneous adverse reactions (SCARs) are rare, life-threatening conditions associated with drug exposure. These conditions differ globally in causative agents, demographics, and drug latency (time between drug administration and adverse reaction). We provide a comprehensive survey of SCAR in a pharmacovigilance database maintained by the FDA.

Methods After sanitization, the FDA Adverse Events Reporting System (FAERS) database from January 2004 to December 2003 was queried for SCAR cases. Disproportionality analyses measured associations between drugs and SCAR. Random forests were used to determine the influence of covariates on latency and mortality.

Findings A total of 56,683 SCAR cases were reported, comprising 0.33% of total reports. 49.4% were female (28008). The median age of SCAR patients was 53 (interquartile range [IQR] 32-68), with significant differences in age between phenotypes. There has been an increase in SCAR reporting over time, particularly biologics and immune checkpoint inhibitors (ICI). Over 200 drugs had positive disproportionality signals. Several cases co-reported two of SJS-TEN, DRESS, and/or AGEP, indicating the possibility of intermediate phenotypes. TTE analysis showed causative drug is the most influential variable on drug latency, followed by number of concomitant drugs. Mortality varied between SCAR and increased with age and number of concomitant drugs.

Interpretation This largest retrospective study of SCAR to date shows the variety of phenotypes, causative agents, demographic variables, drug latencies, and mortality in this population. This study demonstrates the power of using pharmacovigilance databases to study rare reactions. Continued mining of these databases, retrospective analyses of electronic health records, and prospective data can expand upon these results, better characterize variations, and improve recognition and care for patients with SCAR.

Funding National Institutes of Health, Vanderbilt University Medical Center.

Introduction

Adverse drug reactions (ADR) are one of the top 10 causes of death in the developed world, costing up to $130 billion annually in the US.1 Among the most serious ADRs are severe cutaneous adverse reactions (SCARs), including Stevens-Johnson Syndrome/Toxic Epidermal Necrolysis (SJS-TEN), drug reaction with eosinophilia and systemic symptoms (DRESS), acute generalized exanthematous pustulosis (AGEP), and generalized bullous fixed drug eruption (GBFDE). These conditions are distinguished by morphology, causative agents, systemic involvement, and latency. For instance, AGEP and GBFDE manifests within hours to days of initial drug exposure, while DRESS and SJS-TEN can take weeks to manifest.2

SCAR has incidence of 0.4 to 0.6 per million/year in the developed world, with mortality up to 70 percent in high-risk populations.2 SCAR can lead to signficant skin damage, systemic involvement, and long-term sequelae. Because of their rarity, SCAR are difficult to study at scale. The largest retrospective studies of SCAR have 300-400 cases, and the largest meta-analysis of SJS-TEN had 2917 patients.4,5 Thus, there are many unanswered questions, including intermediate phenotypes between different SCAR, variations in mortality and phenotype between causative agents, and geographic and temporal variation.

The FDA Adverse Event Reporting System (FAERS) is a valuable public database collecting information on ADRs, used for post-marketing surveillance. FAERS contains millions of reports submitted by providers, patients, and manufacturers. However, FAERS is hampered by unstructured drug names, duplication, and inconsistent detail.6 In this study, we stringently sanitized FAERS and comprehensively analyzed SCAR reports from 2004 to 2023, focusing on temporal trends, disproportionality, and potential overlap conditions. With the large dataset obtained – the largest real-world retrospective dataset of SCAR to date – we were also able to analyze drug latency and mortality using machine learning (ML) techniques, gaining key insights on time-to-event (TTE) and determinants of mortality.

Methods

Coding was done in RStudio 2024.12.0 Build 467 with R version 4.4.2, Pycharm 2024.2.0.1 with Python 3.12, PostgreSQL 16.3 (with PgAdmin IV), with GPT4o assistance. Further details are found in the online supplement.

Data Source and Preparation

Preparation is summarized in Figure 1. Quarterly ASCII files were fed into AEOLUS, which maps drugs and outcomes to standardized vocabularies and deduplicates by removing cases that match all four demographic fields (event date, age, sex and reporter country).7 Unmapped drugs and outcomes were manually mapped by two researchers (EMM and DP), resulting in 100% of rows being mapped in both tables.

Figure 1.
Figure 1. FAERS Deduplication and Sanitization.

Quarterly FAERS data files are deduplicated and combined, with drugs and outcomes mapped to standard terms. Any unmapped drugs and outcomes were manually mapped to standardized terms. From a data set of n = 17201564 unique patient IDs (PIDs), 56683 have at least one reported SCAR. Of those cases, 13633 (24.1%) had sufficient data for TTE analysis, and 13540 (23.9%) had sufficient data to analyze mortality.

Disproportionality

Contingency tables were computed for each SCAR and for SCAR in aggregate, filtered to primary suspects (PS). Reporting odds ratio (ROR), proportional reporting ratio (PRR), and information component (IC) were computed as disproportionality measures.

Co-Reporting of Multiple SCAR

Overlap between patients with SCAR and other cutaneous conditions (pemphigoid, pemphigus, fixed drug eruption [FDE], linear IgA bullous dermatosis [LABD], symmetrical drug-related intertriginous and flexural exanthema [SDRIFE], and systemic contact dermatitis) were analyzed using the binomial test under assumption of random distribution. P-values were adjusted using the Benjamini-Hochberg procedure to control false discovery rate.

Time-to-Event Analysis and Random Survival Forests

After filtering and outlier removal, 13633 SCAR cases had sufficient data for TTE analysis. Random survival forests (RSFs) are a method of analyzing TTE based on the random forest (RF) classifier. Like RFs, RSFs are bootstrapped aggregates of hundreds of decision trees, well-suited for analyzing high-dimensional data with interacting, nonlinear variables.8 To understand the effects of covariates on TTE, we generated RSFs for each SCAR and for all SCAR aggregated using the Python package scikit-survival (version 0.23.1), after a grid search for hyperparameter optimization. The covariates included were causative drug, sex, region, age (in years), number of concomitant drugs, and Boolean indicator for each SCAR (SJS-TEN, DRESS, AGEP, GBFDE) when applicable. Variable importance (VIMP) for each covariate was determined by bootstrapping. Model fit was assessed using the concordance index (C-index). Survival curves were generated for stratified covariates to analyze their effects.

Mortality Analysis

13540 SCAR cases had both TTE and mortality information. RFs were trained on all cases using age, sex, region, causative drug, TTE, and Boolean variables for each SCAR as covariates. Due to technical limitations, the top 50 causative drugs were included, and the rest were coded as “other”. Separate RFs were trained for SJS-TEN, DRESS, and AGEP, using the same covariates except the binary SCAR variables. All RFs were created after hyperparameter tuning by grid search and trained using five-fold cross-validation.

Role of the Funding Source

The funding source had no role in the work in this publication.

Results

Demographics

A total of 56,683 SCAR cases were identified, comprising 0.33% of total reports during this period. The most frequently reported was SJS-TEN (28,871), then DRESS (22,444), AGEP (6,183), and GBFDE (150). The number of reports over time, age, gender, and geographic distribution are summarized in Figure 2. Reports increase from 2004-2023, steadily starting in 2014, peaking in 2019. While the number of new SJS-TEN cases have stayed relatively flat, new DRESS cases have steadily increased, overtaking new SJS-TEN cases in 2017.

Figure 2.
Figure 2. Demographic Summary of SCAR.

(A) SCAR reporting over time in FAERS, by quarter. (Bl Age distribution by SCAR. Pairwise comparisons done by Wilcoxon signed-rank test. **p < 0.01, ***p < 0.001, ****p < 0.0001. (C) Sex Distribution by SCAR. All patients without “M”or“F” sex markers were classified as “other”. (D) Top 10 Countries by SCAR cases. Number of cases were aggregated over entire time period (2004 to 2023). US – United States, FR – France, JP – Japan, GB – Great Britain, ES – Spain, IT – Italy, DE – Germany, CA – Canada, IN – India, PT – Portugal

Age statistically differs between SCAR, with GBFDE having lowest median (46 years, IQR [35-65]), then SJS-TEN (52 years, [30-68]), DRESS (52 years, [31-66]), and AGEP (59 years, [41-73]). Each SCAR shows female predominance except GBFDE. Analyzing sex and age (Figure S1) shows that males have a bimodal age distribution, while female patients overtake males in adolescence before “catching up” in older age. Reports originate from 140 countries and territories (Figure S2). USA (15607), France (9530), and Japan (5248) contributed the most reports. SJS-TEN outnumbers DRESS in all top 10 countries except for France, Spain, and Portugal. India reports the most GBFDE (44 cases), followed by France (28), the US (15) and Iran (15).

Temporal Trends in Causative Agents

Using criteria from Evans et al (>2 cases, PRR ≥ 2, chi-squared ≥ 4), we find 237 drugs with positive signal when aggregating all SCAR together (Supplementary Data).9 When splitting by phenotype, there were 229 drugs with a positive signal for SJS-TEN, 161 for DRESS, 130 for AGEP, and 16 for GBFDE.

We sought to understand temporal trends in causative agents of SCAR (Figure 3). The top five causative agents during the span of this study (allopurinol, vancomycin, lamotrigine, ibuprofen, and carbamazepine) consistently are responsible for ∼20% of new reports. The proportion of new reports due to biologics has steadily increased over time (p < 0.0001 by Cochrane-Armitage test), comprising 8.2% of new reports in 2023, up from 2.6% in 2015 and 1.2% in 2004. Subdividing biologics by function, we note that since their 2011 FDA approval, cases associated with ICIs have increased steadily (p < 0.0001 by Cochrane-Armitage test), comprising 49.1% of reports with a biologic as the primary suspect in 2023. Conversely, while immunomodulators comprised the majority of biologic-associated SCAR in the early 2000s, their proportion has decreased with the rise of ICIs.

Figure 3.
Figure 3. Trends in SCAR Reporting.

Number of reports were aggregated by approximately-equal-frequency bins. (A) Number of SCAR reports whose primary cause is a “top 5” small molecule (allopurinol, carbamazepine, ibuprofen, lamotrigine, vancomycin), another small molecule, or a bio­ logic. (B) Number of SCAR reports subdivided by type of biologic. ICI = immune checkpoint inhibitor. (C) Proportion of SCAR report by drug category over time. Proportion of biologics has increased over time (p < 0.0001 by Cochrane-Armitage test). (D) Proportion of total biologic-related SCAR reports subdivided by type. Proportion of ICI has increased over time (p < 0.0001 by Cochrane-Armitage test)

Disproportionality Analysis

Figure 4 plots the information content (IC) of the top ≤20 drugs with >2 cases. Figure S3 shows the top 100 drugs when aggregating all SCAR together. Several well-known culprits (e.g. lamotrigine, allopurinol, and phenytoin for SJS-TEN, allopurinol, vancomycin, and lamotrigine for DRESS) are strongly associated (IC ≥ 3) with their respective reaction. Other, less-well-known associations (e.g., enfortumab and zonisamide for SJS-TEN, ciprofloxacin and valacyclovir for DRESS) have hundreds of reports and clear positive signals.10–12

Figure 4.
Figure 4. Disproportionality Analysis.

Tob ≤20 drugs for each SCAR by number of cases (with a minimum of 3 cases) are displayed. Each point represents the IC of a given SCAR combination, with the total number of cases noted just above. Confidence intervals span IC025 to IC975 Red dashed line represents IC of 3, suggesting strong disproportionality.

Interestingly, some suspects have negative IC (Figure S4), meaning they are less frequently associated with SCAR than expected based on frequencies in the entire FAERS database. Many of these (lenalidomide, adalimumab, etanercept, and ocrelizumab) are immunomodulators that have been infrequently (and controversially) implicated in SCAR in individual cases.13–15 Some hormonal medications (testosterone, levonorgestrel), psychiatric medications (methylphenidate, aripiprazole), and anti-osteoporotic agents (denosumab, zoledronic acid) also have negative IC.

Disproportionate Co-Reporting

Overlap conditions between SCAR are rare and difficult to diagnose.16 These conditions are difficult to study retrospectively since SCAR is frequently misdiagnosed and intermediate phenotypes exist. We analyzed cases reporting more than one SCAR, and included other generalized cutaneous ADR (pemphigus, pemphigoid, SDRIFE, FDE, LABD, and systemic contact dermatitis), which can be confused with SCAR, and may require additional testing to diagnose (Figure S5). A binomial model was used to assess whether co-reporting was more frequent than expected by random co-occurrence. There were 231 cases of SJS-TEN/AGEP (p = 2.37e-206), 576 of SJS-

TEN/DRESS (p < 1e-207), and 205 of AGEP/DRESS (p = 3.5e-194). Co-reporting between SCAR and non-SCAR is less pronounced, with the most significant occurring between SJS-TEN and LABD (43 cases, p = 9.65e-41), SJS-TEN and pemphigoid (63 cases, p = 1.45e-28). Pemphigoid and LABD, bullous conditions that can mimic each other, had 14 co-reporting cases (p = 1.29e-15), and pemphigus and pemphigoid had 44 (p = 2.18e-35).

Disproportionality of co-occurring SCAR is shown in Figure 5. For SJS-TEN/DRESS, allopurinol (71 cases), lamotrigine (71), and phenytoin (49) are the most frequent suspects and have IC > 3. Co-reporting with AGEP is rarer; phenytoin (26 cases), ciprofloxacin (18), and allopurinol (14) are the most frequent culprits with co-reported DRESS, all with IC > 3, and phenytoin (34 cases), amlodipine (29), and ibuprofen (13) are the most frequent suspects of co-reported SJS-TEN.

Figure 5.
Figure 5. Disproportionality Analysis and Co-reporting of SCAR.

Information component, with 95% confidence interval (IC025 to IC975) is plotted for all medications with at least 3 cases of co-reported SCAR. Inset: number of patients with co-reported SCAR, pairwise. Elements on the diagonal are the total patients with each SCAR.

Latency Varies Significantly by Causative Drug and SCAR

We conducted TTE analysis on a subset of cases (n = 13,633, Figure 6A). Latency differed by phenotype (p < 0.0001 by log-rank), with GBFDE having the shortest (3 days, IQR [2-3], n = 5), followed by AGEP (4d, [2-9], n = 1541), SJS-TEN (12d, [6-23], n = 7230), and DRESS (20d, [9-31], n = 5021). Median latency varies by causative drug (Figure 6B). For SJS-TEN, NSAIDs (ibuprofen, acetaminophen) and antibiotics (ceftriaxone, amoxicillin, azithromycin) have short latency, while checkpoint inhibitors (pembrolizumab, nivolumab), allopurinol, and some anticonvulsants (zonisamide and phenytoin) have long latency. For DRESS, ibuprofen, amoxicillin, and the kinase inhibitor vemurafenib have short latencies, while the protease inhibitor telaprevir has median latency of 42 days. For AGEP, a few drugs have long latency (vancomycin, hydroxychloroquine, terbinafine).

Figure 6.
Figure 6. Drug Latency.

(A) Survival curves of show clear variation in median TTE/latency per SCAR. P-value by log-rank test. (Bl Drug latency varies by causative drug, within each SCAR. The top 40 drugs (by number of cases) were plotted for each SCAR (GBFDE was excluded due to low number of cases). Each point represents median latency for that drug-SCAR combination, with IQR around the median. The “All Drugs” row for each SCAR shows the mean latency for each SCAR for all drugs (including those not displayed).

To better understand latency, RSFs were constructed on pooled SCAR cases, using SCAR phenotype, age, sex, region, causative drug, and number of concomitant drugs as covariates. The resulting RSF had C-index 0.7791 (Figure 7). When calculating TTE curves by SCAR phenotype from this RSF (thus controlling for all other variables), we found DRESS has the longest median latency (19.84 days, IQR [8.96, 30.95]), followed by SJS-TEN (11.82d, [5.77, 22.91], AGEP (4.19d, [1.87, 8.96]), and GBFDE (3.28d, [1.94, 6.37]). Variable Importance (VIMP) for each covariate showed that causative drug had by far the largest influence on latency, followed by age and number of concomitant medications. To better characterize the determinants of latency within each phenotype, we created separate RSFs for each, with C-indices 0.7953, 0.7977, and 0.7668 for SJS-TEN, DRESS, and AGEP. VIMPs similarly showed that causative agent had the largest influence on TTE, followed by age and concomitant drugs (Figure S6).

Figure 7.
Figure 7. Random Survival Forest Analysis of SCAR.

An RSF was trained on n = 13,633 cases of SCAR, using age, sex, causative drug, number of concomittant drugs, region, and binary variables for each SCAR as covariates. (A) Calculated survival curves for each SCAR. DRESS has the longest median (19.84 days, IQR [8.96, 30.95]), then SJS-TEN (11.82d, [5.77, 22.91], AGEP (4.19d, [1.87, 8.96]), and GBFDE (3.28d, [1.94, 6.37]). P-value by log-rank test. (B) Bootstrapped VIMPs from RSF. Each was calculated using 100 bootstrapped samples.

To display the influence of covariates, we calculated survival curves stratified by each covariate (Figure S7). For all SCAR combined, the log-rank test conducted on each covariate (age, sex, region, and number of concomitant drugs) suggests significant differences. For SJS-TEN and DRESS, however, only region and age are significant (p <0.001), while for AGEP only region is significant (p = 0.0229). For all SCAR, SJS-TEN, and DRESS, the 0–20-year and 80–100-year age group had shortest latency. There is significant regional variation in latency, with the regions showing shortest and longest latency differing by SCAR phenotype.

Mortality

For 13540 cases with sufficient data, mortality was computed by age, concomitant drug count, TTE, sex, region, and SCAR phenotype (Figure 8). Mortality rises with age and number of concomitant drugs, reaching 34.58% for cases with at least 12 drugs. Similar trends characterize individual SCAR (Figure S8-10). SJS-TEN has highest mortality (22.8%), followed by GBFDE (20%), DRESS (5.66%) and AGEP (3.13%). Regionally, North America has lowest mortality (8.61%) and Sub-Saharan Africa has the highest (24.0%). For both overall and individual SCAR, male mortality exceeds female.

Figure 8.
Figure 8. Unadjusted Mortality Rates.

Unadjusted mortality rates were calculated by (A) age group, (B) sex, (C) the number of concomitant drugs recorded, (D) geographic region, (E) TTE/latency, (F) SCAR phenotype. For continuous variables, bins were selected for approximately equal frequency. NS= Not specified.

RFs were used to determine the influence of each variable on the mortality rate, and partial dependence plots (PDPs) were created to assess the influence of a single covariate while averaging the others (Figure 9). Predicted mortality still increased with age and number of concomitant drugs, though with a marked increase for patients <5 years old. Predicted female mortality rate remained slightly lower than the male rate. When considering mortality per SCAR, SJS-TEN is markedly higher than the others, followed by GBFDE, DRESS, and AGEP. VIMPs showed age and concomitant count were the most influence demographic variables, though the binary SJS-TEN variable had highest VIMP overall. Subsequently, we constructed separate RFs on SJS-TEN, DRESS, and AGEP cases (Figure S11-13, respectively). Like SCAR in aggregate, individual SCAR showed increased mortality with age and concomitant drug count. VIMPs from each of these RF similarly show age and concomitant count have the most influence on predicted morality, with TTE and causative drug also having significant importance.

Figure 9.
Figure 9. Partial Dependence Plots (PDP), Variable Importance (VIMP), and Model Performance for Random Forest Classifer of SCAR Mortality.

Marginal dependence of mortality on (A) age, (B) sex, (C) number of concomitant meds administered, (D) geographic region, (E) HE/latency, and (F) SCAR type is obtained from the RF trained on all SCAR cases. (G) VIMP was calculated using 100 boostrapped samples. AnyVIMP over zero contributed to the predictive ability of the RF. GBFDE only had five cases total, but was included for completion. Model performance metrics are also included. OOB – out-of-bag. AUC – area under the curve.

Discussion

This work is one of the most comprehensive retrospective analyses of SCAR to date. FAERS yielded 56,683 cases, 10-100x more than previous studies with notable duplicates removed. Reported causative agents include well-known culprits like beta-lactams, anticonvulsants, and allopurinol, and less-studied agents like immunomodulators. The FDA’s Drug Safety-related Labeling Changes (SrLC) lists nine drugs since 2016 have had label changes due to a strong link with SJS-TEN and/or DRESS – lamotrigine, allopurinol, phenytoin, carbamazepine, amoxicillin, trimethoprim-sulfamethoxazole (TMP-SMX), rifampin, levetiracetam, and vancomycin.3 Importantly, four of these drugs that have had the most label changes since 2016 for SCAR, DRESS, and/or SJS-TEN (allopurinol – 4 changes, phenytoin – 6 changes, rifampin – 5 changes, and TMP-SMX – 4 changes), are the same drugs that we found to be disproportionately represented in our data, providing evidence of the robustness of our methods but also indicating that these label changes may have impacted reporting.

The impact of these findings is underscored by the methods used to prepare FAERS, and by the breadth of medications studied. Prior studies have been limited by smaller sample sizes or focus on a few drugs.17,18 Additionally, prior studies of SCAR have not been as thorough in sanitization, including limited deduplication efforts and lack of stringency in drug naming. The issue of deduplication is well-documented in the literature, and is an area of active improvement.6,19 Even with our efforts, clear limitations remain. FAERS relies on voluntary reporting and has unclear reporting requirements for any given condition (e.g. DRESS or AGEP meeting RegiSCAR or EuroSCAR criteria), leading to potential reporting biases. Detail varies by case, with some omitting concomitant drugs, exact dates, and phenotyping. Furthermore, while our sanitization is stringent, it is likely that duplicated entries and typographical errors remained. Future studies should continue to improve on these methods using more advanced techniques.

The Changing Landscape of SCAR

FAERS reporting of SCAR has been increasing over time. For DRESS, this is likely due to the condition first being named in 1996, common criteria being published in the mid-2000s, and FAERS first reporting in 2003.20 GBFDE, known since the 1980s, has only been reported in FAERS since 2021.21 While the plurality of reports are from high-resource nations, there are many reports from middle– and lower-income nations, reflecting the SCAR’s global nature.

Particularly striking is the fact that immunomodulatory agents are on the rise as a cause of SCAR over the past 20 years, mostly due to ICIs. ICI-induced SCAR presents a mechanistic challenge. While small molecules cause SCAR by perturbing the TCR-peptide-MHC complex, ICIs likely exert their influence through inhibiting tolerance, leading to a variety of organ toxicities including SJS-TEN-like reactions.22 ICIs have a longer latency than other medications in our data, corroborating that they may have a different pathomechanism. Furthermore, ICI-induced SCAR may be truly ICI-induced, or ICIs may facilitate small-molecule-induced SCAR via their mechanism. Further studies of this growing population, with precise phenotyping and clinical information, are needed to understand the mechanisms of ICI-induced SCAR.

Overlap Conditions

Full phenotyping and histopathology are often required to diagnose conditions meeting the criteria of >1 SCAR. While such data is unavailable in FAERS, our work offers some insights. We demonstrate that co-reporting is higher than expected by random assortment and occurs more frequently than with other cutaneous conditions commonly misdiagnosed for each other (e.g. autoimmune bullous disease).

SJS-TEN/DRESS overlap, meaning patients meeting criteria for both diseases, has been reported with allopurinol, anticonvulsants and TMP-SMX, among others.23,24 AGEP/DRESS overlap is more difficult – the condition is recognized as a discrete entity within FAERS; in our data it appears in one case (primary suspect levofloxacin). Pustular DRESS is a known phenotype, with vancomycin, phenytoin, doxycycline and piperacillin-tazobactam implicated in a recent review (noting that these patients didn’t meet full AGEP criteria).25 A retrospective study of 216 patients found one probable case of DRESS-AGEP overlap meeting both criteria.26 Similarly, AGEP with TEN-like features, or AGEP/SJS-TEN overlap, is increasingly recognized.27 Many present with widespread desquamation and mucosal involvement that clinically appears SJS-TEN-like, but with pathology showing pustules. Because histopathology is needed, many cases of both true AGEP/SJS-TEN overlap and AGEP with SJS-TEN-like features are likely misdiagnosed. Further harmonization of criteria and research coordination are needed to characterize these rare cases.

Causative Agents and Latency

Drug latency is a critical issue in the management of SCAR; in tandem with drug timelines, prior knowledge of latency is vital to determining the causative agent in complex cases. We have demonstrated that the causative agent is the strongest influence on latency, and that some agents have unusually long or short latencies, which is valuable knowledge in a clinical setting. The reason for the dependence of latency on causative agent is not well-understood and bears further study, perhaps by considering chemical structure, metabolisms, and time needed for antigen presentation.

A controversy in the field of SCAR is protopathic bias – namely, when a drug is misattributed as causative because it is prescribed early during SCAR to another drug. This is particularly notable for NSAIDs, which are often prescribed in the flu-like prodromal phase of SJS-TEN, and which have low latency (4-5 days) in our data. An internal FDA review suggested acetaminophen caused rash, and mucosal erosion with a 1-3 day delay.28 Some antibiotics like amoxicillin and azithromycin may be protopathic due to being prescribed for subjective malaise. A concerted effort to confirm these cases with follow-up testing would be helpful.

Insights from Machine Learning

For the first time, we have applied random forests to SCAR pharmacovigilance. These methods are particularly suited for complex data with interdependent variables – particularly relevant to SCAR, in which age, causative agent, sex, and number of concomitant drugs likely interact. While previous studies have applied ML to pharmacovigilance, none have used it to predict TTE or mortality, and none have applied them to SCAR.29,30 Using RSFs, we have shown that age, sex, region, and number of concomitant drugs have an influence on latency.

Interestingly, younger patients had shorter latency, even when controlling for other variables (including the causative drug), likely due to faster antigen processing or metabolism. There is also regional variation in latency, which may arise from differences in diagnosis and surveillance. A larger retrospective dataset, with clear criteria for SCAR, well-defined time-of-onset data and causative agent determination, is likely required to gain insights into the role of demographic variables in latency. Such information would be clinically valuable, particularly for difficult cases in which the causative drug is hard to determine from drug timelines.

Rare events like SCAR are ideal subjects of study using large pharmacovigilance databases like FAERS, which aggregate data worldwide to provide large datasets amenable to deeper analyses. However, these databases suffer from issues with duplication, missing data, and reporting bias. As a rare disease, SCAR is particularly complicated to study, with significant phenotypic variation, issues with misdiagnosis, inconsistent criteria, strong genetic influences, and difficulty determining the causative drug. We hope that our study, and the application of more advanced methods of deduplication and statistical analysis, will spur new insights into understanding pathogenesis and improving prediction of SCAR.

Contributions

EMM conceived the project, wrote the manuscript, and conducted and validated all analyses. DP conducted analysis and edited the manuscript. MMP assisted in drafting the manuscript. EP reviewed all analyses and assisted in editing the manuscript.

Data Sharing

Methods and data are available at https://github.com/capuhcheeno/SCARs_ICI-Manuscript-Scripts.

Statement on Use of Artificial Intelligence

GPT4o was used to write, optimize, and debug R, SQL and Python code. Suggestions were cross-referenced in literature, package documentation, and resources like stackoverflow.com. All code and generated data were manually reviewed, tested, and verified by the authors. The authors take responsibility for all code generated.

Acknowledgements

E.J.P. is supported by the following grants from the National Institutes of Health (NIH): NIH U01AI154659, NIH P50GM115305, NIH R01HG010863, NIH R21AI139021, NIH R01AI152183, and NIH 2 D43 TW010559. E.J.P. is also supported by the National Health and Medical Research Council of Australia. E.M.M. is funded by a Vanderbilt University Medical Center internal career development award (Vanderbilt Faculty Research Scholars).

References

  1. 1.↵
    Lazarou J, Pomeranz BH, Corey PN. Incidence of adverse drug reactions in hospitalized patients: a meta-analysis of prospective studies. JAMA. 1998;279(15):1200–1205. doi:10.1001/jama.279.15.1200
    OpenUrlCrossRefPubMedWeb of Science
  2. 2.↵
    Hung SI, Mockenhaupt M, Blumenthal KG, et al. Severe cutaneous adverse reactions. Nat Rev Dis Primers. 2024;10(1):1–25. doi:10.1038/s41572-024-00514-0
    OpenUrlCrossRefPubMed
  3. 3.↵
    Drug Safety-related Labeling Changes (SrLC). Accessed February 26, 2025. https://www.accessdata.fda.gov/scripts/cder/safetylabelingchanges/
  4. 4.↵
    Creadore A, Desai S, Alloo A, et al. Clinical Characteristics, Disease Course, and Outcomes of Patients With Acute Generalized Exanthematous Pustulosis in the US. JAMA Dermatology. 2022;158(2):176–183. doi:10.1001/jamadermatol.2021.5390
    OpenUrlCrossRefPubMed
  5. 5.↵
    Micheletti RG, Chiesa-Fuxench Z, Noe MH, et al. Stevens-Johnson Syndrome/Toxic Epidermal Necrolysis: A Multicenter Retrospective Study of 377 Adult Patients from the United States. J Invest Dermatol. 2018;138(11):2315–2321. doi:10.1016/j.jid.2018.04.027
    OpenUrlCrossRefPubMed
  6. 6.↵
    Giunchi V, Fusaroli M, Hauben M, Raschi E, Poluzzi E. Challenges and Opportunities in Accessing and Analysing FAERS Data: A Call Towards a Collaborative Approach. Drug Saf. 2023;46(10):921–926. doi:10.1007/s40264-023-01345-w
    OpenUrlCrossRefPubMed
  7. 7.↵
    Banda JM, Evans L, Vanguri RS, Tatonetti NP, Ryan PB, Shah NH. A curated and standardized adverse drug event resource to accelerate drug safety research. Sci Data. 2016;3:160026. doi:10.1038/sdata.2016.26
    OpenUrlCrossRef
  8. 8.↵
    Ishwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests. Ann Appl Stat. 2008;2(3). doi:10.1214/08-AOAS169
    OpenUrlCrossRefWeb of Science
  9. 9.↵
    Evans SJ, Waller PC, Davis S. Use of proportional reporting ratios (PRRs) for signal generation from spontaneous adverse drug reaction reports. Pharmacoepidemiol Drug Saf. 2001;10(6):483–486. doi:10.1002/pds.677
    OpenUrlCrossRefPubMedWeb of Science
  10. 10.↵
    Nguyen MN, Reyes M, Jones SC. Postmarketing Cases of Enfortumab Vedotin– Associated Skin Reactions Reported as Stevens-Johnson Syndrome or Toxic Epidermal Necrolysis. JAMA Dermatology. 2021;157(10):1237–1239. doi:10.1001/jamadermatol.2021.3450
    OpenUrlCrossRefPubMed
  11. 11.
    Vivar KL, Mancl K, Seminario-Vidal L. Stevens–Johnson syndrome/toxic epidermal necrolysis associated with zonisamide. Clin Case Rep. 2017;6(2):258–261. doi:10.1002/ccr3.1288
    OpenUrlCrossRefPubMed
  12. 12.↵
    Alkhateeb H, Said S, Cooper CJ, Gaur S, Porres-Aguilar M. DRESS syndrome following ciprofloxacin exposure: An unusual association. Am J Case Rep. 2013;14:526–528. doi:10.12659/AJCR.889703
    OpenUrlCrossRefPubMed
  13. 13.↵
    Allegra A, Alonci A, Penna G, et al. Stevens-Johnson syndrome after lenalidomide therapy for multiple myeloma: a case report and a review of treatment options. Hematol Oncol. 2012;30(1):41–45. doi:10.1002/hon.1000
    OpenUrlCrossRefPubMed
  14. 14.
    Salama M, Lawrance IC. Stevens-Johnson syndrome complicating adalimumab therapy in Crohn’s disease. World J Gastroenterol. 2009;15(35):4449–4452. doi:10.3748/wjg.15.4449
    OpenUrlCrossRefPubMedWeb of Science
  15. 15.↵
    Lowndes S, Darby A, Mead G, Lister A. Stevens-Johnson syndrome after treatment with rituximab. Ann Oncol. 2002;13(12):1948–1950. doi:10.1093/annonc/mdf350
    OpenUrlCrossRefPubMedWeb of Science
  16. 16.↵
    Horcajada-Reales C, Pulido-Pérez A, Suárez-Fernández R. Severe Cutaneous Drug Reactions: Do Overlapping Forms Exist? Actas Dermo-Sifiliográficas (English Edition). 2016;107(1):23–33. doi:10.1016/j.adengl.2015.11.003
    OpenUrlCrossRef
  17. 17.↵
    Zhou L, Yang J, Xiao M, et al. Severe cutaneous adverse reactions due to antibiotics therapy: a pharmacovigilance analysis of FDA adverse event reporting system events. Expert Opinion on Drug Safety. 0(0):1–8. doi:10.1080/14740338.2023.2278685
    OpenUrlCrossRef
  18. 18.↵
    Mao K, Li J, Zhu X, Sun H, Zhong S, Mao W. Signal mining study of severe cutaneous adverse events of valaciclovir or acyclovir based on the FAERS database. Expert Rev Clin Pharmacol. 2024;17(1):101–110. doi:10.1080/17512433.2023.2294002
    OpenUrlCrossRefPubMed
  19. 19.↵
    Kiguba R, Isabirye G, Mayengo J, et al. Navigating duplication in pharmacovigilance databases: a scoping review. BMJ Open. 2024;14(4):e081990. doi:10.1136/bmjopen-2023-081990
    OpenUrlAbstract/FREE Full Text
  20. 20.↵
    Shiohara T, Iijima M, Ikezawa Z, Hashimoto K. The diagnosis of a DRESS syndrome has been sufficiently established on the basis of typical clinical features and viral reactivations. Br J Dermatol. 2007;156(5):1083–1084. doi:10.1111/j.1365-2133.2007.07807.x
    OpenUrlCrossRefPubMedWeb of Science
  21. 21.↵
    Kauppinen K, Stubb S. Fixed eruptions: causative drugs and challenge tests. Br J Dermatol. 1985;112(5):575–578. doi:10.1111/j.1365-2133.1985.tb15266.x
    OpenUrlCrossRefPubMed
  22. 22.↵
    Molina GE, Yu Z, Foreman RK, Reynolds KL, Chen ST. Generalized bullous mucocutaneous eruption mimicking Stevens-Johnson syndrome in the setting of immune checkpoint inhibition: A multicenter case series. J Am Acad Dermatol. 2020;83(5):1475–1477. doi:10.1016/j.jaad.2020.03.029
    OpenUrlCrossRefPubMed
  23. 23.↵
    Van Nguyen K, Van Vu Q, Tran MH, et al. Overlapping Stevens-Johnson Syndrome and DRESS Syndrome Caused by Phenobarbital: A Vietnamese Case Report. Glob Pediatr Health. 2023;10:2333794X231216556. doi:10.1177/2333794X231216556
    OpenUrlCrossRefPubMed
  24. 24.↵
    Chin LD, MacGillivary ML, Purdy KS, Kirshen C. Stevens-Johnson syndrome with overlapping features of DRESS syndrome: A report of two cases. SAGE Open Medical Case Reports. 2024;12:2050313X241307097. doi:10.1177/2050313X241307097
    OpenUrlCrossRefPubMed
  25. 25.↵
    Hansen E, Kirven RM, Gallardo M, et al. Drug Reaction With Eosinophilia and Systemic Symptoms With Pustulosis. JAMA Dermatology. 2024;160(7):777–779. doi:10.1001/jamadermatol.2024.1082
    OpenUrlCrossRefPubMed
  26. 26.↵
    Bouvresse S, Valeyrie-Allanore L, Ortonne N, et al. Toxic epidermal necrolysis, DRESS, AGEP: Do overlap cases exist? Orphanet Journal of Rare Diseases. 2012;7(1):72. doi:10.1186/1750-1172-7-72
    OpenUrlCrossRefPubMed
  27. 27.↵
    Chowdhury TA, Talib KA, Patricia J, Nye KD, Moosa SA. Rare and Complicated Overlap of Stevens-Johnson Syndrome and Acute Generalized Exanthematous Pustulosis. Cureus. 2021;13(6):e15921. doi:10.7759/cureus.15921
    OpenUrlCrossRef
  28. 28.↵
    Amending Over-the-Counter Monograph M013: Internal Analgesic, Antipyretic, and Antirheumatic Drug Products for Over-the-Counter Human Use | FDA. Accessed February 8, 2025. https://dps.fda.gov/omuf/ordersearch/order_otc000035
  29. 29.↵
    Zhao Y, Yu Y, Wang H, et al. Machine Learning in Causal Inference: Application in Pharmacovigilance. Drug Saf. 2022;45(5):459–476. doi:10.1007/s40264-022-01155-6
    OpenUrlCrossRefPubMed
  30. 30.↵
    Farnoush A, Sedighi-Maman Z, Rasoolian B, Heath JJ, Fallah B. Prediction of adverse drug reactions using demographic and non-clinical drug characteristics in FAERS data. Sci Rep. 2024;14(1):23636. doi:10.1038/s41598-024-74505-2
    OpenUrlCrossRefPubMed
Back to top
Posted March 06, 2025.
Download PDF

Supplementary Material

Data/Code
Email
Characterization of Demographics, Drug Latency, and Mortality of Severe Cutaneous Adverse Reactions in an FDA Pharmacovigilance Database
Eric Milan Mukherjee, Dodie Park, Michelle Martin-Pozo, Elizabeth Phillips
medRxiv 2025.03.05.25323441; doi: https://doi.org/10.1101/2025.03.05.25323441
Twitter logo Facebook logo LinkedIn logo Mendeley logo
Citation Tools

Subject Area