Comparison of Artificial Intelligence and Lexicomp for Drug--Drug Interaction Screening in Hospitalized Patients

Comparison of Artificial Intelligence and Lexicomp for Drug--Drug Interaction Screening in Hospitalized Patients | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 23 September 2025 V1 Latest version Share on Comparison of Artificial Intelligence and Lexicomp for Drug--Drug Interaction Screening in Hospitalized Patients Authors : Noemi Santos de Oliveira , Tricila Ribeiro de Matos , Eduardo de Araújo Oliveira , Luiza Fernanda Mendonça Nicolau , Emilyn Barbosa de Oliveira Guerreiro , Fernando Chiodini Machado , and Zaira Fernanda Martinho Nicolau [email protected] Authors Info & Affiliations https://doi.org/10.22541/au.175861495.50863378/v1 424 views 141 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Background and Purpose Adverse drug reactions are a major source of healthcare burden, with drug–drug interactions (DDIs) contributing up to 20% of cases. Conventional electronic checkers often show heterogeneity in interaction detection and severity assessment. This study aimed to compare an LLM-based artificial intelligence system with UpToDate/Lexicomp for DDI detection, severity classification, and processing time in hospitalized patients. Experimental Approach We conducted a comparative, cross-sectional study at Hospital Samel, Brazil, including 60 inpatients (ICU and ward) prescribed ≥2 medications. Active drug lists were assessed with both UTD and an AI system, which integrated demographic, clinical, and laboratory data to classify DDIs on the same severity scale. Key Results Among 284 interaction pairs, AI identified 191 and UTD 184, with 91 detected by both. Severity concordance was 63.7% (κ=0.19). AI tended to assign higher severity, often upgrading category C to D. Compared with UTD, AI detected more interactions with NSAIDs (22 vs 1; q<0.001) and antibiotics (21 vs 3; q=0.003), while UTD identified more non-opioid analgesic and antipsychotic interactions. Post-hoc prompting showed AI acknowledged 71% of initially missed UTD interactions, largely filtered as “low priority.” Processing time was significantly shorter with AI (36.2s vs 102.4s per patient; p<0.001), a 68% reduction. Conclusion and Implications LLM-based AI complemented conventional checkers, detecting additional interactions and delivering faster and contextualized assessments. Despite single-center and small-sample limitations, this study advances understanding by demonstrating AI’s potential to enrich drug–drug interaction evaluation and reduce processing burden in clinical workflows. 1.INTRODUCTION Adverse drug reactions (ADRs) are a leading and persistent cause of patient harm and health-system expenditure. In a large prospective UK cohort, 6.5% of all acute hospital admissions were related to ADRs, with a projected annual cost to the National Health Service of US$847 million (Pirmohamed et al., 2004). Also, 10% of inpatients experience an ADR during their hospital stay (Bouvy et al., 2015). Drug–drug interactions (DDIs) can account up to 20% of ADR (Pirmohamed et al., 2004; Magro et al., 2011). The occurrence of ADRs secondary to DDI can be influenced by many factors, such as age, genetics, renal function, and other comorbidities (Malki et al., 2020). To reduce DDI-related harm, clinicians rely on electronic drug interaction checkers. Among the most widely used are UpToDate (UTD), powered by the Lexicomp database (Wolters Kluwer, 2025), and similar evidence-based platforms. These systems standardize the classification of DDIs and provide recommendations. However, there are inconsistent results in detection and severity classification across databases (Carollo et al., 2024). Recent advances in artificial intelligence (AI), particularly large language models (LLMs), have created new opportunities for clinical decision support. LLMs can process unstructured data, for example, free-text clinical notes, laboratory results, reason across multimodal inputs, and adapt recommendations to patient-specific contexts (Rajkomar et al., 2018; Qi et al., 2025). Nonetheless, direct comparative studies between LLM-based AI systems and established drug interaction checkers remain scarce. We hypothesized that an LLM-based AI could complement conventional drug interaction checkers by providing faster, contextualized, and potentially broader detection of clinically relevant DDIs. Therefore, the objective of this study was to compare the performance of an LLM-based AI system with the UTD/Lexicomp module in DDI detection, severity classification, and processing time among hospitalized patients. 2. MATERIALS and METHODS 2.1. Study Design and Setting This was a comparative, cross-sectional study performed at Hospital Samel, a tertiary care center in Manaus, Brazil. The study compared the performance of the drug interaction module integrated into UTD, which is powered by the Lexicomp database, with SAMIA, a large language model (LLM)–based artificial intelligence system of Hospital Samel. 2.2. Population and Inclusion Criteria Patients were eligible if they were hospitalized in the intensive care unit (ICU) or general ward during the study period and were prescribed two or more medications. Demographic and clinical information, including age, sex, hospital unit, medical history, and laboratory results, were retrieved from the electronic medical record. A total of 60 patients met inclusion criteria and were included in the final analysis. 2.3. Drug Interaction Assessment For each patient, the active medication list was entered into the UTD drug-interaction checker, and results were recorded using Lexicomp’s four-level severity scale (B: no action needed; C: monitor therapy; D: consider therapy modification; X: avoid combination). In parallel, SAMIA generated drug interactions alerts contextualizing its response by patient-specific factors, assigning severity on the same scale for direct comparison. SAMIA’s assessments integrated pharmacokinetic/pharmacodynamic mechanisms, documentation quality, and onset, were contextualized by patient factors (age, renal/hepatic function, clinical history, medication history, laboratory/vital data, and pregnancy/lactation), and considered benefit–risk and therapeutic alternatives. It was instructed to use established sources, regulatory product labels (Anvisa, FDA, EMA), and curated drug knowledge bases, with supplemental checks. For drug pairs flagged by UTD but not initially by AI, we issued a standardized post-hoc prompt listing the pairs, asking whether an interaction existed and, if so, to classify its severity. If AI affirmed an interaction, a second prompt requested one primary reason for the initial miss (knowledge gap, inference, pair ambiguity, low priority, or other). Post-hoc responses were used to characterize recoverability and failure modes and did not alter primary detection outcomes. 2.4. Outcome Measures Outcome measures included concordance between SAMIA and UTD in detecting and classifying drug interactions. Detection concordance was the proportion of interactions identified by both systems. Severity concordance was the proportion of jointly detected interactions assigned to the same severity category. Additional outcomes were the number of interactions detected exclusively by either system, inter-system agreement quantified with Cohen’s kappa, the presence of systematic asymmetry in discordant classifications, and the processing time per patient to complete interaction checks. 2.5. Ethical Considerations The study was conducted in accordance with the Declaration of Helsinki. It received approval from the ethics committee (CAAE: 85059824.2.0000.0005). 2.6. Statistical Analysis Patient demographics and medication counts were summarized as mean (SD) or median (IQR) according to distribution (Shapiro–Wilk). Between-group comparisons (ICU vs ward) used Mann–Whitney U for non-normal data with confirmatory Welch’s t where appropriate. Agreement between AI and UTD was quantified with Cohen’s κ, and concordance rates as simple proportions. Paired proportions of detected interactions were compared with McNemar’s test; directional asymmetry in multi-category classifications was evaluated with Bowker’s test of symmetry. Processing times were analyzed with paired t tests, with Wilcoxon signed-rank and sign tests as nonparametric corroboration; effect sizes were reported as Cohen’s d (between-group) and d_z or r (paired). For class-level analyses, drugs were grouped into therapeutic classes, and two-sided exact McNemar tests were applied with Benjamini–Hochberg FDR (BH-FDR) control across classes. Significance: p<0.05 (two-sided) for single comparisons; q<0.05 (BH–FDR) for class-level multiplicity. Analyses were performed in Python (v3.11) with SciPy (v1.11.3) and Statsmodels (v0.14.0). 2.7. Results A total of 60 patients were included in the analysis. The mean age was 28.2 ± 15.2 years (range: 1–95 years). Regarding sex distribution, 27 patients (45.0%) were male and 33 patients (55.0%) were female. Patients admitted to the Intensive Care Unit (ICU) had a higher mean number of medications (8.1) compared with those in the ward (6.9). The median number of medications was 8.0 (interquartile range [IQR]: 7–9) in the ICU group and 7.0 (IQR: 6–8) in the ward group. The ICU group demonstrated a wider range of medication counts (4–11) than the ward group (2–10). The distribution of the number of medications per patient, stratified by hospital location, is presented in Figure 1 . Normality testing indicated non-normal distributions in both groups (Shapiro–Wilk p<0.001). A Mann–Whitney U test confirmed that patients in the ICU received significantly more medications than those in the ward (U=8986.5; p<0.001). Results were consistent with Welch’s t test (t=4.44; p<0.001), with a Cohen’s d of 0.72, indicating a medium-to-large effect size. A total of 284 drug interaction pairs were analyzed. Patients contributed a median of 4 interactions (IQR 3–6; range 1–15). Concordance between AI and UTD was observed in 91 cases (32.0%). AI identified 191 interactions in total, whereas UTD identified 184. Among these, 100 interactions were detected exclusively by AI and 93 exclusively by UTD. There was no statistically significant difference between the two methods (p = 0.67). A total of 91 drug interaction pairs were detected simultaneously by both AI and UTD. Concordance in severity classification was observed in 58 cases (63.7%). When AI assigned category C, UTD also classified 50 cases as C, while seven were classified as B and 24 as D. When AI assigned category D, UTD agreed in eight cases, while one was classified as C and one as X. No concordant assignments were observed for category B or X. The overall level of agreement between AI and UTD for severity classification was low, with an unweighted Cohen’s kappa coefficient of 0.19, indicating slight agreement beyond chance. In addition, Bowker’s test of symmetry demonstrated a highly significant asymmetry in the discordant classifications (χ² = 29.2, df = 3, p < 0.0001), indicating that AI systematically tended to assign higher severity categories (particularly D instead of C) compared with UTD. Asymmetric detection varied by class. AI detected more interactions than UTD for nonsteroidal anti-inflammatory drugs (NSAIDs) (22 vs 1; p=5.72×10⁻⁶; q=1.20×10⁻⁴; N=27) and antibiotics (21 vs 3; p=2.77×10⁻⁴; q=0.00291; N=26). UTD detected more than AI for non-opioid analgesics (17 vs 37; p=0.0091; q=0.0635; N=64) and antipsychotics (2 vs 11; p=0.0225; q=0.1179; N=28). For GI mucosal protectants, AI detected more than UTD (6 vs 0; p=0.0313; q=0.1313; N=6). After BH–FDR correction, only NSAIDs and antibiotics remained significant (q<0.05). Among the 93 interactions that were initially not detected by the AI but were identified by UTD, the AI subsequently agreed in 66 cases (71.0%) that an interaction was present when asked directly. Of the 56 interactions classified as category C by UTD, the AI post-hoc classification agreed in 51 cases, while one was downgraded to category B and four were upgraded to category D. Among the eight interactions initially classified as category D by UTD, the AI agreed in two cases but downgraded six to category C. For the two interactions classified as category X by UTD, the AI reclassified both as category C. These cross-classifications are depicted in Figure 2 . Agreement beyond chance was low, with Cohen’s kappa = 0.15 (unweighted), and 0.10 (quadratic weighted). AI assigned a higher severity than UTD in four cases (6.1%), and a lower severity in nine cases (13.6%). There was no significant imbalance between “AI higher” and “AI lower” (p = 0.27). Regarding the reasons for initial AI non-detection, the most frequently reported cause was low priority (60/66; 90.9%), followed by inference (5/66; 7.6%) and insufficient knowledge (1/66; 1.5%). There was no significant association between the reason for failure and subsequent agreement in severity category (χ² = 1.61, p = 0.45). However, analysis of reason versus classification direction revealed a statistically significant association (χ² = 12.5, p = 0.02), mainly driven by cases labeled as “inference,” in which AI more frequently assigned a higher severity than UTD after prompting. UTD time to complete drug–interaction checks required a mean of 102.4 seconds (SD 26.3; median 97.0; IQR 81.0–122.0), whereas AI required 36.2 seconds (SD 18.6; median 32.0; IQR 25.0–46.0). The mean paired difference (UTD – AI) was 66.1 seconds (SD 23.0; 95% CI 63.5 to 68.8), as illustrated in Figure 3 , which depicts paired times for each interaction. A paired t test indicated that AI was significantly faster than UTD (t(283)=48.4; p<0.001), with a large standardized effect (Cohen’s d_z=2.87). The Wilcoxon signed-rank test yielded consistent results (p<0.001; r=0.87), and AI was faster in all comparisons (284 of 284; p<0.001). A log-scale analysis yielded an AI/UTD geometric mean ratio of 0.318 (95% CI 0.300–0.337; p<0.001), corresponding to AI requiring approximately 68% less time than UTD (95% CI 66%–70%). 3.DISCUSSION AND CONCLUSION In this comparative study of an LLM-based system with UTD/Lexicomp for DDIs, three findings stand out. First, each system surfaced a substantial set of interactions not identified by the other (AI 100; UTD 93), indicating complementary coverage rather than redundancy. Second, among jointly detected interactions (n=91), severity agreement was only slight (κ=0.19). This limited agreement between sources has also been reported in other comparative studies of DDI checkers (Carollo et al., 2024a; Carollo et al.,2024b). Comparing Lexicomp and Google Bard, researchers found only slight agreement in severity ratings and no agreement in reliability scores (Sulaiman et al., 2023), while Lexicomp and Micromedex showed moderate consistency (Liu et al., 2023). Third, AI completed checks markedly faster, requiring 68% less time per check on average, and it was faster in every paired comparison. Taken together, these results suggest that an LLM can act as a complement to established DDI checkers, adding patient context while substantially reducing time burden. A key contribution of AI in this setting is contextualization. AI took into account demographics, renal/hepatic function, clinical history, and laboratory/vital data to modulate risk, whereas electronic checkers primarily return decontextualized, population-level statements. In our study, AI detected more interactions for NSAIDs and antibiotics. This can be helpful in reducing the risk of clinically significant DDI-related adverse events, such as bleeding, arrhythmias, and renal failure (Day et al., 2017). Conversely, UTD surfaced more non-opioid analgesic and antipsychotic interactions, which may reflect differences in curation depth in those classes. These complementary blind spots can be leveraged in a combined workflow that maximizes each system’s strengths. Disagreement deserves careful interpretation. When the AI was asked post hoc about the 93 UTD-only pairs, it acknowledged 71% as genuine interactions, and 90.9% were not showed because of low priority. In practical terms, LLMs appear to operate with a prioritization layer that filters low-yield signals unless explicitly queried. This property, if appropriately governed, could mitigate alert fatigue. Alert fatigue remains a global concern, with 81% of clinicians perceiving DDI alerts as excessive and 68% reporting reduced focus due to cognitive overload, which increases the risk of overlooking important safety information (Takizawa et al., 2025). The observed time savings are clinically meaningful. Manual compendium checks took about 102 seconds per check versus 36 seconds with AI. In hospital care with many patients, this reduction can lessen the workload for pharmacists and physicians without sacrificing overall detection. Beyond speed, AI produced structured rationales and listed therapeutic alternatives, outputs that can shorten the loop from alert to action and facilitate shared decision-making at the bedside (Garcia-Vidal et al., 2019). Translating these findings into practice, we envision a layered decision-support model. The electronic drug checker analyzes potential DDIs . Then, the AI performs a screening that integrates the patient’s clinical history and provides patient-specific justifications, with suggested next steps. A governance layer aligns information with institutional protocols and scientific evidence. Such a combination uses AI to personalize and accelerate review, while the electronic drug checker stabilizes calibration and ensures reproducibility, all under final governance supervision. Our study limitations include a small sample size and a single-center setting with a relatively young median age. This may limit generalizability to other regions and age ranges. Also, AI is in continuous development, and this study represents the first implementation of this interaction-detection mechanism. Subsequent updates could alter outputs. DDI review should shift from static lookups to patient-specific reasoning. Large language models can serve as assistive layers that convert clinical data and pharmacology into clear, point-of-care guidance. By organizing information and highlighting the most relevant risks and management options, these tools help clinicians prioritize clinically meaningful interactions and act more efficiently. Future studies with real-world evaluation should confirm gains in safety, usability, and value. 4.REFERENCES 1. Pirmohamed M, James S, Meakin S, Green C, Scott AK, Walley TJ, Farrar K, Park BK, Breckenridge AM. Adverse drug reactions as cause of admission to hospital: prospective analysis of 18 820 patients. BMJ 2004;329:15-9. 2. Bouvy J, De Bruin M, Koopmanschap M. Epidemiology of adverse drug reactions in europe: a review of recent observational studies. Drug Saf 2015;38:437–53. 3. Magro L, Moretti U, Leone R. Epidemiology and characteristics of adverse drug reactions caused by drug–drug interactions. Expert Opin Drug Saf 2011;11:83–94. 4. Malki MA, Pearson ER. Drug-drug-gene interactions and adverse drug reactions. Pharmacogenomics J 2020;20:355-66. 5. Wolters Kluwer. Uptodate Drug Interactions. Accessed April, 2025. https://www.uptodate.com/drug-interactions/?source=responsive_home#di-druglist. 6. Carollo M, Crisafulli S, Ciccimarra F, Andò G, Diemberger I, Trifirò G. Exploring the level of agreement among different drug-drug interaction checkers: a comparative study on direct oral anticoagulants. Expert Opin Drug Metab Toxicol 2024;20:157-64. 7. Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, Liu PJ, Liu X, Marcus J, Sun M, Sundberg P, Yee H, Zhang K, Zhang Y, Flores G, Duggan GE, Irvine J, Le Q, Litsch K, Mossin A, Tansuwan J, Wang D, Wexler J, Wilson J, Ludwig D, Volchenboum SL, Chou K, Pearson M, Madabushi S, Shah NH, Butte AJ, Howell MD, Cui C, Corrado GS, Dean J. Scalable and accurate deep learning with electronic health records. NPJ Digit Med 2018;1:18. 8. Qi H, Li X, Zhang C, Zhao T. Improving drug-drug interaction prediction via in-context learning and judging with large language models. Front Pharmacol 2025;16:1589788. 9. Carollo M, Crisafulli S, Selleri M, Piccoli L, L’Abbate L, Trifirò G. Agreement of Different Drug-Drug Interaction Checkers for Proton Pump Inhibitors. JAMA Netw Open 2024;7(7):e2419851. 10. Sulaiman DM, Shaba SS, Almufty HB, Sulaiman AM, Merza MA. Screening the Drug-Drug Interactions Between Antimicrobials and Other Prescribed Medications Using Google Bard and Lexicomp® Online™ Database. Cureus 2023;15(9):e44961. 11. Liu Y, Wang J, Gong H, Li C, Wu J, Xia T, Li C, Li S, Chen M. Prevalence and associated factors of drug-drug interactions in elderly outpatients in a tertiary care hospital: a cross-sectional study based on three databases. Ann Transl Med 2023;11(1):17. 12. Day RO, Snowden L, McLachlan AJ. Life-threatening drug interactions: what the physician needs to know. Intern Med J 2017;47(5):501-512. 13. Takizawa M, Nakayama N, Ohishi Y, Tanaka K, Noguchi R, Saito Y, Hirano K, Komatsu Y. Physicians’ Perspectives on Prescription Alerts: A Journey Towards Reducing Fatigue. Cureus 2025;17(6):e86996. 14. Garcia-Vidal C, Sanjuan G, Puerta-Alcalde P, Moreno-García E, Soriano A. Artificial intelligence to support clinical decision-making processes. EBioMedicine 2019;46:27-29. 5. FIGURES Figure 1. Distribution of the Number of Medications by Hospital Location. Violin plots show the distribution of prescribed medications in intensive care unit (ICU) and ward patients. Boxes represent the interquartile range, black lines the median, and red diamonds the mean. Figure 2. Cross-classification of Interaction Severity by Uptodate® (UTD) and Artificial Intelligence (AI). Sankey diagram illustrating the distribution of drug interaction pairs according to severity categories assigned by UTD and the AI system. The thickness of each flow is proportional to the number of pairs in each category. Figure 3. Paired Dot Plot of Task Completion Times. Each line connects the times for the same drug–interaction check using UpToDate® (blue) and AI (orange). The consistent downward slope indicates that AI was faster in all 284 paired assessments. BULLET POINT SUMMARY What is already known: Electronic drug–drug interaction checkers often provide inconsistent detection and severity classifications. What this study adds: LLM-based AI provided complementary detection and significantly faster processing than UpToDate/Lexicomp Clinical significance: Combining AI with established checkers may improve efficiency, reduce alert fatigue, and enhance safety. AI reduces review time and personalizes recommendations using patient-specific information to guide care. Confict of interest : The authors report no conflicts of interest. Acknowledgments: The authors were solely responsible for data presented in this paper, with no additional acknowledgments to report. Funding: The authors declare that no financial support was provided for the development of this study. Information & Authors Information Version history V1 Version 1 23 September 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Authors Affiliations Noemi Santos de Oliveira Hospital Samel View all articles by this author Tricila Ribeiro de Matos Hospital Samel View all articles by this author Eduardo de Araújo Oliveira Hospital Samel View all articles by this author Luiza Fernanda Mendonça Nicolau Hospital Samel View all articles by this author Emilyn Barbosa de Oliveira Guerreiro Hospital Samel View all articles by this author Fernando Chiodini Machado Hospital Samel View all articles by this author Zaira Fernanda Martinho Nicolau [email protected] Hospital Samel View all articles by this author Metrics & Citations Metrics Article Usage 424 views 141 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Noemi Santos de Oliveira, Tricila Ribeiro de Matos, Eduardo de Araújo Oliveira, et al. Comparison of Artificial Intelligence and Lexicomp for Drug--Drug Interaction Screening in Hospitalized Patients. Authorea . 23 September 2025. DOI: https://doi.org/10.22541/au.175861495.50863378/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.175861495.50863378/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a0034b170e63f047',t:'MTc3OTUzMTI2Mg=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();