Deep Learning-based Prediction of Peptic Ulcer Diseases Caused by Nonsteroidal Anti-inflammatory Drugs Using Longitudinal Electronic Health Records

Deep Learning-based Prediction of Peptic Ulcer Diseases Caused by Nonsteroidal Anti-inflammatory Drugs Using Longitudinal Electronic Health Records | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Deep Learning-based Prediction of Peptic Ulcer Diseases Caused by Nonsteroidal Anti-inflammatory Drugs Using Longitudinal Electronic Health Records Joo Seong Kim, Junmo Kim, Hyunsoo Chung, Chaiho Shin, Sae-Hoon Kim, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5457261/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Nonsteroidal anti-inflammatory drugs (NSAID) are widely used to treat musculoskeletal disorders but are associated with peptic ulcers (PUs). Predicting the risk of PU in NSAID users is essential to minimize serious adverse effects such as bleeding and perforation. We developed and validated a deep learning-based model to predict the occurrence of NSAID-induced PU within 180 days after starting NSAID treatment using longitudinal electronic health records. The cohort included 125,930 patients prescribed NSAID for at least seven days. We used laboratory tests, medication history, and demographic information to train several machine learning and deep learning models, including random forests, gradient boosting machines (GBM), recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), and transformers. Endoscopy reports comprising free-text were used to more accurately determine the incidence of PU. The GRU model achieved the highest performance, with an AUROC of 0.941 for internal validation and 0.964 for external validation. Hemoglobin level, medication duration and aspirin use were significant predictors. Risk scores showed a sharp increase in risk two months before PU. We developed and validated robust predictive models for NSAID-induced PUs using longitudinal EHR data. These models may help inform clinical decision making for NSAID management and prevention of PU. Further studies are needed to improve these models and extend their application to diverse datasets. Health sciences/Diseases/Gastrointestinal diseases/Ulcers/Duodenal ulcers Health sciences/Diseases/Gastrointestinal diseases/Ulcers/Peptic ulcers Health sciences/Health care/Drug regulation Health sciences/Health care Health sciences/Risk factors Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Nonsteroidal anti-inflammatory drugs (NSAIDs) are primarily used as treatments for acute and chronic musculoskeletal disorders.( 1 ) NSAIDs alleviate pain, reduce both local and systemic inflammatory responses, and enhance musculoskeletal function and quality of life.( 2 – 4 ) Due to the aging population, there has been an increase in chronic musculoskeletal diseases.( 5 ) Consequently, NSAID use is rising, heightening concerns about drug-related side effects. NSAIDs inhibit the synthesis of prostaglandins from arachidonic acid by blocking Cyclooxygenase( 6 ), which comprises two isoforms: Cyclooxygenase-1, responsible for producing cytoprotective prostaglandins in the gastrointestinal tract, and cyclooxygenase-2, which is predominant at sites of inflammation.( 7 , 8 ) The adverse effects of NSAIDs can be attributed to these mechanisms. Known side effects of NSAIDs include gastrointestinal complications, cardiovascular events, and renal impairment. Among these, gastrointestinal complications are prevalent, affecting 10 ~ 60% of patients using NSAIDs.( 9 ) Gastrointestinal complications range from mild issues such as dyspepsia to severe complications including peptic ulcer (PU) with bleeding or perforation. The incidence of symptomatic PU caused by NSAIDs ranges from 2.7 ~ 4.5%, and complicated PU (cPU) associated with bleeding or perforation were reported in 1.0 ~ 1.5%.( 10 – 12 ) Risk factors for PU include age ≥ 65 years, a history of PU, high-dose NSAID use, and co-administration of aspirin, an antiplatelet agent, or a steroid.( 13 , 14 ) Because cPU may be associated with increased mortality, it is crucial for patients using NSAIDs to closely monitor for PU, and to detect and treat it early.( 15 ) NSAIDs are predominantly prescribed in outpatient settings, resulting in limited data availability. Consequently, there have been few studies to date focused on developing models to predict NSAID-induced PU.( 16 , 17 ) Recently, a machine learning model for predicting PU in patients using NSAIDs was introduced.( 18 ) The study targeted patients with osteoarthritis and reported an AUROC of 0.896, indicating that performance remains suboptimal. To the best of our knowledge, previous studies have not adequately elucidated crucial features for NSAID-induced PU prediction with rational justifications, nor have they validated results externally across multicenter populations. Furthermore, deep learning models trained on time-series electronic health record (EHR) data have not been utilized to predict PU in NSAID users. Thus, this study aimed to predict the occurrence of PU within 180 days following the initiation of NSAID treatment using longitudinal EHR data, which include laboratory tests and patient information such as demographics, comorbidities, and medications. We trained various machine learning and deep learning models to predict PU and compared the performances of these models. Data were collected from two geographically distinct tertiary hospitals. All trained PU prediction models underwent external validation. Additionally, important features were identified from the trained model, and variations in PU risk over time were compared between the PU and non-PU groups. Results Study cohort and patient characteristics This study included 423,278 patients at SNUH and 314,548 at SNUBH, spanning January 2001 to December 2022 and January 2004 to December 2021, respectively. All patients were aged over 18 and prescribed NSAIDs. Following our cohort criteria, we identified 491 and 77,712 patients in the PU and non-PU groups at SNUH, along with 334 and 47,393 at SNUBH. In the PU group, patients were mostly excluded due to continuous NSAID use, whereas in the non-PU group, exclusions were largely due to missing laboratory records. The detailed population flowcharts are depicted in Fig. 1 . The baseline characteristics of the included patients are presented in Table 1 . In both hospitals, patients in the PU group were older than those in the non-PU group, with males showing a higher prevalence of PU. The prevalence of PU was 0.63% (491/78,203) at SNUH and 0.70% (334/47,727) at SNUBH. The average time to PU occurrence was approximately four months in both facilities. The proportion of cPU was higher in SNUBH (63.17%) compared to SNUH (53.16%). cPU and non-PU classifications used in this study are detailed in Supplementary Table 2. Both hospitals showed similar trends in laboratory tests, with significant differences between the PU and non-PU groups across all tests. Additionally, the PU group at both hospitals exhibited a significantly higher prevalence of most comorbidities compared to the non-PU group. Regarding NSAID usage, both hospitals had significantly higher rates of aspirin use and significantly lower rates of aceclofenac and ibuprofen use in the PU group. SNUH showed a higher tendency toward using aceclofenac, ibuprofen, and naproxen, whereas SNUBH had a higher proportion of aspirin and celecoxib usage. Drug adherence was notably higher in the PU group in both hospitals. Higher usage of anticoagulants, antiplatelets, and antacids (H2-blocker, P-CAB, PPI) was observed in the PU group at both hospitals. Prescriptions primarily originated from internal medicine, followed by orthopedics. Table 1 . Baseline characteristics of the included patients. SNUH (n=78203) SNUBH (n=47727) Development and internal validation set External validation set PU Non-PU P-value PU Non-PU P-value (n=491) (n=77712) (n=334) (n=47393) age 66.50 ± 12.87 56.65 ± 15.22 <0.0001 69.73 ± 11.98 59.78 ± 14.26 <0.0001 male sex 255 (51.93) 36571 (47.06) 0.0330 188 (56.29) 23516 (49.62) 0.0157 Time to PU 118.43 ± 51.38 n/a n/a 123.28 ± 48.71 n/a n/a Complicated PU (cPU) 261 (53.16) n/a n/a 211 (63.17) n/a n/a Laboratory Tests Hemoglobin (g/dL) 12.27 ± 2.19 12.93 ± 1.93 <0.0001 12.47 ± 2.25 13.08 ± 1.95 <0.0001 Platelet (10 3 /μL) 220.87 ± 90.00 246.21 ± 89.53 <0.0001 220.52 ± 82.74 250.35 ± 86.88 <0.0001 BUN (mg/dL) 18.17 ± 8.49 15.39 ± 6.44 <0.0001 18.78 ± 8.12 15.89 ± 6.71 <0.0001 Creatinine (mg/dL) 1.00 ± 0.43 0.89 ± 0.32 <0.0001 1.04 ± 0.47 0.88 ± 0.35 <0.0001 Comorbidities Malignant Tumor 211 (42.97) 30170 (38.82) 0.0632 148 (44.31) 15215 (32.10) <0.0001 Myocardial Infarction 32 (6.52) 1710 (2.20) <0.0001 28 (8.38) 2095 (4.42) 0.0013 Uncomplicated Diabetes 133 (27.09) 5748 (7.40) <0.0001 101 (30.24) 4500 (9.50) <0.0001 Complicated Diabetes 52 (10.59) 1374 (1.77) <0.0001 63 (18.86) 2000 (4.22) <0.0001 Renal Disease 47 (9.57) 1962 (2.52) <0.0001 37 (11.08) 1164 (2.46) <0.0001 Heart Failure 31 (6.31) 1203 (1.55) <0.0001 35 (10.48) 1076 (2.27) <0.0001 Metastatic Carcinoma 63 (12.83) 4392 (5.65) <0.0001 31 (9.28) 2548 (5.38) 0.0033 Dementia 29 (5.91) 611 (0.79) <0.0001 18 (5.39) 519 (1.10) <0.0001 Cerebrovascular Disease 116 (23.63) 5690 (7.32) <0.0001 118 (35.33) 7020 (14.81) <0.0001 Peripheral Vascular Disease 35 (7.13) 1274 (1.64) <0.0001 27 (8.08) 1338 (2.82) <0.0001 Pulmonary Disease 47 (9.57) 2249 (2.89) <0.0001 48 (14.37) 1560 (3.29) <0.0001 Liver Disease 1 (0.20) 295 (0.38) 1.0000 3 (0.90) 61 (0.13) 0.0103 Mild Liver Disease 70 (14.26) 5439 (7.00) <0.0001 41 (12.28) 1950 (4.11) <0.0001 Paraplegia and Hemiplegia 11 (2.24) 198 (0.25) <0.0001 3 (0.90) 361 (0.76) 0.7447 Connective Tissue Disease 23 (4.68) 2755 (3.55) 0.1770 27 (8.08) 1456 (3.07) <0.0001 NSAIDs Usage Aceclofenac 98 (19.96) 24136 (31.06) <0.0001 55 (16.47) 12372 (26.11) <0.0001 Aspirin 249 (50.71) 26605 (34.24) <0.0001 200 (59.88) 18877 (39.83) <0.0001 Celecoxib 86 (17.52) 11865 (15.27) 0.1665 78 (23.35) 11194 (23.62) 0.9485 Diclofenac 0 (0.00) 28 (0.04) 1.0000 1 (0.30) 22 (0.05) 0.1492 Ibuprofen 61 (12.42) 12759 (16.42) 0.0169 18 (5.39) 4452 (9.39) 0.0106 Indomethacin 0 (0.00) 223 (0.29) 0.6538 1 (0.30) 141 (0.30) 1.0000 Ketoprofen 0 (0.00) 0 (0.00) 1.0000 10 (2.99) 2259 (4.77) 0.1542 Ketorolac 2 (0.41) 536 (0.69) 0.7796 0 (0.00) 157 (0.33) 0.6321 Loxoprofen 0 (0.00) 51 (0.07) 1.0000 1 (0.30) 19 (0.04) 0.1311 Meloxicam 26 (5.30) 5481 (7.05) 0.1558 11 (3.29) 3042 (6.42) 0.0178 Naproxen 54 (11.00) 7790 (10.02) 0.4515 16 (4.79) 2726 (5.75) 0.5543 Piroxicam 0 (0.00) 180 (0.23) 0.6336 9 (2.69) 1613 (3.40) 0.6474 Drug Adherence (%) 0.48 ± 0.26 0.39 ± 0.29 <0.0001 0.52 ± 0.26 0.44 ± 0.28 <0.0001 Other Medications Anticoagulants 45 (9.16) 4474 (5.76) 0.0025 10 (2.99) 653 (1.38) 0.0286 Antiplatelets 135 (27.49) 12815 (16.49) <0.0001 115 (34.43) 10372 (21.89) <0.0001 Steroids 167 (34.01) 28171 (36.25) 0.3227 139 (41.62) 18439 (38.91) 0.3114 H2 Blockers 246 (50.10) 38033 (48.94) 0.6186 94 (28.14) 8265 (17.44) <0.0001 P-CAB 9 (1.83) 446 (0.57) 0.0026 4 (1.20) 148 (0.31) 0.0225 Proton Pump Inhibitors 371 (75.56) 25968 (33.42) <0.0001 194 (58.08) 15281 (32.24) <0.0001 Prescribing Departments Anesthetics 5 (1.02) 407 (0.52) 0.1196 4 (1.20) 635 (1.34) 1.0000 General Surgery 16 (3.26) 4688 (6.03) 0.0074 7 (2.10) 897 (1.89) 0.6868 Internal Medicine 230 (46.84) 34895 (44.90) 0.3877 186 (55.69) 22706 (47.91) 0.0050 Neurosurgery 2 (0.41) 2867 (3.69) <0.0001 5 (1.50) 1817 (3.83) 0.0211 Obstetrics and Gynecology 13 (2.65) 1931 (2.48) 0.7704 2 (0.60) 515 (1.09) 0.5934 Orthopedics 53 (10.79) 8254 (10.62) 0.8832 21 (6.29) 7263 (15.33) <0.0001 Pediatrics 2 (0.41) 576 (0.74) 0.5943 0 (0.00) 6 (0.01) 1.0000 Others 170 (34.62) 24094 (31.00) 0.0868 109 (32.63) 13554 (28.60) 0.1141 *All data are shown as mean ± SD and n (%). *Abbreviations: PU=peptic ulcer, BUN=Blood urea nitrogen, P-CAB=potassium-competitive acid blocker PU prediction performance The performance of PU prediction models is presented in Table 2 . All models used in internal and external validations were selected after five-fold grid search cross-validation in the development process. The results of the development process are summarized in Supplementary Fig. 1. RNN with five layers and 256 nodes, LSTM with a single layer and 64 nodes, GRU with a single layer and 128 nodes, Transformer with a single layer and 32 nodes, and RETAIN with three layers and 256 nodes were employed for internal and external validations. Among these models, LSTM demonstrated the best performance. In the validation process, GRU displayed the most accurate prediction results in both internal and external validations with an AUROC of 0.941 and 0.964, respectively. Transformer attained the highest AUPRC of 0.188 for internal validation while GRU exhibited the highest AUPRC of 0.225 for external validation. We calculated all sensitivities, specificities, precisions, and F1-scores using Youden’s index( 19 ). Meanwhile, sensitivity, specificity, precision, and F1-score are influenced by the choice of classification threshold. Therefore, to fairly compare the rest of the metrics, we reported the results with a fixed sensitivity of 0.9 in Supplementary Table 1. The receiver operating characteristic (ROC) and precision-recall curves for all experiments are summarized in Fig. 2 . Table 2 Performance for detecting peptic ulcer. AUROC P-Value AUPRC Sensitivity Specificity Precision F1 Score Internal Validation Tree-Based Model Random Forest 0.834 (0.788–0.881) n/a 0.145 (0.099–0.191) 0.798 (0.792–0.804) 0.747 (0.740–0.754) 0.020 (0.018–0.022) 0.038 (0.036–0.042) GBM 0.860 (0.816–0.903) 0.0308 0.130 (0.086–0.173) 0.778 (0.771–0.784) 0.833 (0.827–0.839) 0.029 (0.026–0.032) 0.056 (0.052–0.059) RNN-Based Model Simple RNN 0.896 (0.860–0.933) 0.2648 0.184 (0.147–0.220) 0.879 (0.874–0.884) 0.806 (0.800-0.812) 0.028 (0.026–0.031) 0.054 (0.051–0.058) LSTM 0.937 (0.921–0.954) 0.0139 0.164 (0.148–0.181) 0.879 (0.874–0.884) 0.847 (0.842–0.853) 0.035 (0.033–0.038) 0.068 (0.064–0.072) GRU 0.941 (0.923–0.959) 0.0130 0.090 (0.072–0.108) 0.909 (0.904–0.913) 0.870 (0.865–0.875) 0.043 (0.040–0.046) 0.082 (0.077–0.086) Attention-Based Model Transformer 0.834 (0.786–0.881) 0.9934 0.188 (0.140–0.236) 0.677 (0.669–0.684) 0.869 (0.863–0.874) 0.032 (0.029–0.035) 0.061 (0.057–0.065) RETAIN 0.765 (0.709–0.820) 0.3016 0.057 (0.002–0.113) 0.616 (0.609–0.624) 0.833 (0.827–0.839) 0.023 (0.021–0.025) 0.044 (0.041–0.048) External Validation Tree-Based Model Random Forest 0.807 (0.782–0.832) n/a 0.106 (0.081–0.131) 0.659 (0.654–0.663) 0.813 (0.810–0.817) 0.024 (0.023–0.026) 0.047 (0.045–0.049) GBM 0.839 (0.816–0.861) 0.0048 0.074 (0.051–0.096) 0.650 (0.645–0.654) 0.873 (0.870–0.876) 0.035 (0.033–0.036) 0.066 (0.064–0.068) RNN-Based Model Simple RNN 0.882 (0.863–0.901) < 0.0001 0.112 (0.093–0.132) 0.808 (0.805–0.812) 0.796 (0.792–0.799) 0.027 (0.026–0.029) 0.053 (0.051–0.055) LSTM 0.908 (0.894–0.922) < 0.0001 0.155 (0.141–0.170) 0.841 (0.838–0.845) 0.830 (0.827–0.834) 0.034 (0.032–0.035) 0.065 (0.063–0.067) GRU 0.964 (0.957–0.970) < 0.0001 0.225 (0.218–0.231) 0.907 (0.905–0.910) 0.897 (0.895-0.900) 0.059 (0.057–0.061) 0.110 (0.108–0.113) Attention-Based Model Transformer 0.814 (0.786–0.842) 0.7930 0.189 (0.161–0.216) 0.674 (0.669–0.678) 0.825 (0.822–0.828) 0.026 (0.025–0.028) 0.051 (0.049–0.053) RETAIN 0.780 (0.752–0.807) 0.1823 0.057 (0.029–0.084) 0.695 (0.690–0.699) 0.736 (0.732–0.740) 0.018 (0.017–0.019) 0.036 (0.034–0.037) *Bold indicates the best performance and underlined denotes the second best. *P-values were derived using the DeLong method. *Abbreviations: AUROC = area under the receiver operating characteristic curve, AUPRC = area under the precision-recall curve, GBM = gradient boosting machine, RNN = recurrent neural network, LSTM = long short-term memory, GRU = gated recurrent unit Feature importance analysis We employed Deep SHAP( 20 ) to identify significant features for deep learning-based PU prediction. The SHAP values of laboratory tests, drugs, and patient information in both hospitals exhibited similar patterns, as illustrated in Fig. 3 . For this analysis, GRU, with a single layer and 128 nodes, was used as the reference model because it was selected as the best model in internal validation. Among laboratory measurements, hemoglobin was the most influential, followed by BUN, creatinine, and platelets. Aspirin had the highest SHAP value among NSAIDs. Antiplatelet agents, H2 blockers, and proton pump inhibitors were also identified as significant. In patient information, the duration of medication was prominently important. Risk variation over time To assess temporal differences in risk variation between PU and non-PU groups, we calculated continuous risk scores by sequentially entering timelines ranging from seven days to 210 days into the trained model. The risk score was the output value of the model. GRU served as a reference model in this process. Risk score variations over time are shown in Fig. 4 . For both hospitals, the non-PU group maintained its initial risk score throughout, while the risk score for the PU group increased from the second half. Initially, the PU group had a higher risk score than the non-PU group. Discussion In this study, we developed and validated several machine learning and deep learning-based PU prediction models in patients prescribed NSAIDs, using longitudinal EHR data for a total of 125,930 patients. For internal and external validation, we used large multicenter datasets from two geographically distinct tertiary hospitals, SNUH and SNUBH. The model trained with GRU exhibited superior prediction performance, with an AUROC of 0.941 for internal validation and 0.964 for external validation. Additionally, we identified influential features for PU prediction through Deep SHAP and assessed the temporal differences in risk variation between PU and non-PU groups. A few studies have reported on PU prediction models using artificial intelligence. Jeong et al. demonstrated PU detection performance with an AUROC of 0.896 using a GBM model based on data from the National Health Insurance Service (NHIS).( 18 ) The PU prediction model developed in this study showed superior performance compared to the previous study, likely due to the use of longitudinal EHR data. Although the NHIS data from South Korea do not include laboratory test results, the longitudinal EHR data, while smaller, provide more detailed information. Among laboratory tests, hemoglobin was identified as the most critical feature, whereas in drug and patient information, the use of aspirin and medication duration were pivotal. Hemoglobin is associated with PU, and anemia may occur, particularly due to bleeding caused by PU( 14 ). Aspirin is a well-documented risk factor for PU, and its use, either alone or in conjunction with NSAIDs, has been shown to increase the risk of PU( 21 , 22 ). The duration of medication also emerged as a crucial risk factor for PU, likely due to the higher prevalence of aspirin use among patients in the PU group. Aspirin is used in treating cardiovascular diseases and, more recently, as chemoprevention in certain cancers( 23 ). Therefore, it is often prescribed for extended periods. Additionally, drug adherence was significantly higher in the PU group. Patients requiring long-term NSAIDs are advised to use gastroprotective agents to prevent PU( 14 , 24 ). Both PPI and H2 blockers were identified as significant predictors of PU. Although PCAB has been recently approved in South Korea, data availability is limited. Nevertheless, PPI and H2RA were more commonly used in the PU group. The occurrence of PU despite prophylaxis with these agents suggests that patient factors such as age, underlying disease, duration and type of medication might play more significant roles in the development of PU. Baseline characteristics also indicated that the PU group was older and had more comorbidities compared to the non-PU group. Additionally, our model revealed a significant difference in risk scores between the PU and non-PU groups, with the PU group consistently exhibiting higher risk scores, and a marked increase two months prior to the PU occurrence in both hospitals. This supports the critical role of medication duration. Moreover, the risk score could facilitate proactive interventions to prevent PU in patients on NSAID therapy. Proactive clinical decisions, such as stopping NSAIDs or using a protective agent like a PPI, could substantially decrease PU incidence. Upon validation in further studies, the model could be considered for real-world clinical implementation. In this study, RNN-based models outperformed tree-based and attention-based models. Although Transformers are increasingly used in fields such as natural language processing (NLP), they proved to be less effective for training on numeric timelines in our dataset compared to RNN-based models. This disparity could be due to the different ways RNNs and Transformers process input data: RNNs handle data sequentially, whereas Transformers process data concurrently using positional encoding and learn variable relationships with less sensitivity to temporal and sequential dependences. Several studies involving numeric time-series data have successfully utilized RNNs to capture sequential changes.( 25 – 27 ) On the other hand, RETAIN, an interpretable attention-based neural network model designed for temporal EHR data with binary variables, did not perform as well on our dataset that featured numerous continuous numeric variables in timelines. This study has several limitations. First, it was a retrospective study, and the prevalence of PU in both cohorts was less than 1%. Despite using loss of focus to address class imbalance, extreme class imbalance was frequently observed in our deep learning model. The incidence of symptomatic PU due to NSAIDs has been reported to range from 2.7–4.5%. The low prevalence of PU in our data may have resulted from the high use of PPIs and H2 blockers in both the PU and non-PU groups, which likely had a protective effect against PU. Nevertheless, consistent trends were observed in both internal and external validations, enhancing the reliability of our findings. Second, our analysis does not include data on the history of PU disease from other healthcare providers. A history of PU is a well-known risk factor for PU, which helps determine prescribing PPI for PU prophylaxis when using NSAIDs( 14 ). However, this variable is challenging to investigate accurately in our EHR-based data. This represents a fundamental limitation of EHR-based data, as obtaining records from other healthcare providers is difficult. However, this limitation can be overcome by combining EHR data with NHIS data, and future studies will be conducted with more complete data using this method. Third, Helicobacter pylori (HP) was not included in the analysis. Eradication of HP is recommended to reduce the incidence of PU in patients using NSAIDs who have tested positive for HP( 28 ). Typically, if an ulcer is discovered during endoscopy in symptomatic patients, HP infection is tested for. However, there is a paucity of data on HP infections, particularly in control groups, due to the fact that individuals frequently do not undergo testing if they do not present with symptoms. Fourth, high-dose NSAIDs are also a risk factor for PU( 14 ), but NSAID dose was not included in our analysis. Since different NSAIDs have various doses, calculating equivalent doses from EHR-based data is problematic. Despite these limitations, this study has several strengths. We utilized large-scale EHR data to predict NSAIDs-induced ulcers. Few studies to date have employed EHR data from patient samples exceeding 100,000, and EHR data may better reflect real-world situations compared to NHIS data. Furthermore, we structured unstructured data, specifically endoscopy records, to accurately identify patients with peptic ulcer. In the PU group at SNUH, 20.2% (99/491) of the patients had a diagnosis of PU identified in their endoscopy record, without a corresponding diagnostic code for PU. In the PU group at SNUBH, 12.6% (42/334) had PU identified from the endoscopy records. Previous studies have used a diagnostic code-based approach to define peptic ulcer( 16 , 18 ). However, this approach is limited because the diagnosis of peptic ulcers may be overlooked if a more severe primary diagnosis is present. This study utilized endoscopic reports and natural language processing techniques to more accurately identify patients with PU, thus enhancing the reliability of the study results. In conclusion, we developed a high-performance, deep learning-based PU prediction model for patients receiving NSAIDs, internally and externally validated using data from two geographically distinct tertiary hospitals. The model can assist clinicians in making informed decisions about PU prevention and NSAID management, thereby improving patient outcomes. However, it is important to acknowledge that this study has limitations related to data collection, class imbalance, and missing data. Further research is necessary to validate these findings and potentially integrate additional data sources, such as NHIS, to enhance the comprehensiveness of the model. Methods Data curation This study utilized EHR data integrated into the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM).( 29 ) Established in the United States in 2008, OMOP is a public-private partnership aimed at enhancing the use of observational healthcare databases to evaluate medical products' effects.( 30 ) The OMOP CDM provides standardized data analysis solutions supporting the conversion of EHR from various sources into a unified data structure, facilitating large-scale data analysis.( 29 ) We extracted data from Seoul National University Hospital (SNUH) from January 2001 to December 2022 and Seoul National University Bundang Hospital (SNUBH) from January 2004 to December 2021. Data from SNUH were randomly divided into development (70% for training, 15% for validation) and internal validation (15%) datasets, while data from SNUBH were designated for external validation. We also used endoscopy reports comprising free-text to identify patients diagnosed with ulcer-related findings not recorded in the EHR system. We extracted all records containing the keyword "ulcer" and applied regular expression-based natural language processing techniques to find actual PU cases. To eliminate ulcers due to other causes, such as cancers and postoperative ulcers at the anastomosis site, we excluded reports with keywords like “ulcerofungating”, “ulceroinfiltrative”, “anastomosis”, “previous ulcer”, etc. Among the filtered reports, we identified ulcers associated with hemorrhage by using bleeding-related keywords (e.g., active bleeding, Forrest, F + Ia). Since negation terms commonly appeared with bleeding keywords, we only considered reports that mentioned bleeding keywords without negation terms as confirmed hemorrhage cases. This process was verified by two gastroenterologists. The Institutional Review Boards ( IRB ) at SNUH ( IRB No. 2308-101-1459) and SNUBH ( IRB No. X-2308-846-906) granted waivers of approval and informed consent, noting that the data were de-identified and sourced from observational electronic health records integrated into the OMOP CDM. This retrospective, multicenter study complied with the Declaration of Helsinki , the Korean Bioethics and Safety Act (Law No. 16372), and the Human Research Protection Program–Standard Operating Procedure of Seoul National University Hospital. Cohort definition and main outcomes Patients prescribed NSAIDs for at least 7 days and aged over 18 were identified and divided into PU and non-PU groups based on whether the first PU occurred within 180 days after NSAIDs usage.( 10 ) For the PU group, the index date was defined as the earliest start date of NSAIDs within the 180-day period prior to the first PU diagnosis, whereas for the non-PU group, the index date was the first start date of NSAIDs. The exclusion criteria for the PU group were as follows: First, patients with PU that occurred within seven days (washout period) after the index date were excluded, assuming that PU occurred due to other reasons rather than NSAIDs. Second, patients who had been taking NSAIDs continuously were also excluded, selecting only individuals who had not been prescribed NSAIDs within the 90 days before the index date, effectively identifying those who newly commenced NSAIDs. For both PU and non-PU groups, patients with a history of prior gastrointestinal surgery involving the stomach or duodenum were also excluded due to potential ulcer development at anastomosis sites.( 31 , 32 ) Additionally, candidates missing laboratory results within the 30 days leading up to the index date and from the index date until the PU diagnosis or 180 days post-index were omitted. For laboratory parameters, we chose variables routinely measured in outpatient settings associated with complicated PU, including hemoglobin, platelet count, blood urea nitrogen (BUN), and creatinine.( 33 ) A brief illustration of the index date definition is presented in Fig. 5 . Data preprocessing We collected laboratory tests, drugs, and patient information for each individual. We monitored not only NSAIDs but also anticoagulants, antiplatelet agents, steroids, H2 blockers, potassium-competitive acid blockers (PCAB), and proton pump inhibitors (PPI). Anticoagulants, antiplatelet agents, and steroids are identified as risk factors for NSAIDs-induced PU, while antacids such as H2 blockers, PCAB, and PPI serve as protective agents( 13 , 14 ). All OMOP CDM concept IDs utilized in this study are summarized in Supplementary Tables 2–6. We constructed a timeline of laboratory tests and medications for each patient, incorporating vectors of patient information. To construct a timeline, we first generated a table with 210 columns, implying the maximum monitoring duration with each column indicating a sequential date, the last column marking the final day. Values for each laboratory item were filled in corresponding dates. Blank parts between tests were filled by linear interpolation, and periods before and after tests were padded with the first and last recorded laboratory tests. For each drug, entries were marked with a one if administered on that day, or a zero otherwise. Patient information vectors included age, gender, comorbidity records, total duration of medication, the number of medications used, and prescribing department (anesthetics, general surgery, internal medicine, neurosurgery, obstetrics and gynecology, orthopedics, pediatrics, and others). All numeric variables were standardized prior to analysis. The adherence to NSAID medication was defined as follows( 18 ): Drug adherence = \(\:\frac{\sum\:Prescribed\:days\:of\:NSAIDs}{Total\:observation\:time}\) Model development We trained three types of models: a tree-based model including random forest and gradient boosting machine (GBM)( 34 ) as a baseline, a recurrent neural network (RNN)-based model including simple RNN, long short-term memory (LSTM)( 35 ), and gated recurrent unit (GRU)( 36 ), and an attention-based model including the Transformer( 37 ) and RETAIN( 38 ). Tree-based models, which were trained with one-dimensional vectors, utilized both the first and last columns of the timeline and patient information vector to incorporate the temporal history of laboratory tests and medication. In the case of RNN and attention-based models, a timeline was fed into the RNN or attention layer and a patient information vector was fed into a dense layer. The outputs from these layers were merged and then fed into another dense layer to classify cases as PU or normal. A brief illustration of PU prediction process using deep learning models is illustrated in Fig. 6 . We employed five-fold grid search cross-validation to identify the optimal model. For tree-based models, hyperparameters included the number of trees (ranging from 20 to 200 in increments of 20), the maximum tree depth (ranging from one to ten and infinite), and the maximum number of features considered for the best split (ranging from one to ten and the number of features). For RNN and attention-based models, the hyperparameters were the number of RNN and attention layers (one to five) and the number of nodes in all hidden layers (32, 64, 128, 256, and 512). In the case of Transformer models, the number of heads for multi-head attention was set to eight. The batch size was set to 256. We applied Focal loss( 39 ) to address the extreme class imbalance (less than 1%) in the dataset and used the Adam optimizer( 40 ) with a learning rate of 0.0001. The model training was limited to a maximum of 100 epochs with early stopping activated at a patience level of 20 based on AUROC performance. Experiments during the development phase were conducted with five different random seeds. The models demonstrating the best average performance were selected for both internal and external validations. Tree-based and deep learning-based models were implemented using Scikit-learn (version 1.0.2) and Pytorch (version 1.12.0) respectively, in Python (version 3.8.10). Identifying important features To identify key features for deep-learning based PU prediction, we utilized Deep SHAP( 20 ), an advanced form of the Deep LIFT algorithm( 41 ). Deep SHAP calculates attribution scores for all nodes and approximates Shapley values, providing insights into feature importance. For each patient's timeline, we computed the total absolute SHAP values for all days and averaged these values across patients to identify significant laboratory items and drugs. For patient information vectors, we averaged the absolute SHAP values across all patients to ascertain crucial patient information. The SHAP (version 0.42.1) package in Python (version 3.8.10) was used for SHAP value calculation. Statistical analysis The characteristics of PU and non-PU groups were compared using the Mann Whitney U test( 42 ) for continuous variables and the Fisher's exact test( 43 ) for categorical variables. To evaluate and compare model performances, we used AUROC and AUPRC. Confidence intervals for AUROC and AUPRC were determined using DeLong's method( 44 ), while those for sensitivity, specificity, precision, and F1-score were computed using Wilson's method( 45 ). Statistical significance was established at α = 0.05. All analyses were conducted using Scikit-learn (version 1.0.2) in Python (version 3.8.10). Data availability The raw data used in this study are not publicly available to preserve participant privacy. The data generated and analyzed during this study are available from the corresponding author upon reasonable request. Code availability The codes for deep learning training and statistical analysis are available at our private anonymous GitHub repository ( https://anonymous.4open.science/r/nsaids_ulcer-84F4 ). The repository will be made publicly available as the official GitHub repository after the manuscript is accepted. Declarations Competing interests The authors disclose no conflicts Author Contribution J.S.K. and J.K.L. contributed to the conceptualization and design of this study. J.K. handled and mainly analyzed the research data. All authors interpreted the results. J.K. constructed machine learning and deep learning models. J.K. and C.S. conducted statistical analysis. J.S.K. and J.K. wrote the original draft of the paper. H.C., S.H.L., S.H.K. and S.Y. revised the paper. J.S.K., H.C., S.H.L, S.H.K. and J.K.L. reviewed clinical evidence of this study. K.S.K. and S.Y. provided the data. J.S.K. and J.K. verified the quality of the data. J.K., K.S.K. and S.Y. had full access to all raw data. All authors had the final responsibility to submit for publication Acknowledgements This research was supported by a grant from Korea Institute of Drug Safety andRiskManagementin2023. J.K.L. received funding from Korea Institute of Drug Safety and Risk Management (No. 22233018800). J.K. was supported by a fellowship program of the AI Institute at Seoul National University (AIIS). The funders played no role in the study design, data collection, data analysis, data interpretation, or writing of this manuscript. References Laine L. Approaches to nonsteroidal anti-inflammatory drug use in the high-risk patient. Gastroenterology. 2001;120(3):594–606. Richette P, Doherty M, Pascual E, Barskova V, Becce F, Castañeda-Sanabria J, et al. 2016 updated EULAR evidence-based recommendations for the management of gout. Ann Rheum Dis. 2017;76(1):29–42. Dahl JB, Nielsen RV, Wetterslev J, Nikolajsen L, Hamunen K, Kontinen VK, et al. Post-operative analgesic effects of paracetamol, NSAIDs, glucocorticoids, gabapentinoids and their combinations: a topical review. Acta Anaesthesiol Scand. 2014;58(10):1165–81. Machado GC, Maher CG, Ferreira PH, Day RO, Pinheiro MB, Ferreira ML. Non-steroidal anti-inflammatory drugs for spinal pain: a systematic review and meta-analysis. Ann Rheum Dis. 2017;76(7):1269–78. Wongrakpanich S, Wongrakpanich A, Melhado K, Rangaswami J. A Comprehensive Review of Non-Steroidal Anti-Inflammatory Drug Use in The Elderly. Aging Dis. 2018;9(1):143–50. Zidar N, Odar K, Glavac D, Jerse M, Zupanc T, Stajer D. Cyclooxygenase in normal human tissues–is COX-1 really a constitutive isoform, and COX-2 an inducible isoform? J Cell Mol Med. 2009;13(9b):3753–63. Ricciotti E, FitzGerald GA. Prostaglandins and inflammation. Arterioscler Thromb Vasc Biol. 2011;31(5):986–1000. Rouzer CA, Marnett LJ. Cyclooxygenases: structural and functional insights. J Lipid Res. 2009;50 Suppl(Suppl):S29-34. Lazzaroni M, Bianchi Porro G. Gastrointestinal side-effects of traditional non-steroidal anti-inflammatory drugs and new formulations. Aliment Pharmacol Ther. 2004;20 Suppl 2:48–58. Silverstein FE, Faich G, Goldstein JL, Simon LS, Pincus T, Whelton A, et al. Gastrointestinal toxicity with celecoxib vs nonsteroidal anti-inflammatory drugs for osteoarthritis and rheumatoid arthritis: the CLASS study: A randomized controlled trial. Celecoxib Long-term Arthritis Safety Study. Jama. 2000;284(10):1247–55. Bombardier C, Laine L, Reicin A, Shapiro D, Burgos-Vargas R, Davis B, et al. Comparison of upper gastrointestinal toxicity of rofecoxib and naproxen in patients with rheumatoid arthritis. VIGOR Study Group. N Engl J Med. 2000;343(21):1520–8, 2 p following 8. Schnitzer TJ, Burmester GR, Mysler E, Hochberg MC, Doherty M, Ehrsam E, et al. Comparison of lumiracoxib with naproxen and ibuprofen in the Therapeutic Arthritis Research and Gastrointestinal Event Trial (TARGET), reduction in ulcer complications: randomised controlled trial. Lancet. 2004;364(9435):665–74. Lanza FL, Chan FK, Quigley EM. Guidelines for prevention of NSAID-related ulcer complications. Am J Gastroenterol. 2009;104(3):728–38. Joo MK, Park CH, Kim JS, Park JM, Ahn JY, Lee BE, et al. Clinical Guidelines for Drug-Related Peptic Ulcer, 2020 Revised Edition. Gut Liver. 2020;14(6):707–26. Kavitt RT, Lipowska AM, Anyane-Yeboa A, Gralnek IM. Diagnosis and Treatment of Peptic Ulcer Disease. Am J Med. 2019;132(4):447–56. Petersen J, Møller Hansen J, de Muckadell OBS, Dall M, Hallas J. A model to predict the risk of aspirin/non-steroidal anti-inflammatory drugs-related upper gastrointestinal bleeding for the individual patient. Basic Clin Pharmacol Toxicol. 2020;126(5):437–43. Guglielmi A, Ruzzenente A, Sandri M, Kind R, Lombardo F, Rodella L, et al. Risk assessment and prediction of rebleeding in bleeding gastroduodenal ulcer. Endoscopy. 2002;34(10):778–86. Jeong J, Han H, Ro DH, Han HS, Won S. Development of Prediction Model Using Machine-Learning Algorithms for Nonsteroidal Anti-inflammatory Drug-Induced Gastric Ulcer in Osteoarthritis Patients: Retrospective Cohort Study of a Nationwide South Korean Cohort. Clin Orthop Surg. 2023;15(4):678–89. Youden WJ. Index for rating diagnostic tests. Cancer. 1950;3(1):32–5. Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Advances in neural information processing systems. 2017;30. Sung JJ, Lau JY, Ching JY, Wu JC, Lee YT, Chiu PW, et al. Continuation of low-dose aspirin therapy in peptic ulcer bleeding: a randomized trial. Ann Intern Med. 2010;152(1):1–9. Chan FK, Ching JY, Suen BY, Tse YK, Wu JC, Sung JJ. Effects of Helicobacter pylori infection on long-term risk of peptic ulcer bleeding in low-dose aspirin users. Gastroenterology. 2013;144(3):528–35. Katona BW, Weiss JM. Chemoprevention of Colorectal Cancer. Gastroenterology. 2020;158(2):368–88. Rostom A, Moayyedi P, Hunt R. Canadian consensus guidelines on long-term nonsteroidal anti-inflammatory drug therapy and the need for gastroprotection: benefits versus risks. Aliment Pharmacol Ther. 2009;29(5):481–96. Li H, Yang S, Song Y, Luo Y, Li J, Zhou T. Spatial dynamic graph convolutional network for traffic flow forecasting. Applied Intelligence. 2023;53(12):14986–98. Choya T, Tamura N, Katayama S, Urano K, Yonezawa T, Kawaguchi N, editors. CrowdFlowTransformer: Capturing Spatio-Temporal Dependence for Forecasting Human Mobility. 2023 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops); 2023 13–17 March 2023. Lim H, Park J-S, Han Y, editors. Rare Sound Event Detection Using 1D Convolutional Recurrent Neural Networks. DCASE; 2017. Tang CL, Ye F, Liu W, Pan XL, Qian J, Zhang GX. Eradication of Helicobacter pylori infection reduces the incidence of peptic ulcer disease in patients using nonsteroidal anti-inflammatory drugs: a meta-analysis. Helicobacter. 2012;17(4):286–96. Hripcsak G, Duke JD, Shah NH, Reich CG, Huser V, Schuemie MJ, et al. Observational Health Data Sciences and Informatics (OHDSI): Opportunities for Observational Researchers. Stud Health Technol Inform. 2015;216:574–8. Zhang X, Wang L, Miao S, Xu H, Yin Y, Zhu Y, et al. Analysis of treatment pathways for three chronic diseases using OMOP CDM. Journal of Medical Systems. 2018;42(12):260. Sulieman I, Strobel O, Scharenberg C, Mihaljevic AM, Müller BM, Diener MK, et al. Symptomatic marginal ulcer after pancreatoduodenectomy. Surgery. 2020;168(1):67–71. Beran A, Shaear M, Al-Mudares S, Sharma I, Matar R, Al-Haddad M, et al. Predictors of marginal ulcer after gastric bypass: a systematic review and meta-analysis. J Gastrointest Surg. 2023;27(6):1066–77. Laursen SB, Oakland K, Laine L, Bieber V, Marmo R, Redondo-Cerezo E, et al. ABC score: a new risk score that accurately predicts mortality in acute upper and lower gastrointestinal bleeding: an international multicentre study. Gut. 2021;70(4):707–16. Friedman JH. Greedy Function Approximation: A Gradient Boosting Machine. The Annals of Statistics. 2001;29(5):1189–232. Hochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Computation. 1997;9(8):1735–80. Chung J, Gulcehre C, Cho K, Bengio Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:14123555. 2014. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is all you need. Advances in neural information processing systems. 2017;30. Choi E, Bahadori MT, Sun J, Kulas J, Schuetz A, Stewart W. Retain: An interpretable predictive model for healthcare using reverse time attention mechanism. Advances in neural information processing systems. 2016;29. Lin T-Y, Goyal P, Girshick R, He K, Dollár P, editors. Focal loss for dense object detection. Proceedings of the IEEE international conference on computer vision; 2017. Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980. 2014. Shrikumar A, Greenside P, Kundaje A. Learning Important Features Through Propagating Activation Differences. In: Doina P, Yee Whye T, editors. Proceedings of the 34th International Conference on Machine Learning; Proceedings of Machine Learning Research: PMLR; 2017. p. 3145–53. Mann HB, Whitney DR. On a Test of Whether one of Two Random Variables is Stochastically Larger than the Other. The Annals of Mathematical Statistics. 1947;18(1):50–60. Fisher RA. On the Interpretation of χ 2 from Contingency Tables, and the Calculation of P. Journal of the Royal Statistical Society. 1922;85(1):87–94. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the Areas under Two or More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach. Biometrics. 1988;44(3):837–45. Wilson EB. Probable Inference, the Law of Succession, and Statistical Inference. Journal of the American Statistical Association. 1927;22(158):209–12. Additional Declarations No competing interests reported. Supplementary Files Supplementaryfile20241017a.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5457261","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":385803666,"identity":"ec26617c-f436-42b5-bb83-2e05b491cfcc","order_by":0,"name":"Joo Seong Kim","email":"","orcid":"","institution":"Dongguk University Ilsan Hospital","correspondingAuthor":false,"prefix":"","firstName":"Joo","middleName":"Seong","lastName":"Kim","suffix":""},{"id":385803667,"identity":"5a7292c3-d159-45d6-ba24-10f3120cf717","order_by":1,"name":"Junmo Kim","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Junmo","middleName":"","lastName":"Kim","suffix":""},{"id":385803670,"identity":"e43b91f9-36df-4f2b-b5c4-1c2da04161a0","order_by":2,"name":"Hyunsoo Chung","email":"","orcid":"","institution":"Seoul National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hyunsoo","middleName":"","lastName":"Chung","suffix":""},{"id":385803672,"identity":"55dad866-e9b9-4436-abf0-3986302d0208","order_by":3,"name":"Chaiho Shin","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Chaiho","middleName":"","lastName":"Shin","suffix":""},{"id":385803673,"identity":"c686b45b-871b-47cd-9c43-1b6bb56e2e50","order_by":4,"name":"Sae-Hoon Kim","email":"","orcid":"","institution":"Seoul National University Bundang Hospital","correspondingAuthor":false,"prefix":"","firstName":"Sae-Hoon","middleName":"","lastName":"Kim","suffix":""},{"id":385803674,"identity":"80a0e854-932f-4231-84f3-2319a978e857","order_by":5,"name":"Sooyoung Yoo","email":"","orcid":"","institution":"Seoul National University Bundang Hospital","correspondingAuthor":false,"prefix":"","firstName":"Sooyoung","middleName":"","lastName":"Yoo","suffix":""},{"id":385803675,"identity":"b08981c8-9306-443e-b7c5-04a630a71da2","order_by":6,"name":"Sang Hyub Lee","email":"","orcid":"","institution":"Seoul National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Sang","middleName":"Hyub","lastName":"Lee","suffix":""},{"id":385803676,"identity":"eb002d68-3016-43f3-a025-215f261b8425","order_by":7,"name":"Kwangsoo Kim","email":"","orcid":"","institution":"Seoul National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Kwangsoo","middleName":"","lastName":"Kim","suffix":""},{"id":385803680,"identity":"1f2da214-0fa5-499c-93b8-9930440afd1a","order_by":8,"name":"Jun Kyu Lee","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAt0lEQVRIiWNgGAWjYPACG8YGKMuAWC1ppGs5TIIWeffeh49uVJyX3XDt8AOGj3sYjM0bCGgxPHPc2DjnzG3jDbfTDBhnPGMwkzlASMuMNDbp3LbbiRtu5zAw8xxgsJEg5DDD+c+AWv6dI0GLvAQbUEvDAbgWM4JaDHjSmI1zjiUbzwT65eCMAxLGhG1pP8b4OKfGTrbvdvLDBx8O2BjOIGjLASQOkE3QDqAtDYTVjIJRMApGwUgHADdjPlMjPZb7AAAAAElFTkSuQmCC","orcid":"","institution":"Dongguk University Ilsan Hospital","correspondingAuthor":true,"prefix":"","firstName":"Jun","middleName":"Kyu","lastName":"Lee","suffix":""}],"badges":[],"createdAt":"2024-11-15 03:23:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5457261/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5457261/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":70492133,"identity":"b7df0b65-43ce-453c-9246-a6f91d6828b1","added_by":"auto","created_at":"2024-12-03 17:23:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":110141,"visible":true,"origin":"","legend":"\u003cp\u003ePopulation flowchart\u003c/p\u003e","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-5457261/v1/3b00e89d501e673757f98d9b.png"},{"id":70492134,"identity":"33cfad1f-35e4-4c7e-8063-2e76570b13fd","added_by":"auto","created_at":"2024-12-03 17:23:40","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":278783,"visible":true,"origin":"","legend":"\u003cp\u003eROC and precision–recall curves of peptic ulcer detection models. Left side: internal validation; Right side: external validation\u003c/p\u003e","description":"","filename":"floatimage23.png","url":"https://assets-eu.researchsquare.com/files/rs-5457261/v1/6f934e50c12e53d7fea487ff.png"},{"id":70492406,"identity":"77e2b50c-64b3-49ff-aa65-765c9efd6589","added_by":"auto","created_at":"2024-12-03 17:31:41","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":182581,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP values of laboratory tests, drugs, and patient information for patients with peptic ulcer.\u003c/p\u003e","description":"","filename":"floatimage36.png","url":"https://assets-eu.researchsquare.com/files/rs-5457261/v1/17604ef6cd7ab9546f1f6751.png"},{"id":70492150,"identity":"b02e63e9-a920-462a-b8c9-138ecc71f83f","added_by":"auto","created_at":"2024-12-03 17:23:41","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":89463,"visible":true,"origin":"","legend":"\u003cp\u003eRisk score variation over time. The shaded part indicates standard deviation.\u003c/p\u003e","description":"","filename":"floatimage44.png","url":"https://assets-eu.researchsquare.com/files/rs-5457261/v1/d56e7d5534e92d05509c353e.png"},{"id":70492937,"identity":"331997f3-465c-4f50-9aca-83e1fa7f9252","added_by":"auto","created_at":"2024-12-03 17:39:40","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":67152,"visible":true,"origin":"","legend":"\u003cp\u003eIndex date definition\u003c/p\u003e","description":"","filename":"floatimage52.png","url":"https://assets-eu.researchsquare.com/files/rs-5457261/v1/0b0a8140f2415a5c2fd35904.png"},{"id":70492137,"identity":"1013fe6d-4a6e-4695-8a0c-568cfce2f3d9","added_by":"auto","created_at":"2024-12-03 17:23:40","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":51367,"visible":true,"origin":"","legend":"\u003cp\u003eThe process of peptic ulcer prediction using deep learning models.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-5457261/v1/660593b2fac738eac8d03197.png"},{"id":72903631,"identity":"a4db0f60-3140-45bf-afef-a1259585226b","added_by":"auto","created_at":"2025-01-03 13:17:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1388879,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5457261/v1/b464e0cc-c736-4386-926d-848b224343f3.pdf"},{"id":70492402,"identity":"98754cc4-58c5-4a52-b66c-389f4d342412","added_by":"auto","created_at":"2024-12-03 17:31:40","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1006667,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryfile20241017a.docx","url":"https://assets-eu.researchsquare.com/files/rs-5457261/v1/ef6b1c3b46498bf355bc3801.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning-based Prediction of Peptic Ulcer Diseases Caused by Nonsteroidal Anti-inflammatory Drugs Using Longitudinal Electronic Health Records","fulltext":[{"header":"Introduction","content":"\u003cp\u003eNonsteroidal anti-inflammatory drugs (NSAIDs) are primarily used as treatments for acute and chronic musculoskeletal disorders.(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) NSAIDs alleviate pain, reduce both local and systemic inflammatory responses, and enhance musculoskeletal function and quality of life.(\u003cspan additionalcitationids=\"CR3\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) Due to the aging population, there has been an increase in chronic musculoskeletal diseases.(\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) Consequently, NSAID use is rising, heightening concerns about drug-related side effects.\u003c/p\u003e \u003cp\u003eNSAIDs inhibit the synthesis of prostaglandins from arachidonic acid by blocking Cyclooxygenase(\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e), which comprises two isoforms: Cyclooxygenase-1, responsible for producing cytoprotective prostaglandins in the gastrointestinal tract, and cyclooxygenase-2, which is predominant at sites of inflammation.(\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e) The adverse effects of NSAIDs can be attributed to these mechanisms. Known side effects of NSAIDs include gastrointestinal complications, cardiovascular events, and renal impairment. Among these, gastrointestinal complications are prevalent, affecting 10\u0026thinsp;~\u0026thinsp;60% of patients using NSAIDs.(\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e) Gastrointestinal complications range from mild issues such as dyspepsia to severe complications including peptic ulcer (PU) with bleeding or perforation. The incidence of symptomatic PU caused by NSAIDs ranges from 2.7\u0026thinsp;~\u0026thinsp;4.5%, and complicated PU (cPU) associated with bleeding or perforation were reported in 1.0\u0026thinsp;~\u0026thinsp;1.5%.(\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e) Risk factors for PU include age\u0026thinsp;\u0026ge;\u0026thinsp;65 years, a history of PU, high-dose NSAID use, and co-administration of aspirin, an antiplatelet agent, or a steroid.(\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e) Because cPU may be associated with increased mortality, it is crucial for patients using NSAIDs to closely monitor for PU, and to detect and treat it early.(\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e)\u003c/p\u003e \u003cp\u003eNSAIDs are predominantly prescribed in outpatient settings, resulting in limited data availability. Consequently, there have been few studies to date focused on developing models to predict NSAID-induced PU.(\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e) Recently, a machine learning model for predicting PU in patients using NSAIDs was introduced.(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e) The study targeted patients with osteoarthritis and reported an AUROC of 0.896, indicating that performance remains suboptimal. To the best of our knowledge, previous studies have not adequately elucidated crucial features for NSAID-induced PU prediction with rational justifications, nor have they validated results externally across multicenter populations. Furthermore, deep learning models trained on time-series electronic health record (EHR) data have not been utilized to predict PU in NSAID users.\u003c/p\u003e \u003cp\u003eThus, this study aimed to predict the occurrence of PU within 180 days following the initiation of NSAID treatment using longitudinal EHR data, which include laboratory tests and patient information such as demographics, comorbidities, and medications. We trained various machine learning and deep learning models to predict PU and compared the performances of these models. Data were collected from two geographically distinct tertiary hospitals. All trained PU prediction models underwent external validation. Additionally, important features were identified from the trained model, and variations in PU risk over time were compared between the PU and non-PU groups.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy cohort and patient characteristics\u003c/h2\u003e \u003cp\u003eThis study included 423,278 patients at SNUH and 314,548 at SNUBH, spanning January 2001 to December 2022 and January 2004 to December 2021, respectively. All patients were aged over 18 and prescribed NSAIDs. Following our cohort criteria, we identified 491 and 77,712 patients in the PU and non-PU groups at SNUH, along with 334 and 47,393 at SNUBH. In the PU group, patients were mostly excluded due to continuous NSAID use, whereas in the non-PU group, exclusions were largely due to missing laboratory records. The detailed population flowcharts are depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe baseline characteristics of the included patients are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. In both hospitals, patients in the PU group were older than those in the non-PU group, with males showing a higher prevalence of PU. The prevalence of PU was 0.63% (491/78,203) at SNUH and 0.70% (334/47,727) at SNUBH. The average time to PU occurrence was approximately four months in both facilities. The proportion of cPU was higher in SNUBH (63.17%) compared to SNUH (53.16%). cPU and non-PU classifications used in this study are detailed in Supplementary Table\u0026nbsp;2. Both hospitals showed similar trends in laboratory tests, with significant differences between the PU and non-PU groups across all tests. Additionally, the PU group at both hospitals exhibited a significantly higher prevalence of most comorbidities compared to the non-PU group. Regarding NSAID usage, both hospitals had significantly higher rates of aspirin use and significantly lower rates of aceclofenac and ibuprofen use in the PU group. SNUH showed a higher tendency toward using aceclofenac, ibuprofen, and naproxen, whereas SNUBH had a higher proportion of aspirin and celecoxib usage. Drug adherence was notably higher in the PU group in both hospitals. Higher usage of anticoagulants, antiplatelets, and antacids (H2-blocker, P-CAB, PPI) was observed in the PU group at both hospitals. Prescriptions primarily originated from internal medicine, followed by orthopedics.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e. Baseline characteristics of the included patients.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"592\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"4\" style=\"width: 146px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" style=\"width: 223px;\"\u003e\n \u003cp\u003eSNUH (n=78203)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" style=\"width: 224px;\"\u003e\n \u003cp\u003eSNUBH (n=47727)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"3\" style=\"width: 223px;\"\u003e\n \u003cp\u003eDevelopment and internal validation set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" style=\"width: 224px;\"\u003e\n \u003cp\u003eExternal validation set\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003ePU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003eNon-PU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 49px;\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003ePU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003eNon-PU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 49px;\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e(n=491)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e(n=77712)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e(n=334)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e(n=47393)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e66.50\u0026nbsp;\u0026plusmn;\u0026nbsp;12.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e56.65\u0026nbsp;\u0026plusmn;\u0026nbsp;15.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e69.73\u0026nbsp;\u0026plusmn;\u0026nbsp;11.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e59.78\u0026nbsp;\u0026plusmn;\u0026nbsp;14.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003emale sex\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e255 (51.93)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e36571 (47.06)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0330\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e188 (56.29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e23516 (49.62)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0157\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eTime to PU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e118.43\u0026nbsp;\u0026plusmn;\u0026nbsp;51.38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003en/a\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003en/a\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e123.28\u0026nbsp;\u0026plusmn;\u0026nbsp;48.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003en/a\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003en/a\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eComplicated PU (cPU)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e261 (53.16)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003en/a\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003en/a\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e211 (63.17)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003en/a\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003en/a\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eLaboratory Tests\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eHemoglobin (g/dL)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e12.27\u0026nbsp;\u0026plusmn;\u0026nbsp;2.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e12.93\u0026nbsp;\u0026plusmn;\u0026nbsp;1.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e12.47\u0026nbsp;\u0026plusmn;\u0026nbsp;2.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e13.08\u0026nbsp;\u0026plusmn;\u0026nbsp;1.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003ePlatelet (10\u003csup\u003e3\u003c/sup\u003e/\u0026mu;L)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e220.87\u0026nbsp;\u0026plusmn;\u0026nbsp;90.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e246.21\u0026nbsp;\u0026plusmn;\u0026nbsp;89.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e220.52\u0026nbsp;\u0026plusmn;\u0026nbsp;82.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e250.35\u0026nbsp;\u0026plusmn;\u0026nbsp;86.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eBUN (mg/dL)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e18.17\u0026nbsp;\u0026plusmn;\u0026nbsp;8.49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e15.39\u0026nbsp;\u0026plusmn;\u0026nbsp;6.44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e18.78\u0026nbsp;\u0026plusmn;\u0026nbsp;8.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e15.89\u0026nbsp;\u0026plusmn;\u0026nbsp;6.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eCreatinine (mg/dL)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1.00\u0026nbsp;\u0026plusmn;\u0026nbsp;0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0.89\u0026nbsp;\u0026plusmn;\u0026nbsp;0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1.04\u0026nbsp;\u0026plusmn;\u0026nbsp;0.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0.88\u0026nbsp;\u0026plusmn;\u0026nbsp;0.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eComorbidities\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eMalignant Tumor\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e211 (42.97)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e30170 (38.82)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0632\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e148 (44.31)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e15215 (32.10)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eMyocardial Infarction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e32 (6.52)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1710 (2.20)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e28 (8.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2095 (4.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eUncomplicated Diabetes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e133 (27.09)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e5748 (7.40)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e101 (30.24)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e4500 (9.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eComplicated Diabetes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e52 (10.59)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1374 (1.77)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e63 (18.86)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2000 (4.22)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eRenal Disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e47 (9.57)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1962 (2.52)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e37 (11.08)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1164 (2.46)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eHeart Failure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e31 (6.31)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1203 (1.55)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e35 (10.48)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1076 (2.27)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eMetastatic Carcinoma\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e63 (12.83)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e4392 (5.65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e31 (9.28)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2548 (5.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0033\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eDementia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e29 (5.91)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e611 (0.79)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e18 (5.39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e519 (1.10)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eCerebrovascular Disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e116 (23.63)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e5690 (7.32)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e118 (35.33)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e7020 (14.81)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003ePeripheral Vascular Disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e35 (7.13)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1274 (1.64)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e27 (8.08)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1338 (2.82)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003ePulmonary Disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e47 (9.57)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2249 (2.89)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e48 (14.37)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1560 (3.29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eLiver Disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1 (0.20)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e295 (0.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e3 (0.90)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e61 (0.13)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0103\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eMild Liver Disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e70 (14.26)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e5439 (7.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e41 (12.28)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1950 (4.11)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eParaplegia and Hemiplegia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e11 (2.24)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e198 (0.25)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e3 (0.90)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e361 (0.76)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.7447\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eConnective Tissue Disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e23 (4.68)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2755 (3.55)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.1770\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e27 (8.08)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1456 (3.07)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eNSAIDs Usage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eAceclofenac\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e98 (19.96)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e24136 (31.06)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e55 (16.47)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e12372 (26.11)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eAspirin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e249 (50.71)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e26605 (34.24)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e200 (59.88)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e18877 (39.83)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eCelecoxib\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e86 (17.52)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e11865 (15.27)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.1665\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e78 (23.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e11194 (23.62)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.9485\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eDiclofenac\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0 (0.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e28 (0.04)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1 (0.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e22 (0.05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.1492\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eIbuprofen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e61 (12.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e12759 (16.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0169\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e18 (5.39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e4452 (9.39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0106\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eIndomethacin\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0 (0.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e223 (0.29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.6538\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1 (0.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e141 (0.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eKetoprofen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0 (0.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0 (0.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e10 (2.99)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2259 (4.77)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.1542\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eKetorolac\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2 (0.41)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e536 (0.69)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.7796\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0 (0.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e157 (0.33)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.6321\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eLoxoprofen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0 (0.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e51 (0.07)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1 (0.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e19 (0.04)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.1311\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eMeloxicam\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e26 (5.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e5481 (7.05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.1558\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e11 (3.29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e3042 (6.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0178\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eNaproxen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e54 (11.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e7790 (10.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.4515\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e16 (4.79)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2726 (5.75)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.5543\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003ePiroxicam\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0 (0.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e180 (0.23)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.6336\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e9 (2.69)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1613 (3.40)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.6474\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eDrug Adherence (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0.48\u0026nbsp;\u0026plusmn;\u0026nbsp;0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0.39\u0026nbsp;\u0026plusmn;\u0026nbsp;0.29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0.52\u0026nbsp;\u0026plusmn;\u0026nbsp;0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0.44\u0026nbsp;\u0026plusmn;\u0026nbsp;0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eOther Medications\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eAnticoagulants\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e45 (9.16)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e4474 (5.76)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0025\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e10 (2.99)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e653 (1.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0286\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eAntiplatelets\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e135 (27.49)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e12815 (16.49)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e115 (34.43)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e10372 (21.89)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eSteroids\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e167 (34.01)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e28171 (36.25)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.3227\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e139 (41.62)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e18439 (38.91)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.3114\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eH2 Blockers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e246 (50.10)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e38033 (48.94)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.6186\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e94 (28.14)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e8265 (17.44)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eP-CAB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e9 (1.83)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e446 (0.57)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0026\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e4 (1.20)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e148 (0.31)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0225\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eProton Pump Inhibitors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e371 (75.56)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e25968 (33.42)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e194 (58.08)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e15281 (32.24)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003ePrescribing Departments\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eAnesthetics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e5 (1.02)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e407 (0.52)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.1196\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e4 (1.20)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e635 (1.34)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eGeneral Surgery\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e16 (3.26)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e4688 (6.03)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0074\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e7 (2.10)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e897 (1.89)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.6868\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eInternal Medicine\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e230 (46.84)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e34895 (44.90)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.3877\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e186 (55.69)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e22706 (47.91)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0050\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eNeurosurgery\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2 (0.41)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2867 (3.69)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e5 (1.50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1817 (3.83)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0211\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eObstetrics and Gynecology\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e13 (2.65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e1931 (2.48)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.7704\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2 (0.60)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e515 (1.09)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.5934\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eOrthopedics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e53 (10.79)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e8254 (10.62)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.8832\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e21 (6.29)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e7263 (15.33)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u0026lt;0.0001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003ePediatrics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e2 (0.41)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e576 (0.74)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.5943\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e0 (0.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e6 (0.01)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1.0000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 146px;\"\u003e\n \u003cp\u003eOthers\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e170 (34.62)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e24094 (31.00)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.0868\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e109 (32.63)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\n \u003cp\u003e13554 (28.60)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e0.1141\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 0px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*All data are shown as mean \u0026plusmn; SD and n (%).\u003cbr\u003e\u0026nbsp;*Abbreviations: PU=peptic ulcer, BUN=Blood urea nitrogen, P-CAB=potassium-competitive acid blocker\u003c/p\u003e\n\u003ch3\u003ePU prediction performance\u003c/h3\u003e\n\u003cp\u003eThe performance of PU prediction models is presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. All models used in internal and external validations were selected after five-fold grid search cross-validation in the development process. The results of the development process are summarized in Supplementary Fig.\u0026nbsp;1. RNN with five layers and 256 nodes, LSTM with a single layer and 64 nodes, GRU with a single layer and 128 nodes, Transformer with a single layer and 32 nodes, and RETAIN with three layers and 256 nodes were employed for internal and external validations. Among these models, LSTM demonstrated the best performance. In the validation process, GRU displayed the most accurate prediction results in both internal and external validations with an AUROC of 0.941 and 0.964, respectively. Transformer attained the highest AUPRC of 0.188 for internal validation while GRU exhibited the highest AUPRC of 0.225 for external validation. We calculated all sensitivities, specificities, precisions, and F1-scores using Youden\u0026rsquo;s index(\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e). Meanwhile, sensitivity, specificity, precision, and F1-score are influenced by the choice of classification threshold. Therefore, to fairly compare the rest of the metrics, we reported the results with a fixed sensitivity of 0.9 in Supplementary Table\u0026nbsp;1. The receiver operating characteristic (ROC) and precision-recall curves for all experiments are summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance for detecting peptic ulcer.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eP-Value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUPRC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInternal Validation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTree-Based Model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.834\u003c/p\u003e \u003cp\u003e(0.788\u0026ndash;0.881)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003en/a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.145\u003c/p\u003e \u003cp\u003e(0.099\u0026ndash;0.191)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.798\u003c/p\u003e \u003cp\u003e(0.792\u0026ndash;0.804)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.747\u003c/p\u003e \u003cp\u003e(0.740\u0026ndash;0.754)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.020\u003c/p\u003e \u003cp\u003e(0.018\u0026ndash;0.022)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.038\u003c/p\u003e \u003cp\u003e(0.036\u0026ndash;0.042)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGBM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.860\u003c/p\u003e \u003cp\u003e(0.816\u0026ndash;0.903)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0308\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.130\u003c/p\u003e \u003cp\u003e(0.086\u0026ndash;0.173)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.778\u003c/p\u003e \u003cp\u003e(0.771\u0026ndash;0.784)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003cp\u003e(0.827\u0026ndash;0.839)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.029\u003c/p\u003e \u003cp\u003e(0.026\u0026ndash;0.032)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.056\u003c/p\u003e \u003cp\u003e(0.052\u0026ndash;0.059)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRNN-Based Model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSimple RNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.896\u003c/p\u003e \u003cp\u003e(0.860\u0026ndash;0.933)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.2648\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.184\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.147\u0026ndash;0.220)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.879\u003c/p\u003e \u003cp\u003e(0.874\u0026ndash;0.884)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.806\u003c/p\u003e \u003cp\u003e(0.800-0.812)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.028\u003c/p\u003e \u003cp\u003e(0.026\u0026ndash;0.031)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.054\u003c/p\u003e \u003cp\u003e(0.051\u0026ndash;0.058)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.937\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.921\u0026ndash;0.954)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0139\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.164\u003c/p\u003e \u003cp\u003e(0.148\u0026ndash;0.181)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.879\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.874\u0026ndash;0.884)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.847\u003c/p\u003e \u003cp\u003e(0.842\u0026ndash;0.853)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.035\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.033\u0026ndash;0.038)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.068\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.064\u0026ndash;0.072)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e0.941\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.923\u0026ndash;0.959)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0130\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.090\u003c/p\u003e \u003cp\u003e(0.072\u0026ndash;0.108)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.909\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.904\u0026ndash;0.913)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e0.870\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.865\u0026ndash;0.875)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e0.043\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.040\u0026ndash;0.046)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e0.082\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.077\u0026ndash;0.086)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAttention-Based Model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTransformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.834\u003c/p\u003e \u003cp\u003e(0.786\u0026ndash;0.881)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.9934\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.188\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.140\u0026ndash;0.236)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.677\u003c/p\u003e \u003cp\u003e(0.669\u0026ndash;0.684)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.869\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.863\u0026ndash;0.874)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.032\u003c/p\u003e \u003cp\u003e(0.029\u0026ndash;0.035)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.061\u003c/p\u003e \u003cp\u003e(0.057\u0026ndash;0.065)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRETAIN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.765\u003c/p\u003e \u003cp\u003e(0.709\u0026ndash;0.820)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.3016\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.057\u003c/p\u003e \u003cp\u003e(0.002\u0026ndash;0.113)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.616\u003c/p\u003e \u003cp\u003e(0.609\u0026ndash;0.624)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003cp\u003e(0.827\u0026ndash;0.839)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.023\u003c/p\u003e \u003cp\u003e(0.021\u0026ndash;0.025)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.044\u003c/p\u003e \u003cp\u003e(0.041\u0026ndash;0.048)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eExternal Validation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTree-Based Model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.807\u003c/p\u003e \u003cp\u003e(0.782\u0026ndash;0.832)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003en/a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.106\u003c/p\u003e \u003cp\u003e(0.081\u0026ndash;0.131)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.659\u003c/p\u003e \u003cp\u003e(0.654\u0026ndash;0.663)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.813\u003c/p\u003e \u003cp\u003e(0.810\u0026ndash;0.817)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.024\u003c/p\u003e \u003cp\u003e(0.023\u0026ndash;0.026)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.047\u003c/p\u003e \u003cp\u003e(0.045\u0026ndash;0.049)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGBM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.839\u003c/p\u003e \u003cp\u003e(0.816\u0026ndash;0.861)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.0048\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.074\u003c/p\u003e \u003cp\u003e(0.051\u0026ndash;0.096)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.650\u003c/p\u003e \u003cp\u003e(0.645\u0026ndash;0.654)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.873\u003c/p\u003e \u003cp\u003e(0.870\u0026ndash;0.876)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.035\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.033\u0026ndash;0.036)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.066\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.064\u0026ndash;0.068)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRNN-Based Model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSimple RNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.882\u003c/p\u003e \u003cp\u003e(0.863\u0026ndash;0.901)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.0001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.112\u003c/p\u003e \u003cp\u003e(0.093\u0026ndash;0.132)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.808\u003c/p\u003e \u003cp\u003e(0.805\u0026ndash;0.812)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.796\u003c/p\u003e \u003cp\u003e(0.792\u0026ndash;0.799)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.027\u003c/p\u003e \u003cp\u003e(0.026\u0026ndash;0.029)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.053\u003c/p\u003e \u003cp\u003e(0.051\u0026ndash;0.055)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLSTM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.908\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.894\u0026ndash;0.922)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.0001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.155\u003c/p\u003e \u003cp\u003e(0.141\u0026ndash;0.170)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.841\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.838\u0026ndash;0.845)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.830\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.827\u0026ndash;0.834)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.034\u003c/p\u003e \u003cp\u003e(0.032\u0026ndash;0.035)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.065\u003c/p\u003e \u003cp\u003e(0.063\u0026ndash;0.067)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGRU\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e0.964\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.957\u0026ndash;0.970)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.0001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.225\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.218\u0026ndash;0.231)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.907\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.905\u0026ndash;0.910)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e0.897\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.895-0.900)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e0.059\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.057\u0026ndash;0.061)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e0.110\u003c/b\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003e(0.108\u0026ndash;0.113)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAttention-Based Model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTransformer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.814\u003c/p\u003e \u003cp\u003e(0.786\u0026ndash;0.842)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.7930\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e0.189\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e(0.161\u0026ndash;0.216)\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.674\u003c/p\u003e \u003cp\u003e(0.669\u0026ndash;0.678)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.825\u003c/p\u003e \u003cp\u003e(0.822\u0026ndash;0.828)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.026\u003c/p\u003e \u003cp\u003e(0.025\u0026ndash;0.028)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.051\u003c/p\u003e \u003cp\u003e(0.049\u0026ndash;0.053)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRETAIN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.780\u003c/p\u003e \u003cp\u003e(0.752\u0026ndash;0.807)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.1823\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.057\u003c/p\u003e \u003cp\u003e(0.029\u0026ndash;0.084)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.695\u003c/p\u003e \u003cp\u003e(0.690\u0026ndash;0.699)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.736\u003c/p\u003e \u003cp\u003e(0.732\u0026ndash;0.740)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.018\u003c/p\u003e \u003cp\u003e(0.017\u0026ndash;0.019)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.036\u003c/p\u003e \u003cp\u003e(0.034\u0026ndash;0.037)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e*Bold indicates the best performance and underlined denotes the second best.\u003c/p\u003e \u003cp\u003e*P-values were derived using the DeLong method.\u003c/p\u003e \u003cp\u003e*Abbreviations: AUROC\u0026thinsp;=\u0026thinsp;area under the receiver operating characteristic curve, AUPRC\u0026thinsp;=\u0026thinsp;area under the precision-recall curve, GBM\u0026thinsp;=\u0026thinsp;gradient boosting machine, RNN\u0026thinsp;=\u0026thinsp;recurrent neural network, LSTM\u0026thinsp;=\u0026thinsp;long short-term memory, GRU\u0026thinsp;=\u0026thinsp;gated recurrent unit\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eFeature importance analysis\u003c/h3\u003e\n\u003cp\u003eWe employed Deep SHAP(\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e) to identify significant features for deep learning-based PU prediction. The SHAP values of laboratory tests, drugs, and patient information in both hospitals exhibited similar patterns, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. For this analysis, GRU, with a single layer and 128 nodes, was used as the reference model because it was selected as the best model in internal validation. Among laboratory measurements, hemoglobin was the most influential, followed by BUN, creatinine, and platelets. Aspirin had the highest SHAP value among NSAIDs. Antiplatelet agents, H2 blockers, and proton pump inhibitors were also identified as significant. In patient information, the duration of medication was prominently important.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eRisk variation over time\u003c/h3\u003e\n\u003cp\u003eTo assess temporal differences in risk variation between PU and non-PU groups, we calculated continuous risk scores by sequentially entering timelines ranging from seven days to 210 days into the trained model. The risk score was the output value of the model. GRU served as a reference model in this process. Risk score variations over time are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. For both hospitals, the non-PU group maintained its initial risk score throughout, while the risk score for the PU group increased from the second half. Initially, the PU group had a higher risk score than the non-PU group.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed and validated several machine learning and deep learning-based PU prediction models in patients prescribed NSAIDs, using longitudinal EHR data for a total of 125,930 patients. For internal and external validation, we used large \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"multicenter\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003emulticenter\u003c/em\u003e datasets from two geographically distinct tertiary hospitals, SNUH and SNUBH. The model trained with GRU exhibited superior prediction performance, with an AUROC of 0.941 for internal validation and 0.964 for external validation. Additionally, we identified influential features for PU prediction through Deep SHAP and assessed the temporal differences in risk variation between PU and non-PU groups.\u003c/p\u003e \u003cp\u003eA few studies have reported on PU prediction models using artificial intelligence. Jeong et al. demonstrated PU detection performance with an AUROC of 0.896 using a GBM model based on data from the National Health Insurance Service (NHIS).(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e) The PU prediction model developed in this study showed superior performance compared to the previous study, likely due to the use of longitudinal EHR data. Although the NHIS data from South Korea do not include laboratory test results, the longitudinal EHR data, while smaller, provide more detailed information.\u003c/p\u003e \u003cp\u003eAmong laboratory tests, hemoglobin was identified as the most critical feature, whereas in drug and \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"patient\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003epatient\u003c/em\u003e information, the use of aspirin and medication duration were pivotal. Hemoglobin is associated with PU, and anemia may occur, particularly due to bleeding caused by PU(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). Aspirin is a well-documented risk factor for PU, and its use, either alone or in conjunction with NSAIDs, has been shown to increase the risk of PU(\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). The duration of medication also emerged as a crucial risk factor for PU, likely due to the higher prevalence of aspirin use among patients in the PU group. Aspirin is used in treating cardiovascular diseases and, more recently, as chemoprevention in certain cancers(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). Therefore, it is often prescribed for extended periods. Additionally, drug adherence was significantly higher in the PU group. Patients requiring long-term NSAIDs are advised to use gastroprotective agents to prevent PU(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). Both PPI and H2 blockers were identified as significant predictors of PU. Although PCAB has been recently \u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"approved\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eapproved\u003c/em\u003e in South Korea, data availability is limited. Nevertheless, PPI and H2RA were more commonly used in the PU group. The occurrence of PU despite prophylaxis with these agents suggests that \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"patient\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003epatient\u003c/em\u003e factors such as age, underlying disease, duration and type of medication might play more significant roles in the development of PU. Baseline characteristics also indicated that the PU group was older and had more comorbidities compared to the non-PU group.\u003c/p\u003e \u003cp\u003eAdditionally, our model revealed a significant difference in risk scores between the PU and non-PU groups, with the PU group consistently exhibiting higher risk scores, and a marked increase two months prior to the PU occurrence in both hospitals. This supports the critical role of medication duration. Moreover, the risk score could facilitate proactive interventions to prevent PU in patients on NSAID therapy. Proactive clinical decisions, such as stopping NSAIDs or using a protective agent like a PPI, could substantially decrease PU incidence. Upon validation in further studies, the model could be considered for real-world clinical implementation.\u003c/p\u003e \u003cp\u003eIn this study, RNN-based models outperformed tree-based and attention-based models. Although Transformers are increasingly used in fields such as natural language processing (NLP), they proved to be less effective for training on numeric timelines in our dataset compared to RNN-based models. This disparity could be due to the different ways RNNs and Transformers process input data: RNNs handle data sequentially, whereas Transformers process data concurrently using positional encoding and learn variable relationships with less sensitivity to temporal and sequential dependences. Several studies involving numeric time-series data have successfully utilized RNNs to capture sequential changes.(\u003cspan additionalcitationids=\"CR26\" citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e–\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e) On the other hand, RETAIN, an interpretable attention-based neural network model designed for temporal EHR data with binary variables, did not perform as well on our dataset that featured numerous continuous numeric variables in timelines.\u003c/p\u003e \u003cp\u003eThis study has several limitations. First, it was a \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"retrospective*\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eretrospective\u003c/em\u003e study, and the prevalence of PU in both cohorts was less than 1%. Despite using loss of focus to address class imbalance, extreme class imbalance was frequently observed in our deep learning model. The incidence of symptomatic PU due to NSAIDs has been reported to range from 2.7–4.5%. The low prevalence of PU in our data may have resulted from the high use of PPIs and H2 blockers in both the PU and non-PU groups, which likely had a protective effect against PU. Nevertheless, consistent trends were observed in both internal and external validations, enhancing the reliability of our findings. Second, our analysis does not include data on the history of PU disease from other healthcare providers. A history of PU is a well-known risk factor for PU, which helps determine prescribing PPI for PU prophylaxis when using NSAIDs(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). However, this variable is challenging to investigate accurately in our EHR-based data. This represents a fundamental limitation of EHR-based data, as obtaining records from other healthcare providers is difficult. However, this limitation can be overcome by combining EHR data with NHIS data, and future studies will be conducted with more complete data using this method. Third, Helicobacter pylori (HP) was not included in the analysis. Eradication of HP is recommended to reduce the incidence of PU in patients using NSAIDs who have tested positive for HP(\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e). Typically, if an ulcer is discovered during endoscopy in symptomatic patients, HP infection is tested for. However, there is a paucity of data on HP infections, particularly in control groups, due to the fact that individuals frequently do not undergo testing if they do not present with symptoms. Fourth, high-dose NSAIDs are also a risk factor for PU(\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e), but NSAID dose was not included in our analysis. Since different NSAIDs have various doses, calculating equivalent doses from EHR-based data is problematic.\u003c/p\u003e \u003cp\u003eDespite these limitations, this study has several strengths. We utilized large-scale EHR data to predict NSAIDs-induced ulcers. Few studies to date have employed EHR data from \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"patient\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003epatient\u003c/em\u003e samples exceeding 100,000, and EHR data may better reflect real-world situations compared to NHIS data. Furthermore, we structured unstructured data, specifically endoscopy records, to accurately identify patients with peptic ulcer. In the PU group at SNUH, 20.2% (99/491) of the patients had a diagnosis of PU identified in their endoscopy record, without a corresponding diagnostic code for PU. In the PU group at SNUBH, 12.6% (42/334) had PU identified from the endoscopy records. Previous studies have used a diagnostic code-based approach to define peptic ulcer(\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). However, this approach is limited because the diagnosis of peptic ulcers may be overlooked if a more severe primary diagnosis is present. This study utilized endoscopic reports and natural language processing techniques to more accurately identify patients with PU, thus enhancing the reliability of the study results.\u003c/p\u003e \u003cp\u003eIn conclusion, we developed a high-performance, deep learning-based PU prediction model for patients receiving NSAIDs, internally and externally validated using data from two geographically distinct tertiary hospitals. The model can assist clinicians in making \u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"informed\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003einformed\u003c/em\u003e decisions about PU prevention and NSAID management, thereby improving \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"patient\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003epatient\u003c/em\u003e outcomes. However, it is important to acknowledge that this study has limitations related to data collection, class imbalance, and missing data. Further research is necessary to validate these findings and potentially integrate additional data sources, such as NHIS, to enhance the comprehensiveness of the model.\u003c/p\u003e "},{"header":"Methods","content":"\u003ch2\u003eData curation\u003c/h2\u003e\u003cp\u003eThis study utilized EHR data integrated into the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM).(\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e) Established in the United States in 2008, OMOP is a public-private partnership aimed at enhancing the use of observational healthcare databases to evaluate medical products' effects.(\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e) The OMOP CDM provides standardized data analysis solutions supporting the conversion of EHR from various sources into a unified data structure, facilitating large-scale data analysis.(\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e) We extracted data from Seoul National University Hospital (SNUH) from January 2001 to December 2022 and Seoul National University Bundang Hospital (SNUBH) from January 2004 to December 2021. Data from SNUH were randomly divided into development (70% for training, 15% for validation) and internal validation (15%) datasets, while data from SNUBH were designated for external validation.\u003c/p\u003e\u003cp\u003eWe also used endoscopy reports comprising free-text to identify patients diagnosed with ulcer-related findings not recorded in the EHR system. We extracted all records containing the keyword \"ulcer\" and applied regular expression-based natural language processing techniques to find actual PU \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"case*\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003ecases.\u003c/em\u003e To eliminate ulcers due to other causes, such as cancers and postoperative ulcers at the anastomosis site, we excluded reports with keywords like “ulcerofungating”, “ulceroinfiltrative”, “anastomosis”, “previous ulcer”, etc. Among the filtered reports, we identified ulcers associated with hemorrhage by using bleeding-related keywords (e.g., active bleeding, Forrest, F + Ia). Since negation terms commonly appeared with bleeding keywords, we only considered reports that mentioned bleeding keywords without negation terms as confirmed hemorrhage \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"case*\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003ecases.\u003c/em\u003e This process was verified by two gastroenterologists.\u003c/p\u003e\u003cp\u003eThe \u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"institutional review board*\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eInstitutional Review Boards\u003c/em\u003e (\u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"irb\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eIRB\u003c/em\u003e) at SNUH (\u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"irb\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eIRB\u003c/em\u003e No. 2308-101-1459) and SNUBH (\u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"irb\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eIRB\u003c/em\u003e No. X-2308-846-906) granted waivers of \u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"approval\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eapproval\u003c/em\u003e and \u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"informed\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003einformed\u003c/em\u003e \u003cem class=\"Highlight htf340ff0d-a602-4893-ae8d-b60ea075e112\" highlight=\"true\" htmatch=\"consent*\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003econsent,\u003c/em\u003e noting that the data were de-identified and sourced from observational electronic health records integrated into the OMOP CDM. This \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"retrospective*\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eretrospective,\u003c/em\u003e \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"multicenter\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003emulticenter\u003c/em\u003e study complied with the \u003cem class=\"Highlight ht29216696-c42e-4f00-932a-aea34347df6a\" highlight=\"true\" htmatch=\"declaration of helsinki\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eDeclaration of Helsinki\u003c/em\u003e, the Korean Bioethics and Safety Act (Law No. 16372), and the \u003cem class=\"Highlight ht71194251-f7a6-4c2d-a145-3d9f25b46662\" highlight=\"true\" htmatch=\"human*\" htloopnumber=\"879486669\" style=\"font-style: inherit;\"\u003eHuman\u003c/em\u003e Research Protection Program–Standard Operating Procedure of Seoul National University Hospital.\u003c/p\u003e\n\u003ch3\u003eCohort definition and main outcomes\u003c/h3\u003e\n\u003cp\u003ePatients prescribed NSAIDs for at least 7 days and aged over 18 were identified and divided into PU and non-PU groups based on whether the first PU occurred within 180 days after NSAIDs usage.(\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e) For the PU group, the index date was defined as the earliest start date of NSAIDs within the 180-day period prior to the first PU diagnosis, whereas for the non-PU group, the index date was the first start date of NSAIDs. The exclusion criteria for the PU group were as follows: First, patients with PU that occurred within seven days (washout period) after the index date were excluded, assuming that PU occurred due to other reasons rather than NSAIDs. Second, patients who had been taking NSAIDs continuously were also excluded, selecting only individuals who had not been prescribed NSAIDs within the 90 days before the index date, effectively identifying those who newly commenced NSAIDs. For both PU and non-PU groups, patients with a history of prior gastrointestinal surgery involving the stomach or duodenum were also excluded due to potential ulcer development at anastomosis sites.(\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e) Additionally, candidates missing laboratory results within the 30 days leading up to the index date and from the index date until the PU diagnosis or 180 days post-index were omitted. For laboratory parameters, we chose variables routinely measured in outpatient settings associated with complicated PU, including hemoglobin, platelet count, blood urea nitrogen (BUN), and creatinine.(\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e) A brief illustration of the index date definition is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eData preprocessing\u003c/h2\u003e \u003cp\u003eWe collected laboratory tests, drugs, and patient information for each individual. We monitored not only NSAIDs but also anticoagulants, antiplatelet agents, steroids, H2 blockers, potassium-competitive acid blockers (PCAB), and proton pump inhibitors (PPI). Anticoagulants, antiplatelet agents, and steroids are identified as risk factors for NSAIDs-induced PU, while antacids such as H2 blockers, PCAB, and PPI serve as protective agents(\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). All OMOP CDM concept IDs utilized in this study are summarized in Supplementary Tables\u0026nbsp;2\u0026ndash;6. We constructed a timeline of laboratory tests and medications for each patient, incorporating vectors of patient information. To construct a timeline, we first generated a table with 210 columns, implying the maximum monitoring duration with each column indicating a sequential date, the last column marking the final day. Values for each laboratory item were filled in corresponding dates. Blank parts between tests were filled by linear interpolation, and periods before and after tests were padded with the first and last recorded laboratory tests. For each drug, entries were marked with a one if administered on that day, or a zero otherwise. Patient information vectors included age, gender, comorbidity records, total duration of medication, the number of medications used, and prescribing department (anesthetics, general surgery, internal medicine, neurosurgery, obstetrics and gynecology, orthopedics, pediatrics, and others). All numeric variables were standardized prior to analysis. The adherence to NSAID medication was defined as follows(\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e):\u003c/p\u003e \u003cp\u003eDrug adherence = \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{\\sum\\:Prescribed\\:days\\:of\\:NSAIDs}{Total\\:observation\\:time}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eModel development\u003c/h2\u003e \u003cp\u003eWe trained three types of models: a tree-based model including random forest and gradient boosting machine (GBM)(\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e) as a baseline, a recurrent neural network (RNN)-based model including simple RNN, long short-term memory (LSTM)(\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e), and gated recurrent unit (GRU)(\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e), and an attention-based model including the Transformer(\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e) and RETAIN(\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e). Tree-based models, which were trained with one-dimensional vectors, utilized both the first and last columns of the timeline and patient information vector to incorporate the temporal history of laboratory tests and medication. In the case of RNN and attention-based models, a timeline was fed into the RNN or attention layer and a patient information vector was fed into a dense layer. The outputs from these layers were merged and then fed into another dense layer to classify cases as PU or normal. A brief illustration of PU prediction process using deep learning models is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe employed five-fold grid search cross-validation to identify the optimal model. For tree-based models, hyperparameters included the number of trees (ranging from 20 to 200 in increments of 20), the maximum tree depth (ranging from one to ten and infinite), and the maximum number of features considered for the best split (ranging from one to ten and the number of features). For RNN and attention-based models, the hyperparameters were the number of RNN and attention layers (one to five) and the number of nodes in all hidden layers (32, 64, 128, 256, and 512). In the case of Transformer models, the number of heads for multi-head attention was set to eight. The batch size was set to 256. We applied Focal loss(\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e) to address the extreme class imbalance (less than 1%) in the dataset and used the Adam optimizer(\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e) with a learning rate of 0.0001. The model training was limited to a maximum of 100 epochs with early stopping activated at a patience level of 20 based on AUROC performance. Experiments during the development phase were conducted with five different random seeds. The models demonstrating the best average performance were selected for both internal and external validations. Tree-based and deep learning-based models were implemented using Scikit-learn (version 1.0.2) and Pytorch (version 1.12.0) respectively, in Python (version 3.8.10).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eIdentifying important features\u003c/h2\u003e \u003cp\u003eTo identify key features for deep-learning based PU prediction, we utilized Deep SHAP(\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e), an advanced form of the Deep LIFT algorithm(\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e). Deep SHAP calculates attribution scores for all nodes and approximates Shapley values, providing insights into feature importance. For each patient's timeline, we computed the total absolute SHAP values for all days and averaged these values across patients to identify significant laboratory items and drugs. For patient information vectors, we averaged the absolute SHAP values across all patients to ascertain crucial patient information. The SHAP (version 0.42.1) package in Python (version 3.8.10) was used for SHAP value calculation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eThe characteristics of PU and non-PU groups were compared using the Mann Whitney U test(\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e) for continuous variables and the Fisher's exact test(\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e) for categorical variables. To evaluate and compare model performances, we used AUROC and AUPRC. Confidence intervals for AUROC and AUPRC were determined using DeLong's method(\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e), while those for sensitivity, specificity, precision, and F1-score were computed using Wilson's method(\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e). Statistical significance was established at α\u0026thinsp;=\u0026thinsp;0.05. All analyses were conducted using Scikit-learn (version 1.0.2) in Python (version 3.8.10).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eData availability\u003c/h2\u003e \u003cp\u003eThe raw data used in this study are not publicly available to preserve participant privacy. The data generated and analyzed during this study are available from the corresponding author upon reasonable request.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eCode availability\u003c/h2\u003e \u003cp\u003eThe codes for deep learning training and statistical analysis are available at our private anonymous GitHub repository (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://anonymous.4open.science/r/nsaids_ulcer-84F4\u003c/span\u003e\u003cspan address=\"https://anonymous.4open.science/r/nsaids_ulcer-84F4\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The repository will be made publicly available as the official GitHub repository after the manuscript is accepted.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors disclose no conflicts\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eJ.S.K. and J.K.L. contributed to the conceptualization and design of this study. J.K. handled and mainly analyzed the research data. All authors interpreted the results. J.K. constructed machine learning and deep learning models. J.K. and C.S. conducted statistical analysis. J.S.K. and J.K. wrote the original draft of the paper. H.C., S.H.L., S.H.K. and S.Y. revised the paper. J.S.K., H.C., S.H.L, S.H.K. and J.K.L. reviewed clinical evidence of this study. K.S.K. and S.Y. provided the data. J.S.K. and J.K. verified the quality of the data. J.K., K.S.K. and S.Y. had full access to all raw data. All authors had the final responsibility to submit for publication\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eThis research was supported by a grant from Korea Institute of Drug Safety andRiskManagementin2023. J.K.L. received funding from Korea Institute of Drug Safety and Risk Management (No. 22233018800). J.K. was supported by a fellowship program of the AI Institute at Seoul National University (AIIS). The funders played no role in the study design, data collection, data analysis, data interpretation, or writing of this manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLaine L. Approaches to nonsteroidal anti-inflammatory drug use in the high-risk patient. Gastroenterology. 2001;120(3):594\u0026ndash;606.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRichette P, Doherty M, Pascual E, Barskova V, Becce F, Casta\u0026ntilde;eda-Sanabria J, et al. 2016 updated EULAR evidence-based recommendations for the management of gout. Ann Rheum Dis. 2017;76(1):29\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDahl JB, Nielsen RV, Wetterslev J, Nikolajsen L, Hamunen K, Kontinen VK, et al. Post-operative analgesic effects of paracetamol, NSAIDs, glucocorticoids, gabapentinoids and their combinations: a topical review. Acta Anaesthesiol Scand. 2014;58(10):1165\u0026ndash;81.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMachado GC, Maher CG, Ferreira PH, Day RO, Pinheiro MB, Ferreira ML. Non-steroidal anti-inflammatory drugs for spinal pain: a systematic review and meta-analysis. Ann Rheum Dis. 2017;76(7):1269\u0026ndash;78.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWongrakpanich S, Wongrakpanich A, Melhado K, Rangaswami J. A Comprehensive Review of Non-Steroidal Anti-Inflammatory Drug Use in The Elderly. Aging Dis. 2018;9(1):143\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZidar N, Odar K, Glavac D, Jerse M, Zupanc T, Stajer D. Cyclooxygenase in normal human tissues\u0026ndash;is COX-1 really a constitutive isoform, and COX-2 an inducible isoform? J Cell Mol Med. 2009;13(9b):3753\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRicciotti E, FitzGerald GA. Prostaglandins and inflammation. Arterioscler Thromb Vasc Biol. 2011;31(5):986\u0026ndash;1000.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRouzer CA, Marnett LJ. Cyclooxygenases: structural and functional insights. J Lipid Res. 2009;50 Suppl(Suppl):S29-34.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLazzaroni M, Bianchi Porro G. Gastrointestinal side-effects of traditional non-steroidal anti-inflammatory drugs and new formulations. Aliment Pharmacol Ther. 2004;20 Suppl 2:48\u0026ndash;58.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSilverstein FE, Faich G, Goldstein JL, Simon LS, Pincus T, Whelton A, et al. Gastrointestinal toxicity with celecoxib vs nonsteroidal anti-inflammatory drugs for osteoarthritis and rheumatoid arthritis: the CLASS study: A randomized controlled trial. Celecoxib Long-term Arthritis Safety Study. Jama. 2000;284(10):1247\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBombardier C, Laine L, Reicin A, Shapiro D, Burgos-Vargas R, Davis B, et al. Comparison of upper gastrointestinal toxicity of rofecoxib and naproxen in patients with rheumatoid arthritis. VIGOR Study Group. N Engl J Med. 2000;343(21):1520\u0026ndash;8, 2 p following 8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchnitzer TJ, Burmester GR, Mysler E, Hochberg MC, Doherty M, Ehrsam E, et al. Comparison of lumiracoxib with naproxen and ibuprofen in the Therapeutic Arthritis Research and Gastrointestinal Event Trial (TARGET), reduction in ulcer complications: randomised controlled trial. Lancet. 2004;364(9435):665\u0026ndash;74.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLanza FL, Chan FK, Quigley EM. Guidelines for prevention of NSAID-related ulcer complications. Am J Gastroenterol. 2009;104(3):728\u0026ndash;38.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJoo MK, Park CH, Kim JS, Park JM, Ahn JY, Lee BE, et al. Clinical Guidelines for Drug-Related Peptic Ulcer, 2020 Revised Edition. Gut Liver. 2020;14(6):707\u0026ndash;26.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKavitt RT, Lipowska AM, Anyane-Yeboa A, Gralnek IM. Diagnosis and Treatment of Peptic Ulcer Disease. Am J Med. 2019;132(4):447\u0026ndash;56.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePetersen J, M\u0026oslash;ller Hansen J, de Muckadell OBS, Dall M, Hallas J. A model to predict the risk of aspirin/non-steroidal anti-inflammatory drugs-related upper gastrointestinal bleeding for the individual patient. Basic Clin Pharmacol Toxicol. 2020;126(5):437\u0026ndash;43.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuglielmi A, Ruzzenente A, Sandri M, Kind R, Lombardo F, Rodella L, et al. Risk assessment and prediction of rebleeding in bleeding gastroduodenal ulcer. Endoscopy. 2002;34(10):778\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJeong J, Han H, Ro DH, Han HS, Won S. Development of Prediction Model Using Machine-Learning Algorithms for Nonsteroidal Anti-inflammatory Drug-Induced Gastric Ulcer in Osteoarthritis Patients: Retrospective Cohort Study of a Nationwide South Korean Cohort. Clin Orthop Surg. 2023;15(4):678\u0026ndash;89.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYouden WJ. Index for rating diagnostic tests. Cancer. 1950;3(1):32\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg SM, Lee S-I. A unified approach to interpreting model predictions. Advances in neural information processing systems. 2017;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSung JJ, Lau JY, Ching JY, Wu JC, Lee YT, Chiu PW, et al. Continuation of low-dose aspirin therapy in peptic ulcer bleeding: a randomized trial. Ann Intern Med. 2010;152(1):1\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChan FK, Ching JY, Suen BY, Tse YK, Wu JC, Sung JJ. Effects of Helicobacter pylori infection on long-term risk of peptic ulcer bleeding in low-dose aspirin users. Gastroenterology. 2013;144(3):528\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKatona BW, Weiss JM. Chemoprevention of Colorectal Cancer. Gastroenterology. 2020;158(2):368\u0026ndash;88.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRostom A, Moayyedi P, Hunt R. Canadian consensus guidelines on long-term nonsteroidal anti-inflammatory drug therapy and the need for gastroprotection: benefits versus risks. Aliment Pharmacol Ther. 2009;29(5):481\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi H, Yang S, Song Y, Luo Y, Li J, Zhou T. Spatial dynamic graph convolutional network for traffic flow forecasting. Applied Intelligence. 2023;53(12):14986\u0026ndash;98.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoya T, Tamura N, Katayama S, Urano K, Yonezawa T, Kawaguchi N, editors. CrowdFlowTransformer: Capturing Spatio-Temporal Dependence for Forecasting Human Mobility. 2023 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops); 2023 13\u0026ndash;17 March 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLim H, Park J-S, Han Y, editors. Rare Sound Event Detection Using 1D Convolutional Recurrent Neural Networks. DCASE; 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang CL, Ye F, Liu W, Pan XL, Qian J, Zhang GX. Eradication of Helicobacter pylori infection reduces the incidence of peptic ulcer disease in patients using nonsteroidal anti-inflammatory drugs: a meta-analysis. Helicobacter. 2012;17(4):286\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHripcsak G, Duke JD, Shah NH, Reich CG, Huser V, Schuemie MJ, et al. Observational Health Data Sciences and Informatics (OHDSI): Opportunities for Observational Researchers. Stud Health Technol Inform. 2015;216:574\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, Wang L, Miao S, Xu H, Yin Y, Zhu Y, et al. Analysis of treatment pathways for three chronic diseases using OMOP CDM. Journal of Medical Systems. 2018;42(12):260.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSulieman I, Strobel O, Scharenberg C, Mihaljevic AM, M\u0026uuml;ller BM, Diener MK, et al. Symptomatic marginal ulcer after pancreatoduodenectomy. Surgery. 2020;168(1):67\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeran A, Shaear M, Al-Mudares S, Sharma I, Matar R, Al-Haddad M, et al. Predictors of marginal ulcer after gastric bypass: a systematic review and meta-analysis. J Gastrointest Surg. 2023;27(6):1066\u0026ndash;77.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLaursen SB, Oakland K, Laine L, Bieber V, Marmo R, Redondo-Cerezo E, et al. ABC score: a new risk score that accurately predicts mortality in acute upper and lower gastrointestinal bleeding: an international multicentre study. Gut. 2021;70(4):707\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFriedman JH. Greedy Function Approximation: A Gradient Boosting Machine. The Annals of Statistics. 2001;29(5):1189\u0026ndash;232.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Computation. 1997;9(8):1735\u0026ndash;80.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChung J, Gulcehre C, Cho K, Bengio Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:14123555. 2014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is all you need. Advances in neural information processing systems. 2017;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoi E, Bahadori MT, Sun J, Kulas J, Schuetz A, Stewart W. Retain: An interpretable predictive model for healthcare using reverse time attention mechanism. Advances in neural information processing systems. 2016;29.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin T-Y, Goyal P, Girshick R, He K, Doll\u0026aacute;r P, editors. Focal loss for dense object detection. Proceedings of the IEEE international conference on computer vision; 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980. 2014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShrikumar A, Greenside P, Kundaje A. Learning Important Features Through Propagating Activation Differences. In: Doina P, Yee Whye T, editors. Proceedings of the 34th International Conference on Machine Learning; Proceedings of Machine Learning Research: PMLR; 2017. p. 3145\u0026ndash;53.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMann HB, Whitney DR. On a Test of Whether one of Two Random Variables is Stochastically Larger than the Other. The Annals of Mathematical Statistics. 1947;18(1):50\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFisher RA. On the Interpretation of χ \u0026lt;\u0026thinsp;sup\u0026thinsp;\u0026gt;\u0026thinsp;2\u0026lt;/sup\u0026thinsp;\u0026gt;\u0026thinsp;from Contingency Tables, and the Calculation of P. Journal of the Royal Statistical Society. 1922;85(1):87\u0026ndash;94.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the Areas under Two or More Correlated Receiver Operating Characteristic Curves: A Nonparametric Approach. Biometrics. 1988;44(3):837\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilson EB. Probable Inference, the Law of Succession, and Statistical Inference. Journal of the American Statistical Association. 1927;22(158):209\u0026ndash;12.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-5457261/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5457261/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eNonsteroidal anti-inflammatory drugs (NSAID) are widely used to treat musculoskeletal disorders but are associated with peptic ulcers (PUs). Predicting the risk of PU in NSAID users is essential to minimize serious adverse effects such as bleeding and perforation. We developed and validated a deep learning-based model to predict the occurrence of NSAID-induced PU within 180 days after starting NSAID treatment using longitudinal electronic health records. The cohort included 125,930 patients prescribed NSAID for at least seven days. We used laboratory tests, medication history, and demographic information to train several machine learning and deep learning models, including random forests, gradient boosting machines (GBM), recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), and transformers. Endoscopy reports comprising free-text were used to more accurately determine the incidence of PU. The GRU model achieved the highest performance, with an AUROC of 0.941 for internal validation and 0.964 for external validation. Hemoglobin level, medication duration and aspirin use were significant predictors. Risk scores showed a sharp increase in risk two months before PU. We developed and validated robust predictive models for NSAID-induced PUs using longitudinal EHR data. These models may help inform clinical decision making for NSAID management and prevention of PU. Further studies are needed to improve these models and extend their application to diverse datasets.\u003c/p\u003e","manuscriptTitle":"Deep Learning-based Prediction of Peptic Ulcer Diseases Caused by Nonsteroidal Anti-inflammatory Drugs Using Longitudinal Electronic Health Records","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-03 17:23:29","doi":"10.21203/rs.3.rs-5457261/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7ecb1100-1f8c-4839-83b3-6c4124f76ddc","owner":[],"postedDate":"December 3rd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":41103825,"name":"Health sciences/Diseases/Gastrointestinal diseases/Ulcers/Duodenal ulcers"},{"id":41103826,"name":"Health sciences/Diseases/Gastrointestinal diseases/Ulcers/Peptic ulcers"},{"id":41103827,"name":"Health sciences/Health care/Drug regulation"},{"id":41103828,"name":"Health sciences/Health care"},{"id":41103829,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2025-01-03T13:09:00+00:00","versionOfRecord":[],"versionCreatedAt":"2024-12-03 17:23:29","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5457261","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5457261","identity":"rs-5457261","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}