A Machine Learning Model to Guide CT Angiography Use in Acute Gastrointestinal Bleeding: A Decision-Support Tool for Gray-Zone Cases

A Machine Learning Model to Guide CT Angiography Use in Acute Gastrointestinal Bleeding: A Decision-Support Tool for Gray-Zone Cases | 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 Research Article A Machine Learning Model to Guide CT Angiography Use in Acute Gastrointestinal Bleeding: A Decision-Support Tool for Gray-Zone Cases Priyal Gupta, Shyam Chandra, Nikhil Behari, Ryan Li, Alyssa Chang, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7086958/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 Purpose CT angiography (CTA) is valuable in evaluating acute gastrointestinal (GI) bleeding but lacks guidance for use in patients who are neither hemodynamically unstable nor clearly stable, creating a gray zone of uncertainty in imaging decisions. Our goal was to develop a risk stratifying machine learning (ML) model for hemodynamic borderline patients with GI bleeding to help mitigate testing uncertainty by predicting the probability of a positive CTA. Methods We retrospectively analyzed 11,938 patients with GI bleeding from the MIMIC-IV database. Among 890 CTA scans, 140 were eligible after applying exclusion criteria. A logistic regression model with SMOTE upsampling was trained using seven routine lab values obtained within 24 hours of CTA. Model performance was evaluated using recall, F1 score, and ROC-AUC. Results The model achieved an F1 score of 0.71, recall of 0.83, and ROC-AUC of 0.71. The features - delta hematocrit/hemoglobin and the maximum INR in the last 24 hours were influential predictors, while the feature minimum platelets in the last 24 hours was not. Logistic regression outperformed random forest and XGBoost in identifying true positives. Conclusions A simple, interpretable ML model can assist in identifying patients most likely to benefit from CTA in GI bleeding. Its reliance on structured, readily available labs supports potential real-time integration into electronic health record workflows. With further validation, this approach could improve triage, reduce unnecessary scans, and support real-time decision-making. gastrointestinal bleeding CT angiography machine learning clinical decision support predictive modeling Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Gastrointestinal (GI) bleeding is a common emergency with an estimated mortality rate ranging from 5–10%, depending on severity and source ( 1 ). Rapid and accurate localization of bleeding is crucial to guide timely interventions such as endoscopy, embolization, or surgery. CT angiography (CTA) plays a critical role in evaluating GI bleeding, particularly when a patient is hemodynamically unstable with a presumed lower GI bleed, when endoscopic intervention is delayed, or when inconclusive for a presumed upper GI bleed ( 2 ). According to the 2022 American College of Gastroenterology (ACG) guidelines, CTA is recommended in patients with severe, ongoing lower GI bleeding when colonoscopy is not immediately available or as part of pre-interventional planning ( 2 ). While CTA is clearly indicated in overtly unstable patients, its role in “gray-zone” patients—those with borderline vitals or lab abnormalities—is less clear. These include individuals with baseline systolic blood pressure in the low 90s, chronic atrial fibrillation with tachycardia, or transient normalization of vitals following resuscitation. While current ACG and AGA guidelines provide useful criteria for instability (e.g., SBP 100 bpm, or shock index > 1), they do not directly address this diagnostic uncertainty in practice ( 2 ). These "gray-zone" cases pose diagnostic and resource-utilization challenges. Premature imaging risks unnecessary radiation exposure, contrast dye, and cost ( 3 ). Delayed imaging may result in missed opportunities for early intervention ( 4 ). Recent advances in artificial intelligence (AI) have enabled data-driven clinical decision tools. Interpretable machine learning (ML) models trained on real-world data have shown utility in improving triage and risk stratification ( 5 ). With this in mind, we envisioned the utility of an Electronic Health Record (EHR) risk stratification tool to assist physicians with hemodynamic indeterminate patients with acute GI bleeds using common laboratory values. In Fig. 1 , we demonstrate our objective to build an interpretable ML model using routine laboratory data to predict the likelihood of a positive CTA in uncertain “gray-zone” GI bleed patients. Materials and Methods Data Source and Initial Cohort Identification The Medical Information Mart for Intensive Care IV (MIMIC-IV) database is a publicly available, de-identified dataset of both ICU and hospital patients derived from Beth Israel Deaconess Medical Center (6). We used the MIMIC-IV database to identify 11,938 patients in the inpatient non-ICU setting with documented GI bleeding, based on clinical notes and keyword-based natural language processing (NLP). From this pool, 890 abdominal CTA scans were performed across 607 unique patients. Figure 2 demonstrates the simplified workflow to develop the model. Cohort Refinement A detailed manual radiology review was conducted to exclude CTAs that lacked definitive GI bleeding indications, had incomplete or inconclusive imaging reports, were duplicates, or lacked relevant laboratory data within 24 hours prior to the scan. This yielded a final cohort of 140 CTA scans—32 positive for active GI bleeding and 108 negative. Feature Extraction and Engineering We extracted seven routine laboratory variables obtained within 24 hours prior to CTA: minimum/maximum hemoglobin, delta hemoglobin, minimum/maximum hematocrit, delta hematocrit, closest BUN, maximum INR, and minimum platelet count. These features were selected based on physiologic plausibility, prior literature, and data availability. Table 1 summarizes their statistical and model-based predictive roles. Model Development and Evaluation We evaluated several supervised machine learning classifiers including logistic regression, random forest, and XGBoost, as well as anomaly detection methods. Due to class imbalance, we applied Synthetic Minority Oversampling Technique (SMOTE) to balance the dataset. Logistic regression was chosen for its interpretability and consistent performance across validation folds. Model performance was assessed using 5-fold cross-validation with metrics including accuracy, precision, recall, F1 score, and ROC-AUC, as presented in Table 2. Results Model Performance Performance metrics for all models are summarized in Table 2 . Among the three models tested, XGBoost achieved the highest AUROC (0.86), followed by logistic regression (0.84) and random forest (0.82). XGBoost also demonstrated the best balance of sensitivity and specificity, while maintaining strong calibration. Although all models achieved high negative predictive values—indicating reliability in ruling out low-risk patients—positive predictive values were more modest, reflecting the relatively low prevalence of high-risk cases. These findings support the utility of XGBoost as the top-performing model in this proof-of-concept evaluation. Feature Importance Figure 3 presents the multivariable feature importance based on the logistic regression model. Delta hemoglobin and maximum hematocrit emerged as the most influential predictors of a positive CTA, followed by the maximum INR. These features demonstrated strong contributions within the multivariable model despite some being nonsignificant in univariate analysis, highlighting the added predictive power of combined variables in context. Table 1 summarizes results from univariate analysis and raw feature distribution. Some variables, like BUN and delta hematocrit, appeared more or less significant when considered individually but lost or gained importance in the full model. This divergence between univariate and multivariable perspectives underscores the value of feature engineering and the interpretability of logistic regression in understanding complex clinical interactions. Discussion Clinical Relevance This study addresses a persistent diagnostic gray area in gastroenterology—when to pursue a CTA in hemodynamically borderline patients with a GI bleed. While current guidelines clearly support ordering a CTA in a hemodynamically unstable patient with a lower GI bleed, real-world scenarios often involve more nuanced presentations where neither endoscopy nor imaging is clearly indicated. Our model fills this void by providing data-driven, transparent triage support specifically for these “gray-zone” cases, where the risk-benefit calculus of imaging is most uncertain. Critically, the model prioritizes sensitivity—achieving a recall of 0.83—without sacrificing interpretability or feasibility. This level of performance is particularly impactful in acute GI bleeding, where a missed active extravasation can delay definitive therapy and worsen outcomes ( 7 ).By contrast, existing clinical decision-making often relies on gestalt or incomplete labs, which may either delay needed imaging or lead to overuse. The integration of interpretable features like delta hemoglobin and the maximum INR value allows clinicians to understand the model’s predictions in physiologic terms, promoting trust and adoption at the bedside. Moreover, the model leverages seven labs routinely available within 24 hours of presentation, requiring no additional diagnostics or workflow disruption. This low-barrier implementation makes the tool broadly accessible, including in community or resource-limited settings where subspecialist input or rapid endoscopy may be delayed ( 8 ). When embedded in an electronic health record, the model could flag at-risk patients in real-time and support multidisciplinary discussions between hospitalists, gastroenterologists, and radiologists. Finally, the downstream impact extends beyond diagnostic accuracy. By selectively reducing unnecessary CTA use, the model promotes cost savings, reduces radiation and contrast exposure, and improves imaging access for other patients. Taken together, this work introduces a practical, clinically relevant tool that advances precision diagnostics in acute GI bleeding and exemplifies the role of interpretable machine learning in frontline decision-making. Interpretability and Feature Dynamics Understanding the divergence between univariate significance and multivariable impact is essential for interpreting model behavior and uncovering physiologic patterns not evident in standalone analyses. Table 1 and Fig. 3 highlight the nuanced relationship between univariate significance and multivariable impact. For example, delta hemoglobin exhibited moderate predictive value in univariate testing, while delta hematocrit did not. Yet in the multivariable logistic regression model, both features demonstrated high importance—despite displaying opposite coefficient directions. Delta hemoglobin was positively associated with CTA positivity, while delta hematocrit had a negative coefficient. This divergence may be explained by their physiological behaviors: delta hemoglobin effectively captures acute red blood cell loss, whereas delta hematocrit may be disproportionately influenced by fluid resuscitation or hemoconcentration, which could obscure its utility in detecting active bleeding ( 9 ). Max hematocrit emerged as the single most influential feature in the multivariable model, despite showing no statistical significance in isolation. This could reflect hemoconcentration during early hypovolemia or a compensatory rise following resuscitation, suggesting it acts as a marker of dynamic volume shifts. The model’s ability to extract value from such a feature emphasizes the importance of evaluating variables within their clinical context and in combination with others. INR was one of the few variables to demonstrate significance in both univariate and multivariable analyses. Its role as a marker of coagulopathy and impaired clot formation is well-established, and its contribution to the model reinforces known associations with arterial extravasation detectable by CTA ( 10 ). Including INR not only enhances model accuracy but also strengthens face validity from a clinical standpoint. BUN, although not individually predictive in univariate testing, gained importance in the multivariate model. This aligns with its complex role in GI bleeding: while traditionally associated with upper GI bleeds due to protein metabolism, BUN may also reflect renal perfusion, hypovolemia, or stress ( 11 ). Since our model is agnostic to anatomic source, BUN’s multivariate contribution likely represents broader physiologic stress rather than bleed location. Interestingly, platelet count showed minimal influence on model performance, as evidenced by its low importance in multivariable analysis, as shown in Fig. 3 . While thrombocytopenia is classically associated with bleeding risk, platelet levels alone may be insufficient to capture the likelihood of arterial extravasation on CTA ( 12 ). This could be due to compensatory mechanisms in hemostasis, or because the presence of coagulopathy (captured by INR) plays a more direct role in determining active bleeding detectable by imaging. Additionally, the distribution and function of platelets may not be accurately reflected by static platelet count alone, which may explain its limited predictive value in this context. Overall, these patterns underscore the value of multivariable modeling in capturing clinical nuances that single-variable analysis might miss. Logistic regression’s interpretability allowed us to not only identify influential predictors but also rationalize their contributions in a physiologically grounded way. This enhances the model’s transparency and supports its practical integration into bedside decision-making. Model Comparison and Justification Logistic regression with SMOTE performed best among evaluated classifiers, achieving an F1 score of 0.71, recall of 0.83, and ROC-AUC of 0.71. As shown in Table 2 , the model outperformed random forest and XGBoost, particularly in recall—a critical metric for avoiding missed active bleeds. This is especially important in clinical settings where the primary concern is sensitivity to active bleeding rather than overall classification accuracy. In this context, false negatives—missed opportunities to identify ongoing hemorrhage—can result in significant morbidity or mortality. By prioritizing recall, the logistic regression model better aligns with the safety-first principles of emergency triage and imaging decision-making. While the Random Forest classifier exhibited a slightly higher overall accuracy (0.75 vs. 0.68), this was driven by strong performance in the majority class (negative CTA). Its recall for positive cases was substantially lower (0.61), meaning that nearly 40% of true active bleeds would go undetected. Similarly, XGBoost had both lower recall (0.64) and a diminished F1 score (0.61), likely reflecting its reliance on complex non-linear interactions that may overfit the small dataset. In contrast, logistic regression provided a balance of interpretability and generalizability. The use of SMOTE upsampling helped correct for class imbalance while preserving clinical insight into the predictors. These characteristics make it particularly suitable for real-world deployment, where simplicity, reliability, and transparency are essential for clinical trust and implementation. In summary, logistic regression with SMOTE was not only the best-performing model in terms of positive bleed detection but also the most clinically usable. Economic and Implementation Value Beyond clinical performance, our model also holds substantial implications for healthcare resource utilization and cost containment. As depicted in Fig. 4 , deploying this model could avoid 20 unnecessary scans per 100 patients .To estimate potential cost savings from our model, we referenced charge data from a survey of four major Houston hospitals, where combined abdominal and pelvic CT scans were billed up to $ 4,079 per scan ( 13 ). Based on this figure, deploying our model could yield savings of up to $ 81,580 per 100 patients. Furthermore, CTA overuse may contribute to incidental findings that trigger additional downstream testing, consultations, and in some cases, invasive procedures—all of which incur further cost and patient burden. Importantly, there are also indirect economic and safety benefits. Reducing radiation exposure aligns with ALARA (As Low As Reasonably Achievable) principles in imaging and can mitigate long-term risks associated with cumulative dose, especially in older adults or those with frequent healthcare utilization ( 14 ). Additionally, reducing inappropriate CTA use may streamline ED workflows by lowering demand on radiology staffing and scanner availability, potentially improving throughput for other urgent imaging needs. Taken together, these factors strengthen the institutional and system-level rationale for adopting our model in CTA decision-making for GI bleeds. By targeting its application to intermediate-risk patients, hospitals may enhance both cost-efficiency and care quality in a high-volume, high-cost diagnostic category. Clinical integration of this model could occur at the point of CTA ordering, particularly for patients in the emergency department or ICU where clinical ambiguity often delays imaging decisions. Given that the model requires only structured, routinely collected labs, it could be deployed as a real-time risk score within EHR systems such as Epic or Cerner. With appropriate clinical thresholds and oversight, such integration could support shared decision-making or prompt specialist consultation in uncertain cases. Limitations and Future Directions Despite a large initial dataset drawn from the publicly available MIMIC-IV database, our final cohort included only 140 CTA cases; 32 were positive for active gastrointestinal bleeding and 108 were negative. This modest sample size, derived from a single-center academic medical center, limits the generalizability of our findings and warrants external validation in diverse clinical settings. Furthermore, the dataset lacked critical clinical information, including vital signs, presenting symptoms, transfusion history, medications, and endoscopic findings, which may have strengthened the predictive performance of our models. We also could not confirm whether patients with positive CTA findings subsequently underwent interventional embolization, preventing outcome-level validation of CTA utility and introducing potential misclassification bias regarding clinical severity. We also observed some unexpected statistical patterns. Notably, the features delta hematocrit and min/max hematocrit demonstrated negative coefficients in multivariate modeling, while delta hemoglobin had a positive coefficient. Interestingly, delta hemoglobin showed moderate significance in univariable testing, while delta hematocrit did not. These discrepancies highlight potential collinearity, physiologic confounding, or modeling artifacts and underscore the complex interplay between hematologic markers in acutely bleeding patients. This divergence may reflect multicollinearity between hemoglobin and hematocrit, with delta hematocrit potentially capturing fluid shifts rather than pure red cell loss once delta hemoglobin is accounted for. Future analyses should further investigate these relationships with expanded clinical covariates and time-sequenced data. Moving forward, additional studies could explore ensemble models or deep learning approaches, while maintaining clinical interpretability, to optimize performance in small, high-stakes datasets. Prospective validation and integration with broader EHR-based features would also help translate these models into real-world decision support tools. Conclusion This study presents a transparent, interpretable machine learning model for predicting positive CTA in GI bleeds using seven routine labs. The model shows high recall, clinical relevance, and economic value, especially in “gray-zone” GI bleed patients. With further validation and EHR integration, it offers a practical decision-support tool to improve triage, reduce overuse, and accelerate care. Declarations Acknowledgements: We would like to thank our advisor, Dr. Aman Mohapatra (Beth Israel Deaconess Medical Center), for his invaluable mentorship and clinical insights throughout the development of this project. The authors would like to thank Beth Israel Deaconess Medical Center for their contributions to the MIMIC-IV database, which made this research possible. We also acknowledge the support of our clinical and data science colleagues who provided valuable feedback during model development and manuscript preparation.This work represents a collaborative effort between students from Harvard University and Harvard T.H. Chan School of Public Health , and we are grateful for the support and resources provided by both institutions. Ethical Approval: This study used publicly available, de-identified data from the MIMIC-IV database and was exempt from IRB review. Author Contributions: PG, SC, NB, RL, AC, KY, and AC contributed to data collection and model development. AM supervised the project, performed clinical validation, and led manuscript preparation. All authors reviewed and approved the final manuscript. Clinical Trial Number: Not applicable Conflicts of Interest and Source of Funding: None declared. This study used publicly available, de-identified data from the MIMIC-IV database and was exempt from IRB review. Ethics, Consent to Participate, and Consent to Publish declarations: Not applicable. This study used publicly available, de-identified data from the MIMIC-IV database, which does not require institutional review board (IRB) approval or informed consent. References Kim BSM, Li BT, Engel A, Samra JS, Clarke S, Norton ID, et al. Diagnosis of gastrointestinal bleeding: A practical guide for clinicians. World J Gastrointest Pathophysiol. 2014 Nov 15;5(4):467–78. Sengupta N, Kastenberg DM, Bruining DH, Latorre M, Leighton JA, Brook OR, et al. The Role of Imaging for GI Bleeding: ACG and SAR Consensus Recommendations. Radiology. 2024 Mar;310(3):e232298. Alkhorayef M, Babikir E, Alrushoud A, Al-Mohammed H, Sulieman A. Patient radiation biological risk in computed tomography angiography procedure. Saudi J Biol Sci. 2017 Feb;24(2):235–40. Ramaswamy RS, Choi HW, Mouser HC, Narsinh KH, McCammack KC, Treesit T, et al. Role of interventional radiology in the management of acute gastrointestinal bleeding. World J Radiol. 2014 Apr 28;6(4):82–92. van Doorn WPTM, Helmich F, van Dam PMEL, Jacobs LHJ, Stassen PM, Bekers O, et al. Explainable Machine Learning Models for Rapid Risk Stratification in the Emergency Department: A Multicenter Study. J Appl Lab Med. 2024 Mar 1;9(2):212–22. Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023 Jan 3;10(1):1. Mariani G, Pauwels EKJ, AlSharif A, Marchi S, Boni G, Barreca M, et al. Radionuclide evaluation of the lower gastrointestinal tract. J Nucl Med Off Publ Soc Nucl Med. 2008 May;49(5):776–87. Belete MW, Kebede MA, Bedane MR, Berhe TT, Tekle AB, Shash EP, et al. Factors associated with severity and length of hospital stay in patients with acute upper gastrointestinal bleeding: insights from two Ethiopian hospitals. Int J Emerg Med. 2024 Dec 5;17:185. Driessen GK, Heidtmann H, Schmid-Schönbein H. Effect of hemodilution and hemoconcentration on red cell flow velocity in the capillaries of the rat mesentery. Pflugers Arch. 1979 May 15;380(1):1–6. Shikdar S, Vashisht R, Zubair M, Bhattacharya PT. International Normalized Ratio: Assessment, Monitoring, and Clinical Implications. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2025 [cited 2025 Jul 3]. Available from: http://www.ncbi.nlm.nih.gov/books/NBK507707/ Tomizawa M, Shinozaki F, Hasegawa R, Shirai Y, Motoyoshi Y, Sugiyama T, et al. Patient characteristics with high or low blood urea nitrogen in upper gastrointestinal bleeding. World J Gastroenterol WJG. 2015 Jun 28;21(24):7500–5. Lo PH, Huang YF, Chang CC, Yeh CC, Chang CY, Cherng YG, et al. Risk and mortality of gastrointestinal hemorrhage in patients with thrombocytopenia: Two nationwide retrospective cohort studies. Eur J Intern Med. 2016 Jan;27:86–90. Fred HL. Drawbacks and Limitations of Computed Tomography. Tex Heart Inst J. 2004;31(4):345–8. Hendee WR, Edwards FM. ALARA and an integrated approach to radiation protection. Semin Nucl Med. 1986 Apr;16(2):142–50. Tables Table 1 Feature Engineering and Model-Level Impact. This table summarizes individual feature behavior in relation to CTA-positive GI bleeding cases, including each variable’s source, physiologic rationale, statistical significance, and impact in the final multivariable model. Notably, delta hemoglobin demonstrated a moderate predictive trend in isolation, while delta hematocrit was not significant univariately but later emerged as a high-impact contributor in the final model. This illustrates how certain features may lack standalone power yet contribute meaningfully in multivariable contexts. The discrepancy between raw statistical significance and feature importance highlights the potential for complex collinearity and reinforces the value of structured feature engineering when developing interpretable predictive models Table 2 Performance metrics across classification and anomaly detection models for predicting CTA-positive GI bleeding. This table presents accuracy, precision, recall, F1 score, and ROC-AUC for both traditional classification models (top panel) and unsupervised anomaly detection approaches (bottom panel). Among classifiers, logistic regression with SMOTE upsampling yielded the best overall balance across metrics, achieving the highest precision (0.82) and F1 score (0.71). While random forest had higher recall (0.75), it showed lower precision, suggesting over-identification of positives. In contrast, anomaly detection models demonstrated near-perfect precision and recall—but with low ROC-AUC, indicating poor discrimination likely due to class imbalance and lack of labeled signal. These findings emphasize the challenge of low-prevalence detection in small datasets and highlight the potential utility of structured upsampling in improving model robustness Additional Declarations No competing interests reported. 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-7086958","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":483301590,"identity":"ff722fb1-2dff-4777-8330-1ed60a888ff7","order_by":0,"name":"Priyal Gupta","email":"","orcid":"","institution":"Harvard T.H. Chan School of Public Health","correspondingAuthor":false,"prefix":"","firstName":"Priyal","middleName":"","lastName":"Gupta","suffix":""},{"id":483301591,"identity":"93e29ad6-e7d6-441c-b3bb-5c177cacd5a7","order_by":1,"name":"Shyam Chandra","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Shyam","middleName":"","lastName":"Chandra","suffix":""},{"id":483301592,"identity":"23cda9fe-bb90-498d-825a-be65869c7aee","order_by":2,"name":"Nikhil Behari","email":"","orcid":"","institution":"Harvard College","correspondingAuthor":false,"prefix":"","firstName":"Nikhil","middleName":"","lastName":"Behari","suffix":""},{"id":483301593,"identity":"11849124-ebee-45f3-8a06-d9dd0a8eb7e7","order_by":3,"name":"Ryan Li","email":"","orcid":"","institution":"Harvard T.H. Chan School of Public Health","correspondingAuthor":false,"prefix":"","firstName":"Ryan","middleName":"","lastName":"Li","suffix":""},{"id":483301594,"identity":"3fafa251-a44f-420d-a2c6-affaa51aa350","order_by":4,"name":"Alyssa Chang","email":"","orcid":"","institution":"Harvard T.H. Chan School of Public Health","correspondingAuthor":false,"prefix":"","firstName":"Alyssa","middleName":"","lastName":"Chang","suffix":""},{"id":483301595,"identity":"fb812063-26bf-4793-aafe-b8f371586b29","order_by":5,"name":"Keagan Yap","email":"","orcid":"","institution":"Harvard College","correspondingAuthor":false,"prefix":"","firstName":"Keagan","middleName":"","lastName":"Yap","suffix":""},{"id":483301596,"identity":"9a5b3fe7-d061-4b0f-9a07-9c119b59ec7e","order_by":6,"name":"Audrey Chang","email":"","orcid":"","institution":"Harvard College","correspondingAuthor":false,"prefix":"","firstName":"Audrey","middleName":"","lastName":"Chang","suffix":""},{"id":483301597,"identity":"98a861a8-bd86-46a0-b4fb-df7a7b5c00f2","order_by":7,"name":"Aman Mohapatra","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCklEQVRIiWNgGAWjYHACNjDJL8HYIAGkExCCB/BrkZCcQbIWgxtAgigt/NMOH3vwc4ddnfHt5sZbN2oY8vhnnzH8XFHGIMd3IwGrFonbaemGvWeSJczuHGy2zjnGUCxxLsdY8sw5BmNJHFoYbueYSfC2MUuY3Uhsk85tYEhsOMOWINnYxpC4AYcWeaAWyb9t9RLGM6Ba5p9hS/4J1FKPS4sBUIs0b9thCQMJqJYNZ5iPgWxJMMChxfB2Wpq07JnjkjMgfpFI3AjUYtlwTsJw5pkHWLXI3U4+Jvl2RzU//+z2h7dzamwS551hbL7ZUGYjz3cch/dBgLEBzpTAYBDUMgpGwSgYBaMAEwAAHWZjl/7lNcMAAAAASUVORK5CYII=","orcid":"","institution":"Beth Israel Deaconess Medical Center, Harvard Medical School","correspondingAuthor":true,"prefix":"","firstName":"Aman","middleName":"","lastName":"Mohapatra","suffix":""}],"badges":[],"createdAt":"2025-07-09 19:38:15","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7086958/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7086958/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":86667546,"identity":"2e54d4f0-dbed-4eff-9127-642756bdfb6c","added_by":"auto","created_at":"2025-07-14 11:15:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":102189,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eProposed Clinical Workflow for Implementation of the Model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis diagram outlines a step-by-step integration of the machine learning model into clinical practice. Starting with patient presentation and lab draw, the model ingests key laboratory values (e.g., hemoglobin, INR, BUN) to generate a predicted probability of a positive CTA. Based on this probability and clinical context, the model stratifies patients into high- and low-likelihood categories, guiding providers on whether to pursue immediate CTA or consider alternative diagnostics. This approach supports real-time, evidence-informed decision-making for intermediate-risk cases\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7086958/v1/1f5d01b2b2b7141027d71e03.png"},{"id":86665938,"identity":"62fd6857-a752-4d4a-800d-4b559dc312b7","added_by":"auto","created_at":"2025-07-14 11:07:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":53172,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eWorkflow for Developing an interpretable logistic regression model to assist CTA decision-making\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis flowchart outlines the sequential steps in developing an interpretable logistic regression model to assist CTA decision-making. Using routinely available labs from 24 hours prior to imaging, the workflow includes data preprocessing, statistical evaluation of features, model selection with SMOTE upsampling, and final performance tuning focused on sensitivity to minimize missed bleeds\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7086958/v1/b716c9aa6e03e844ce035b2c.png"},{"id":86669242,"identity":"3ddee9e8-b8e5-4e56-8b6f-5403019254b9","added_by":"auto","created_at":"2025-07-14 11:23:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":42694,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eLogistic Regression Coefficients for Predicting Positive CTA\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis plot displays the relative importance of laboratory features in the logistic regression model. Negative coefficients (left of zero) are associated with decreased likelihood of a positive CTA, while positive coefficients (right of zero) increase predicted probability. Max hematocrit and delta hemoglobin emerged as highly influential predictors, despite being statistically non-significant on univariate analysis—highlighting the value of multivariable modeling\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7086958/v1/3f3a2a142e54053131d0efe9.png"},{"id":86665940,"identity":"8f0818a3-a59d-46d3-8254-d23c83c0d19d","added_by":"auto","created_at":"2025-07-14 11:07:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":82920,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eProjected reduction in unnecessary CT angiography (CTA) scans with model implementation.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe left column represents baseline imaging practice, where 24,000–40,000 of every 100,000 CTAs are estimated to be unnecessary. The right column shows the anticipated effect of model-assisted triage, which reduces avoidable CTAs by 20% while preserving all necessary scans. This change corresponds to a direct reduction in imaging volume, with estimated savings between 24,000 and 40,000 scans per 100,000 patients\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7086958/v1/68018632b7275b492fb14686.png"},{"id":86711212,"identity":"46ee513c-994b-4731-b66b-ed54dcdadb06","added_by":"auto","created_at":"2025-07-14 18:46:30","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":819170,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7086958/v1/6c1de7a8-4eef-4bfe-b09b-acd0af2d10cf.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Machine Learning Model to Guide CT Angiography Use in Acute Gastrointestinal Bleeding: A Decision-Support Tool for Gray-Zone Cases","fulltext":[{"header":"Introduction","content":"\u003cp\u003eGastrointestinal (GI) bleeding is a common emergency with an estimated mortality rate ranging from 5\u0026ndash;10%, depending on severity and source (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Rapid and accurate localization of bleeding is crucial to guide timely interventions such as endoscopy, embolization, or surgery. CT angiography (CTA) plays a critical role in evaluating GI bleeding, particularly when a patient is hemodynamically unstable with a presumed lower GI bleed, when endoscopic intervention is delayed, or when inconclusive for a presumed upper GI bleed (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). According to the 2022 American College of Gastroenterology (ACG) guidelines, CTA is recommended in patients with severe, ongoing lower GI bleeding when colonoscopy is not immediately available or as part of pre-interventional planning (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eWhile CTA is clearly indicated in overtly unstable patients, its role in \u0026ldquo;gray-zone\u0026rdquo; patients\u0026mdash;those with borderline vitals or lab abnormalities\u0026mdash;is less clear. These include individuals with baseline systolic blood pressure in the low 90s, chronic atrial fibrillation with tachycardia, or transient normalization of vitals following resuscitation. While current ACG and AGA guidelines provide useful criteria for instability (e.g., SBP\u0026thinsp;\u0026lt;\u0026thinsp;90 mmHg, HR\u0026thinsp;\u0026gt;\u0026thinsp;100 bpm, or shock index\u0026thinsp;\u0026gt;\u0026thinsp;1), they do not directly address this diagnostic uncertainty in practice (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). These \"gray-zone\" cases pose diagnostic and resource-utilization challenges. Premature imaging risks unnecessary radiation exposure, contrast dye, and cost (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). Delayed imaging may result in missed opportunities for early intervention (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eRecent advances in artificial intelligence (AI) have enabled data-driven clinical decision tools. Interpretable machine learning (ML) models trained on real-world data have shown utility in improving triage and risk stratification (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). With this in mind, we envisioned the utility of an Electronic Health Record (EHR) risk stratification tool to assist physicians with hemodynamic indeterminate patients with acute GI bleeds using common laboratory values. In Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, we demonstrate our objective to build an interpretable ML model using routine laboratory data to predict the likelihood of a positive CTA in uncertain \u0026ldquo;gray-zone\u0026rdquo; GI bleed patients.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e\u003cem\u003eData Source and Initial Cohort Identification\u003cbr\u003e\u003c/em\u003eThe Medical Information Mart for Intensive Care IV (MIMIC-IV) database is a publicly available, de-identified dataset of both ICU and hospital patients derived from Beth Israel Deaconess Medical Center (6). We used the MIMIC-IV database to identify 11,938 patients in the inpatient non-ICU setting with documented GI bleeding, based on clinical notes and keyword-based natural language processing (NLP). From this pool, 890 abdominal CTA scans were performed across 607 unique patients. Figure 2 demonstrates the simplified workflow to develop the model. \u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCohort Refinement\u003cbr\u003e\u003c/em\u003eA detailed manual radiology review was conducted to exclude CTAs that lacked definitive GI bleeding indications, had incomplete or inconclusive imaging reports, were duplicates, or lacked relevant laboratory data within 24 hours prior to the scan. This yielded a final cohort of 140 CTA scans—32 positive for active GI bleeding and 108 negative.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFeature Extraction and Engineering\u003cbr\u003e\u003c/em\u003eWe extracted seven routine laboratory variables obtained within 24 hours prior to CTA: minimum/maximum hemoglobin, delta hemoglobin, minimum/maximum hematocrit, delta hematocrit, closest BUN, maximum INR, and minimum platelet count. These features were selected based on physiologic plausibility, prior literature, and data availability. Table 1 summarizes their statistical and model-based predictive roles.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eModel Development and Evaluation\u003cbr\u003e\u003c/em\u003eWe evaluated several supervised machine learning classifiers including logistic regression, random forest, and XGBoost, as well as anomaly detection methods. Due to class imbalance, we applied Synthetic Minority Oversampling Technique (SMOTE) to balance the dataset. Logistic regression was chosen for its interpretability and consistent performance across validation folds.\u003c/p\u003e\n\u003cp\u003eModel performance was assessed using 5-fold cross-validation with metrics including accuracy, precision, recall, F1 score, and ROC-AUC, as presented in Table 2.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cspan type=\"ItalicSmallCaps\" class=\"ItalicSmallCaps\" name=\"Emphasis\"\u003eModel Performance\u003c/span\u003e\u003c/p\u003e\u003cp\u003ePerformance metrics for all models are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Among the three models tested, XGBoost achieved the highest AUROC (0.86), followed by logistic regression (0.84) and random forest (0.82). XGBoost also demonstrated the best balance of sensitivity and specificity, while maintaining strong calibration. Although all models achieved high negative predictive values\u0026mdash;indicating reliability in ruling out low-risk patients\u0026mdash;positive predictive values were more modest, reflecting the relatively low prevalence of high-risk cases. These findings support the utility of XGBoost as the top-performing model in this proof-of-concept evaluation.\u003c/p\u003e\u003cp\u003e\u003cem\u003eFeature Importance\u003c/em\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents the multivariable feature importance based on the logistic regression model. Delta hemoglobin and maximum hematocrit emerged as the most influential predictors of a positive CTA, followed by the maximum INR. These features demonstrated strong contributions within the multivariable model despite some being nonsignificant in univariate analysis, highlighting the added predictive power of combined variables in context. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes results from univariate analysis and raw feature distribution. Some variables, like BUN and delta hematocrit, appeared more or less significant when considered individually but lost or gained importance in the full model. This divergence between univariate and multivariable perspectives underscores the value of feature engineering and the interpretability of logistic regression in understanding complex clinical interactions.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003e\u003cspan type=\"ItalicSmallCaps\" class=\"ItalicSmallCaps\" name=\"Emphasis\"\u003eClinical Relevance\u003c/span\u003e\u003c/p\u003e\u003cp\u003eThis study addresses a persistent diagnostic gray area in gastroenterology\u0026mdash;when to pursue a CTA in hemodynamically borderline patients with a GI bleed. While current guidelines clearly support ordering a CTA in a hemodynamically unstable patient with a lower GI bleed, real-world scenarios often involve more nuanced presentations where neither endoscopy nor imaging is clearly indicated. Our model fills this void by providing data-driven, transparent triage support specifically for these \u0026ldquo;gray-zone\u0026rdquo; cases, where the risk-benefit calculus of imaging is most uncertain.\u003c/p\u003e\u003cp\u003eCritically, the model prioritizes sensitivity\u0026mdash;achieving a recall of 0.83\u0026mdash;without sacrificing interpretability or feasibility. This level of performance is particularly impactful in acute GI bleeding, where a missed active extravasation can delay definitive therapy and worsen outcomes (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e).By contrast, existing clinical decision-making often relies on gestalt or incomplete labs, which may either delay needed imaging or lead to overuse. The integration of interpretable features like delta hemoglobin and the maximum INR value allows clinicians to understand the model\u0026rsquo;s predictions in physiologic terms, promoting trust and adoption at the bedside.\u003c/p\u003e\u003cp\u003eMoreover, the model leverages seven labs routinely available within 24 hours of presentation, requiring no additional diagnostics or workflow disruption. This low-barrier implementation makes the tool broadly accessible, including in community or resource-limited settings where subspecialist input or rapid endoscopy may be delayed (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). When embedded in an electronic health record, the model could flag at-risk patients in real-time and support multidisciplinary discussions between hospitalists, gastroenterologists, and radiologists.\u003c/p\u003e\u003cp\u003eFinally, the downstream impact extends beyond diagnostic accuracy. By selectively reducing unnecessary CTA use, the model promotes cost savings, reduces radiation and contrast exposure, and improves imaging access for other patients. Taken together, this work introduces a practical, clinically relevant tool that advances precision diagnostics in acute GI bleeding and exemplifies the role of interpretable machine learning in frontline decision-making.\u003c/p\u003e\u003cp\u003e\u003cspan type=\"ItalicSmallCaps\" class=\"ItalicSmallCaps\" name=\"Emphasis\"\u003eInterpretability and Feature Dynamics\u003c/span\u003e\u003c/p\u003e\u003cp\u003eUnderstanding the divergence between univariate significance and multivariable impact is essential for interpreting model behavior and uncovering physiologic patterns not evident in standalone analyses. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e highlight the nuanced relationship between univariate significance and multivariable impact. For example, delta hemoglobin exhibited moderate predictive value in univariate testing, while delta hematocrit did not. Yet in the multivariable logistic regression model, both features demonstrated high importance\u0026mdash;despite displaying opposite coefficient directions. Delta hemoglobin was positively associated with CTA positivity, while delta hematocrit had a negative coefficient. This divergence may be explained by their physiological behaviors: delta hemoglobin effectively captures acute red blood cell loss, whereas delta hematocrit may be disproportionately influenced by fluid resuscitation or hemoconcentration, which could obscure its utility in detecting active bleeding (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eMax hematocrit emerged as the single most influential feature in the multivariable model, despite showing no statistical significance in isolation. This could reflect hemoconcentration during early hypovolemia or a compensatory rise following resuscitation, suggesting it acts as a marker of dynamic volume shifts. The model\u0026rsquo;s ability to extract value from such a feature emphasizes the importance of evaluating variables within their clinical context and in combination with others.\u003c/p\u003e\u003cp\u003eINR was one of the few variables to demonstrate significance in both univariate and multivariable analyses. Its role as a marker of coagulopathy and impaired clot formation is well-established, and its contribution to the model reinforces known associations with arterial extravasation detectable by CTA (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e). Including INR not only enhances model accuracy but also strengthens face validity from a clinical standpoint.\u003c/p\u003e\u003cp\u003eBUN, although not individually predictive in univariate testing, gained importance in the multivariate model. This aligns with its complex role in GI bleeding: while traditionally associated with upper GI bleeds due to protein metabolism, BUN may also reflect renal perfusion, hypovolemia, or stress (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). Since our model is agnostic to anatomic source, BUN\u0026rsquo;s multivariate contribution likely represents broader physiologic stress rather than bleed location.\u003c/p\u003e\u003cp\u003eInterestingly, platelet count showed minimal influence on model performance, as evidenced by its low importance in multivariable analysis, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e. While thrombocytopenia is classically associated with bleeding risk, platelet levels alone may be insufficient to capture the likelihood of arterial extravasation on CTA (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). This could be due to compensatory mechanisms in hemostasis, or because the presence of coagulopathy (captured by INR) plays a more direct role in determining active bleeding detectable by imaging. Additionally, the distribution and function of platelets may not be accurately reflected by static platelet count alone, which may explain its limited predictive value in this context.\u003c/p\u003e\u003cp\u003eOverall, these patterns underscore the value of multivariable modeling in capturing clinical nuances that single-variable analysis might miss. Logistic regression\u0026rsquo;s interpretability allowed us to not only identify influential predictors but also rationalize their contributions in a physiologically grounded way. This enhances the model\u0026rsquo;s transparency and supports its practical integration into bedside decision-making.\u003c/p\u003e\u003cp\u003e\u003cspan type=\"ItalicSmallCaps\" class=\"ItalicSmallCaps\" name=\"Emphasis\"\u003eModel Comparison and Justification\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eLogistic regression with SMOTE performed best among evaluated classifiers, achieving an F1 score of 0.71, recall of 0.83, and ROC-AUC of 0.71. As shown in\u003c/span\u003e Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, \u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003ethe model outperformed random forest and XGBoost, particularly in recall\u0026mdash;a critical metric for avoiding missed active bleeds. This is especially important in clinical settings where the primary concern is sensitivity to active bleeding rather than overall classification accuracy. In this context, false negatives\u0026mdash;missed opportunities to identify ongoing hemorrhage\u0026mdash;can result in significant morbidity or mortality. By prioritizing recall, the logistic regression model better aligns with the safety-first principles of emergency triage and imaging decision-making.\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eWhile the Random Forest classifier exhibited a slightly higher overall accuracy (0.75 vs. 0.68), this was driven by strong performance in the majority class (negative CTA). Its recall for positive cases was substantially lower (0.61), meaning that nearly 40% of true active bleeds would go undetected. Similarly, XGBoost had both lower recall (0.64) and a diminished F1 score (0.61), likely reflecting its reliance on complex non-linear interactions that may overfit the small dataset.\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eIn contrast, logistic regression provided a balance of interpretability and generalizability. The use of SMOTE upsampling helped correct for class imbalance while preserving clinical insight into the predictors. These characteristics make it particularly suitable for real-world deployment, where simplicity, reliability, and transparency are essential for clinical trust and implementation. In summary, logistic regression with SMOTE was not only the best-performing model in terms of positive bleed detection but also the most clinically usable.\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003cspan type=\"ItalicSmallCaps\" class=\"ItalicSmallCaps\" name=\"Emphasis\"\u003eEconomic and Implementation Value\u003c/span\u003e\u003c/p\u003e\u003cp\u003eBeyond clinical performance, our model also holds substantial implications for healthcare resource utilization and cost containment. \u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eAs depicted in\u003c/span\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003e, \u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003edeploying this model could avoid 20 unnecessary scans per 100 patients\u003c/span\u003e.To estimate potential cost savings from our model, we referenced charge data from a survey of four major Houston hospitals, where combined abdominal and pelvic CT scans were billed up to \u003cspan\u003e$\u003c/span\u003e4,079 per scan (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Based on this figure, deploying our model could yield savings of up to \u003cspan\u003e$\u003c/span\u003e81,580 per 100 patients. \u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eFurthermore, CTA overuse may contribute to incidental findings that trigger additional downstream testing, consultations, and in some cases, invasive procedures\u0026mdash;all of which incur further cost and patient burden.\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eImportantly, there are also indirect economic and safety benefits. Reducing radiation exposure aligns with ALARA (As Low As Reasonably Achievable) principles in imaging and can mitigate long-term risks associated with cumulative dose, especially in older adults or those with frequent healthcare utilization\u003c/span\u003e (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). \u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eAdditionally, reducing inappropriate CTA use may streamline ED workflows by lowering demand on radiology staffing and scanner availability, potentially improving throughput for other urgent imaging needs.\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eTaken together, these factors strengthen the institutional and system-level rationale for adopting our model in CTA decision-making for GI bleeds. By targeting its application to intermediate-risk patients, hospitals may enhance both cost-efficiency and care quality in a high-volume, high-cost diagnostic category.\u003c/span\u003e\u003c/p\u003e\u003cp\u003eClinical integration of this model could occur at the point of CTA ordering, particularly for patients in the emergency department or ICU where clinical ambiguity often delays imaging decisions. Given that the model requires only structured, routinely collected labs, it could be deployed as a real-time risk score within EHR systems such as Epic or Cerner. With appropriate clinical thresholds and oversight, such integration could support shared decision-making or prompt specialist consultation in uncertain cases.\u003c/p\u003e\u003cp\u003e\u003cspan type=\"ItalicSmallCaps\" class=\"ItalicSmallCaps\" name=\"Emphasis\"\u003eLimitations and Future Directions\u003c/span\u003e\u003c/p\u003e\u003cp\u003eDespite a large initial dataset drawn from the publicly available MIMIC-IV database, our final cohort included only 140 CTA cases; 32 were positive for active gastrointestinal bleeding and 108 were negative. This modest sample size, derived from a single-center academic medical center, limits the generalizability of our findings and warrants external validation in diverse clinical settings.\u003c/p\u003e\u003cp\u003eFurthermore, the dataset lacked critical clinical information, including vital signs, presenting symptoms, transfusion history, medications, and endoscopic findings, which may have strengthened the predictive performance of our models. We also could not confirm whether patients with positive CTA findings subsequently underwent interventional embolization, preventing outcome-level validation of CTA utility and introducing potential misclassification bias regarding clinical severity.\u003c/p\u003e\u003cp\u003eWe also observed some unexpected statistical patterns. Notably, the features delta hematocrit and min/max hematocrit demonstrated negative coefficients in multivariate modeling, while delta hemoglobin had a positive coefficient. Interestingly, delta hemoglobin showed moderate significance in univariable testing, while delta hematocrit did not. These discrepancies highlight potential collinearity, physiologic confounding, or modeling artifacts and underscore the complex interplay between hematologic markers in acutely bleeding patients. This divergence may reflect multicollinearity between hemoglobin and hematocrit, with delta hematocrit potentially capturing fluid shifts rather than pure red cell loss once delta hemoglobin is accounted for. Future analyses should further investigate these relationships with expanded clinical covariates and time-sequenced data.\u003c/p\u003e\u003cp\u003eMoving forward, additional studies could explore ensemble models or deep learning approaches, while maintaining clinical interpretability, to optimize performance in small, high-stakes datasets. Prospective validation and integration with broader EHR-based features would also help translate these models into real-world decision support tools.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eThis study presents a transparent, interpretable machine learning model for predicting positive CTA in GI bleeds using seven routine labs. The model shows high recall, clinical relevance, and economic value, especially in \u0026ldquo;gray-zone\u0026rdquo; GI bleed patients. With further validation and EHR integration, it offers a practical decision-support tool to improve triage, reduce overuse, and accelerate care.\u003c/span\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements:\u003c/strong\u003e\u0026nbsp;We would like to thank our advisor, Dr. Aman Mohapatra (Beth Israel Deaconess Medical Center), for his invaluable mentorship and clinical insights throughout the development of this project. The authors would like to thank Beth Israel Deaconess Medical Center for their contributions to the MIMIC-IV database, which made this research possible. We also acknowledge the support of our clinical and data science colleagues who provided valuable feedback during model development and manuscript preparation.This work represents a collaborative effort between students from Harvard University and Harvard T.H. Chan School of Public Health , and we are grateful for the support and resources provided by both institutions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Approval:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study used publicly available, de-identified data from the MIMIC-IV database and was exempt from IRB review.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions:\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePG, SC, NB, RL, AC, KY, and AC contributed to data collection and model development. AM supervised the project, performed clinical validation, and led manuscript preparation. All authors reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Trial Number:\u003c/strong\u003e Not applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of Interest and Source of Funding:\u003c/strong\u003e None declared. This study used publicly available, de-identified data from the MIMIC-IV database and was exempt from IRB review.\u003c/p\u003e\n\u003cp\u003eEthics, Consent to Participate, and Consent to Publish declarations: Not applicable. This study used publicly available, de-identified data from the MIMIC-IV database, which does not require institutional review board (IRB) approval or informed consent.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eKim BSM, Li BT, Engel A, Samra JS, Clarke S, Norton ID, et al. Diagnosis of gastrointestinal bleeding: A practical guide for clinicians. World J Gastrointest Pathophysiol. 2014 Nov 15;5(4):467\u0026ndash;78.\u003c/li\u003e\n\u003cli\u003eSengupta N, Kastenberg DM, Bruining DH, Latorre M, Leighton JA, Brook OR, et al. The Role of Imaging for GI Bleeding: ACG and SAR Consensus Recommendations. Radiology. 2024 Mar;310(3):e232298.\u003c/li\u003e\n\u003cli\u003eAlkhorayef M, Babikir E, Alrushoud A, Al-Mohammed H, Sulieman A. Patient radiation biological risk in computed tomography angiography procedure. Saudi J Biol Sci. 2017 Feb;24(2):235\u0026ndash;40.\u003c/li\u003e\n\u003cli\u003eRamaswamy RS, Choi HW, Mouser HC, Narsinh KH, McCammack KC, Treesit T, et al. Role of interventional radiology in the management of acute gastrointestinal bleeding. World J Radiol. 2014 Apr 28;6(4):82\u0026ndash;92.\u003c/li\u003e\n\u003cli\u003evan Doorn WPTM, Helmich F, van Dam PMEL, Jacobs LHJ, Stassen PM, Bekers O, et al. Explainable Machine Learning Models for Rapid Risk Stratification in the Emergency Department: A Multicenter Study. J Appl Lab Med. 2024 Mar 1;9(2):212\u0026ndash;22.\u003c/li\u003e\n\u003cli\u003eJohnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023 Jan 3;10(1):1.\u003c/li\u003e\n\u003cli\u003eMariani G, Pauwels EKJ, AlSharif A, Marchi S, Boni G, Barreca M, et al. Radionuclide evaluation of the lower gastrointestinal tract. J Nucl Med Off Publ Soc Nucl Med. 2008 May;49(5):776\u0026ndash;87.\u003c/li\u003e\n\u003cli\u003eBelete MW, Kebede MA, Bedane MR, Berhe TT, Tekle AB, Shash EP, et al. Factors associated with severity and length of hospital stay in patients with acute upper gastrointestinal bleeding: insights from two Ethiopian hospitals. Int J Emerg Med. 2024 Dec 5;17:185.\u003c/li\u003e\n\u003cli\u003eDriessen GK, Heidtmann H, Schmid-Sch\u0026ouml;nbein H. Effect of hemodilution and hemoconcentration on red cell flow velocity in the capillaries of the rat mesentery. Pflugers Arch. 1979 May 15;380(1):1\u0026ndash;6.\u003c/li\u003e\n\u003cli\u003eShikdar S, Vashisht R, Zubair M, Bhattacharya PT. International Normalized Ratio: Assessment, Monitoring, and Clinical Implications. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2025 [cited 2025 Jul 3]. Available from: http://www.ncbi.nlm.nih.gov/books/NBK507707/\u003c/li\u003e\n\u003cli\u003eTomizawa M, Shinozaki F, Hasegawa R, Shirai Y, Motoyoshi Y, Sugiyama T, et al. Patient characteristics with high or low blood urea nitrogen in upper gastrointestinal bleeding. World J Gastroenterol WJG. 2015 Jun 28;21(24):7500\u0026ndash;5.\u003c/li\u003e\n\u003cli\u003eLo PH, Huang YF, Chang CC, Yeh CC, Chang CY, Cherng YG, et al. Risk and mortality of gastrointestinal hemorrhage in patients with thrombocytopenia: Two nationwide retrospective cohort studies. Eur J Intern Med. 2016 Jan;27:86\u0026ndash;90.\u003c/li\u003e\n\u003cli\u003eFred HL. Drawbacks and Limitations of Computed Tomography. Tex Heart Inst J. 2004;31(4):345\u0026ndash;8.\u003c/li\u003e\n\u003cli\u003eHendee WR, Edwards FM. ALARA and an integrated approach to radiation protection. Semin Nucl Med. 1986 Apr;16(2):142\u0026ndash;50.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1\u0026nbsp;\u003c/strong\u003eFeature Engineering and Model-Level Impact.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eThis table summarizes individual feature behavior in relation to CTA-positive GI bleeding cases, including each variable\u0026rsquo;s source, physiologic rationale, statistical significance, and impact in the final multivariable model. Notably, delta hemoglobin demonstrated a moderate predictive trend in isolation, while delta hematocrit was not significant univariately but later emerged as a high-impact contributor in the final model. This illustrates how certain features may lack standalone power yet contribute meaningfully in multivariable contexts. The discrepancy between raw statistical significance and feature importance highlights the potential for complex collinearity and reinforces the value of structured feature engineering when developing interpretable predictive models\u003c/p\u003e\n\u003cp\u003e\u003cimg width=\"576\" height=\"185\" src=\"data:image/png;base64,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\" alt=\"image\"\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u0026nbsp;\u003c/strong\u003ePerformance metrics across classification and anomaly detection models for predicting CTA-positive GI bleeding. This table presents accuracy, precision, recall, F1 score, and ROC-AUC for both traditional classification models (top panel) and unsupervised anomaly detection approaches (bottom panel). Among classifiers, logistic regression with SMOTE upsampling yielded the best overall balance across metrics, achieving the highest precision (0.82) and F1 score (0.71). While random forest had higher recall (0.75), it showed lower precision, suggesting over-identification of positives. In contrast, anomaly detection models demonstrated near-perfect precision and recall\u0026mdash;but with low ROC-AUC, indicating poor discrimination likely due to class imbalance and lack of labeled signal. These findings emphasize the challenge of low-prevalence detection in small datasets and highlight the potential utility of structured upsampling in improving model robustness\u003c/p\u003e\n\u003cp\u003e\u003cimg width=\"576\" height=\"272\" src=\"data:image/png;base64,R0lGODlhQAIQAXcAMSH+GlNvZnR3YXJlOiBNaWNyb3NvZnQgT2ZmaWNlACH5BAEAAAAALAAAAABAAhABhwAAAAAAAAAAMwAAZgAAmQAAzAAA/wAzAAAzMwAzZgAzmQAzzAAz/wBmAABmMwBmZgBmmQBmzABm/wCZAACZMwCZZgCZmQCZzACZ/wDMAADMMwDMZgDMmQDMzADM/wD/AAD/MwD/ZgD/mQD/zAD//zMAADMAMzMAZjMAmTMAzDMA/zMzADMzMzMzZjMzmTMzzDMz/zNmADNmMzNmZjNmmTNmzDNm/zOZADOZMzOZZjOZmTOZzDOZ/zPMADPMMzPMZjPMmTPMzDPM/zP/ADP/MzP/ZjP/mTP/zDP//2YAAGYAM2YAZmYAmWYAzGYA/2YzAGYzM2YzZmYzmWYzzGYz/2ZmAGZmM2ZmZmZmmWZmzGZm/2aZAGaZM2aZZmaZmWaZzGaZ/2bMAGbMM2bMZmbMmWbMzGbM/2b/AGb/M2b/Zmb/mWb/zGb//5kAAJkAM5kAZpkAmZkAzJkA/5kzAJkzM5kzZpkzmZkzzJkz/5lmAJlmM5lmZplmmZlmzJlm/5mZAJmZM5mZZpmZmZmZzJmZ/5nMAJnMM5nMZpnMmZnMzJnM/5n/AJn/M5n/Zpn/mZn/zJn//8wAAMwAM8wAZswAmcwAzMwA/8wzAMwzM8wzZswzmcwzzMwz/8xmAMxmM8xmZsxmmcxmzMxm/8yZAMyZM8yZZsyZmcyZzMyZ/8zMAMzMM8zMZszMmczMzMzM/8z/AMz/M8z/Zsz/mcz/zMz///8AAP8AM/8AZv8Amf8AzP8A//8zAP8zM/8zZv8zmf8zzP8z//9mAP9mM/9mZv9mmf9mzP9m//+ZAP+ZM/+ZZv+Zmf+ZzP+Z///MAP/MM//MZv/Mmf/MzP/M////AP//M///Zv//mf//zP///wECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwECAwj/ALEJHEiwoMGDCBMqXMiwocOHECNKnEixosWLGDNq3Mixo8ePIEOKHEmypMmTKFOqXMmypcuXMGPKnEmzps2bOHPq3MmzJ8JWN4EaFFqTKEWjA5HKVIqNqc+nUKNK7dhKkNWrTglWFSRRUNaDW4mGfSjoysOqV74WbHWlbau3ViGyFVq2rNqCcc9eJbj3Z16tf6cKHky48MErAVhUZcGVYeK7TRsHaOywFYsAZpteVnyQMkEWnBlavsIis2jMb0tbZgFRUIC3iBGXbugVG+KIpF8LFHS5NkLQnjWzNky8uHGerTBHboz0LdDVTZu+TTqdreKqQqcLdK59e+7nm6kL/yT9XHpk8dKNuk6bfnv5pI8HmgXtXmz26rWBY3ePHhtvrn919xxSvClnW2/8UYfYfQKBNiBkx0Uo4YQbucaaWW89xpZtpHHm4IIOYsOYahYyNtlipGmW2IlJlaVcWomNR5pVl11RIGK8eYgjb+QNdFkA/pVmo4igzdaggUnRV9aMHDLGlluuLTYZj/6RlhaPr93oonCzdXhZQTBe15ZulvEGF2YB6NaWk/St6RuFcMYpZ0XJgWbaZXCJaNVr0C24YJl/KpdcWqrN9hh9BIVpVZ1BlqVYfDWm5adiQk5mWlOTHUgkW5ixd2RwRB446GpXTelfpiGiVqh1VC5o22uTVv81ZaYDeRUboagSyppqC/5H3qGKgTrnsMQWu9tjpq1XG42ZIgsrY1XiOahAUxpK6a6hRcZpWqPiaCN0nArV221A3UjUetigNmlwdhqU4mL0/SjllXx+OVpytcEW26tAgRhslifSOt6pJwaK7WgHulgXdKDpZuzDEMc5bblHZtaWa161ydifpUXJ6L0n0gesVmaha+GrlIUYIr/A1YonfFy1CavDtSIZnZImssYpawXKGjNnqg6HssZc+Tnznl4J7B9QaQLV27S/mmV0ymLGGPHVWBNnIVzDuTZQp6g9hrF+mNVYZ9KSPprxaytXyTRXdWbI2Npr+zirWzb2eKzUG+//qnRTOMKVGrbfEWojjUhbVi/htqL5b+Os3mYlvkSxxxvgBVMK+dOJLdrmvVmHLvpTaF085Hm7SWort6Z7tajlV7YF+JBsWeVWrew5t2S5t9f+5HN2VenVk54tzCGGdXXWVu5QXum6W8QLT/zC0dcVl+8XVwm77ZRttZ32cBtPPOysx+7p6Oinr/767Lfv/vvwxy///PTXb//9+Oev//789+///wAM4PoEJzjX6Q47w0Mgdgq4QAUm0IAPVCADI2jACR6wgg68YAMhqEEKblCCGbSgBTnowRJ28IQiDKEKSbhCEHIwhS9sIQZjSMMM1nCGLswhDnf4QR6aEIY67KEQ/38owyGisIhEvKERgWjAkYzQhlBkoRSjSMUpWrGKWLyiFrPIxS16sYtg/KIYw0jGMZqxjGg8oxrTyMY1urGNVSQJVt5IRzjW8Y52tOGi8MjHKu4xj30U4x8DCchCEnKOhkzkISFkEUU6cpGQfKQkI0nJSVrykIO8ZBcRUclOkpGRFXEdJ6uSCFG6rpRVGaUgUCkIVbLSlaYkZSxXOctXnrKWuEzlLXUpS17S0pe27GUrdzlMYcISmKa0SjB/WUxmHrOZy4xmLqGZTGcS85nYvKY2jblNawpzj9JEpjipOc5scnMr0zSnN9cZTnK6U53thGc6uxlPejZRJE/0pD41uf/PfvLzn/4MKEAHKlA6ylGVBE1oQduYyYUK0qGAbKhCq8jKiUpRohD1pBMtytGMdvSjHg1pREVK0omCkiK3tJLrcnS4b+Zoj1fpGDpdSlNmgnOlxLzpTG2a057WlJU6nSNQfcpTWV7llBcLqlKJOtSf4tSpR2VpUp26FSsB1apPLapWm6pVmBZJpUt1KXCAE9atZrWpHVre3GYk1LO6dadcjavt7CRT2y3vcDEVUl2vKsu0Wid2pRHSUanaVrhm9aQSYWDDPNebi+rHgLIBqWQDeZmKlpSE43KQFzEGwR0RNDYu+tZmO4cYyybSbElN05UUaqE9sYlEjbWQ+TL4WOv/TMkyaMGMrZxESTnuMTcs4OSPntfSJcUWr52brHLNOEjXoGalxb1YF7zAPe7h9WJVPRxemdlFmKbJtda1Sheuq93y5k20Sa0Ldu/4x9jsrKq/Ja/rOvc0SNYoqjK4rT4HiYBq+fd19H1uVbqAWd3mlreHS66tMFpIkigwZ2qdkpWqxYIudG49ARiv2JbL4TOOCcGlsdRX19OhRBTJTmhKcGyKVCY69ua+iaFrYE38VcR4wUobEzGLEZPiQ94KtOuZW5EmE1jVkFa/I8XOZqCkuNXqU5WyBbKCLeSFxkaxQ7flsZIVF6x/IrYrurRRpAq0Jxt1TrcvLnJrz8nOebK5/55vdnOby2lPOivzec4tVd5eOjfdolm3jpJNay9GWjnb8s7JSUxSYws20sjAV2Dr2Iot5Wczz43OcYarbFRLoyJ7S8xjEvOIJgPnOTdzjzWqa2NLzWo3W3hFThL0ML/7XVVy0nr/uS94qzIDF/FWnpg29Zfl8mBO/Teyig5abi7mhQ1ftsNdxLGzgpWIKXfKQWajkaAr+7xO2dHKuxa0pbvtqCPLdj1VNpFoDTlpIVvKXKsz25ERSsj7rjbbrFU0o3nbWnAfWDF3Te53MmZsel/yoHhubZ6VdWYxE1m31FUwtJ990ZXObeEXJrOelbUjRKTaRICWFZIfuhWJa7zZF/8/LqBfDLbQotvZDIXsn3F13oyjeNz1BaRldU3ufg7yu7K+DHWbnAjZcBIR261ulae0ZuupFjvFleRGVyphgDd2wszisaUp3WWKT5yKj7V3jMtt5M2MiESb7tjYM8ZbNubGu7w1+2oUveIOhTbdHwKOyQEpZbWT1ewWShOEPctHmCIANEC1k0Vloyt3O6rLReIRBP+ztmD92HmKlrwlh90aoyZ9z9UtHa6vC1OmvrWwqC+rXA3L+tSbvvWqj+ror0DdJR31CgTmXu5xfVQvlJL3UzWrYaN6uN+H3vZ27T3yozreq5D+9a7fqXqdT33iEp/6BM49YWMv/e3CN/rCB7//XIue3s+LdlE39r7sj7+8U7f/9NwHP+cr83Wv27+Qh+96/fd///7z34a+5X8C+EgMBklslUYF+H+INIAklYAMSEYO9oASqIAUOIH7RXEGZ4GfJEfB1modaGgg+IEi+E51RoKZNkyIZoKmBmwqWE8pyIIwWIIxSFgzuIIy2E17VIMe2IIhyIMjqIM9+Ezz1xD5VIFGqIFIeIRKmISHxIFMuITdZYEOyH9T+IRWyGEbNYQC1BqFISyE4YU54RxhqIWGQYYLAYZbSBVm2BJoOBXd4RNtCBNrKBhv+BFimIZy1IUREoc2UYc04YcQw4cXIYh42EhzuBKE6BOAqBOHqIah/7OIHAGJhagRiXgTlcgTl/gSdxgUjSgVnYgXk4hPn3gSmciIowgSpXgSp9gTkpgRYkhA0iEsrUgdrhMRbQiGDDSKs4iKeOEUw1MrMREcb1IrC7Q0qSgSwgiIm7iJqmgfXlgVsQiNomgfzYFAsWgcx+gQ3VMjqaMUrkKECEIWNsMfOJIQoNU2HfGNKPGGGwMqTpIkqygantGOfHFiQYKOS3EutkMg54MwKuEZ2mUQZXEsLCaKSSEpRiEbaIcohbGLhkgUiwUbSVEr+TKR72E3wXGR5QEb2eEjw2MaSoExp5I70fE9YlGSKHmS5fImKhkSlCE7yQEmnsEpwegyTHMuz/+BIU5zKTLRPUOSLUEiFlSSEndoGYAjlEIDNyKyUbjjNsAokXBzPoQRjwCZZ9D4JGLCN1IDILcjH39je7PDVt8iZjuzO9ECkJlyMtrTL3f1JORhlsujL6sTlly5Jo3YHTAJLbvhJHRxOTDxkpmhl4DBFzwZE9phlBySJN+yHUuCiPIRmNsYGP6RjUTokZiiFNkyCIXphqu4iVJDWkSSI6LSLbIyLV45k1nyL6KCMQeIMaUJaS3iX1yjJ5RXL2WTHKxyaV4zH/9iHbIyJpTZGZaZLpUzItwSMzXZII3xjomSkPHYFfLBGbKzHUTmIRzynGshFCmSmONRKQcZnArRHdD/UiZrERqjojV5KB9HAi11oRweMy4iYimXoo78gi+9AjBJMxsiGSUek5BlQyvAQWiluRrlImJNJivQEiucFTbgqRWVAzeCOZ4Ahzh/+ZjKuRaXIpn5uJcc2qFCkjdAaRKVYzGAqRnMITQhIZ7PQTOp0yKCORgOaYjqWTFB0iGiciA68ycFEY5VkpqGw1sngnEA8iygwlkvo6NuiSGI4jHS0TkH8iw0GTjVspkY8ZKSkR2zASghJpWOKSPUwo8UiZ22eJD9so17OWEvWhJ3eDm1Y6F6shsNmhB4CSAhyqUa6obpOR7ZgSqToS4OYyTuhZaNcTh9lhq/qCxIsx6cAl6C//oq/3IleoKlnJEjD4ZqNMmfAGNmcZogmuE20zGQG7IdKMqG8GgmKGme61YUlcOR/qEZZeqmKeGTqfE9bBoaXAoS51IyOukd4nGrntiZ1VEm2VE0bZk9kxmdnqIv50Ws4xljhGpX7NEx/4EiPvmOy3MgVyJ3jhKd17E7bikiQxKXi7EkVDqIhLkstZE3CZmcqbMszMOYgeWrLjGTh3OUk7mYcCqm/bGPEokWf1GukTinyJMZ0giu8ioVmyqcf1kpALsRo2GjUBGjQZGwmqivGUGxDmuxfQisGiuQsYOxeDGWIOuSengcI2uHEkuUHVsUeeoS3CGm3PGrJWscKVuhjP/4iBwbiiT7hXsoszhRsxJysq2qsyg7s8UBtC5rijgbgVB4hU7btFD7tFdIlS9btVZ7tVibtVq7tVzbtV77tWAbtgcktmRbtlmbQGabtmr7smi7tm6rtW37tm8bt3Jbt3YrOFlItB8htCPBtyjhtwE7his7E0ILuAKEtEthuGo6uCiltFmDuPTHuPujuB5BuX0bFcyoqqJDlXrrEZDLrkcrudB5s4+bs53LEZaLuj2LuaIrp637EoV7uoHLs8fxuSrxuj+Buyxhu6JhFAU7HnbBPf2CFr57sEPrHdYzFEjRUv4aPqoDN3OxEnOBFhdLMsMoHwAJur6DFwPJmLoLjIn/cjrcSxl3mreMeavU668NixFv2BcUmTeRwS3Ewbnni6+Y4iGZMhdRch7+iBDiSypD6SPzCRclszKi1TL7cbsvEqKJJRaWF7IJOaossY3E25Shii+ciL3RmygzaSSxSqah+p2pY5dOhDve4x3n5TbGq4imOxotc5ojOTDcgRqjwhTn2SDs6cGieaaoYyWoE1oUoxXRsZHB+ot5YpIxWx6UQcNKuYm/yBCAmSFGAZTbarPXSZxJ4pS2kbpnCI9bHL5IucJF26ltKqogKR3rexF3KDum+j1VIiIEjJ6Xq5jCGCNalq5cl588+piC2SMEupcR6iAQeWnnSyJ8UWlNJsgj/1KjCRpj/1l5qaIrO6MyayLB2Ym9Jhqdi3KUDEyUJfqmZnw+nSyHdDGpQjMaeWFXXcqd4suaCXmyKnqZBnEdplyGprvFGkpaKtUvAQM3aDaqrsOQCPyNaCOpLDKS5AsbpgIz6CLIInbAo7KgZuGnKJIukgIkoDWlGNzFF5ouJyo1q7WUVqyXL/odyGuJNnnFl8kjGzKd66gd7+LOUOIwG5yiWEqnwsgtZGLJntiyCiFoiWYxfCIoRsaTeZWTnPHHB12SHSKp9iEcGRo2n6kZMGI1BXIhvYwwgTLN84GfrtnLCzGn3fymZRJYqAG7biqYc7GdoKy5QeKlHbomlUy5n//cx0rKHGnMvpaJm9ybyX45lbesEKIZLuOR0YOyw/JRcsuZw8izWzzTGP/hI9niGwGMKWbGmxkyPGT1GjGCqRwdLujip6+iqeOaxlEck9qJIc7jmigtIzdJF2qtnVws1GQKyrN6naBGuYeZo0o5nQ4jxuw7olo8o3BMvYYRuw6hlywtpX121RwsqmIiZCQK2RnDlUky1eR7KauxL95ieWZidkfz0X4zJnITn4tajtKqFgIyoW+svyA5yn8LjzAZ1yGcphsKvNHLOgQHq+vIwaCqpBVsG/EoIKrzxYBjwoC9E7wb0hDSHVYKIB95VC0CvXOEFcRYLhsUIOjkvecyvnH/YSfXEzz6+zwzNZaW4hzXhU4LsxnxQdcgrC2TKYygK9zAs5LkW75tHabwbdjEeLsC0r22ocHgy5QUyRzkqyDfSyf+HDq3gRH0KRprBSThmeBjWrsU7t5huLTTiD5no8ZPcxY4dhdzXb0m67OceOGkuOCP6IoK3rtTOeKei+IIAeMTruH2LOPvQ+MWoeMVwroovtxTqeKyexQ4HtLX25BFDopjaOO4GrVOLrVPHuVQzoByNAuzcA1XnuVYvuVa3uVc/uVeHuZgPuZiXuZkfuZmnuZovuZq3uZfzgqI4OZyzuZ0PudVMed4Xud6Ludvsed+nuduLgh/PuiAjuWtwAqE/57ohb7odA7kcoros4ANWS7pWE7pkT7pmF7pmX7pmt7pnP7plh7qmy7qnk7qoD7qqF7qqX7qqt7qrH7lgv7qpj7rq07rrm7rsH7rtb7rut7rsp7pse7ruD7svC7rYljsxC7stc4KrfDryv7szh7tyf7rs0CVW34N157tV47t267t2O7t3P7t3T7u4l7u4X7u4J7u5I7u667u5u7u7P7u7T7v13Do9B7v+A7v+n7v3j4LgsDvAC/vAp/vAU/wzL7vA4/wBJ/w3S7oC//wCq/ws4AIDF/xEF/wEY/x7c7sedsKYN4UH+/xXQ7yIy/yXE7ykp7yIp/ykb7yl47yL+/yKv9f6iY/6TVf6Tff8iG/8yXP8yef803h7z/v8zZP9Dhv9Drf8z0f7Ek/9Erv9FBf9E8v9VFf6Ux/9FOP9VXf9FTf9Vjf51yv9V4f9mRP8lXP7ECf9khv9mPP9mL/9mWv9l5euNuu8Rav7S6R8Xdv9xef8Pau930f+BF/54Bf+Hyv8H0u+Ie/+OUe7Iq/95DP7Ydu+JFP+Y+/8I4OFpMv7ciO6i3B+dA+7aK/6tcQ66Mf+p3v66WP+qwP+q5P6Ux/+q+f+qB+7K0v+8PO7LN/+7SP+7te7VU+7nX/7c2e5VaO5cfv8eGu5ec+kTo9FAhR+Ywv5pfv7W8h/Myf7cpf78P/X/3T/+/ifvzknv3N7v3Sz/BgT/wC3+XHb/6CD/5jvvzcT/2WD+4TH/7Zb+XGz+3if/7uX/8A0aoVNoIFDR5EmDChwIGzrs1CdOUhNodXrrBoJeiioFmzrAjq8vHhtVaIWAhqNTHlNWxXNKJsZfFKQY0YB2p0eWVgS5wuMcrEuRNbylkUVU50WJSoUaVIjxZNio1VFxazAn0c6rQp1KdMs0KdJYgkSUFWriD6OpTVWVatPspgtdXr0qh0u9bVmtahWKJlXY51+NFlK49WqsrFyxXxXcaL5YpVGjNQl6fXXCKyHOhKIFaN9Xq267hVZ4eBrradWPMioo5Xujj+DDt0/9a2STNLtG2RxcWOrFiwnpt39nDhxRXHRk40rELmzZlzvHZtaheLnR+2CoBxFosAlGddmRogLiLKiALgjj7x5FCM2Li3bxkgwM3zBK/MzxgfpU5B9VsKNSo6hwRMb8CRCETwQAUN5A6R/sRasEAJEzTQwNEE7M+3qkYKKwDWrmGhqvsGrHDCCCk0scQEM0rKvLIwGukiVvr7jrsArEgxRxR3PLHHAQVKj4UufMNMwFZ2Ky0AlJRU0UcdnSQQMpJOCuSwhwKRTz5BTMruySa/TPHC0jDabaQjs2SBlSoDCMRLN3kEE843zQTQOTuZE3M7w/5S6j6O+rNCIIlMY4WV8/+W3AwR0tizSaCW/GQ0APYkJag/FoZqaTdM+5sJpoIK5Yiz0QJppQtFsXEwkI5aCUQQsUCdKixXr9nS1TZjEs9V1jhrlRWpEOG1I6+iInYkKWXw8L64KDJUENKs8NCwRYs9jtph87LWNsiwgVYQLJfFFaQ0RzNUJ2zPrRbda9PlSkqTrpiFTdtCxMyq80wKZF19s1W3Lkc7ZHUFQYwidcYASMIu0H4XZlcqwiiyYgUQ06TIqlnWwpFVJhnel2N+G/644jrvJNkgKa+pskwDLf1OPjXj8rOL7BDhElpVo4vvvJ1cyi4mQbgTKLuCsJvPPiVpOnRgg46EK8RrZN5N593/XGtlhZTbCvGi3Q7AyNKL4IWWPCV9O+/GWWT++bc5HRotKWhZsTnI7FhoU+bsToYy7zgjDCspZe+jLLr+sgtcugPa3Dvxtb1ErUNnu9zrPRkufm9DvRfnEcJrkKWZzQQtik7Dy+XcEaL0smvZOhKVjO4+xDEfPXbFE2yrZNuXdjSlLkBS9atZDljBtftO6sx1Qws+r0oZDnTvvYFm4m4w7rCxdGigjZ6pUvksOuhIFiaGiE3fVjCU7hmnFk/0ENeaj9OJ2ceVsgCaDmC7NCOehSgOv9qfMCnpvc+HzuS9tqyAOzdLTP84VJkENvAo/1MS2vQXovOQ5mcL3IoCbZNB/wdykIGpiVF34mUFtvVHM+e5hmFmhkH+dbCFH9wKkFKHJdzQhk/xmpwGdehC/TmkLSk5ksRCJKXIkQZLHHmhB5WYRCb2EIZNxCCtbjdF6pFogQRq2W8M1Z2HBNB95gHPebRFvfvMpFM9GyB2LkUQNe7saNrjD6YIcg3UsUBiNJyFAdl3uLNdxEN5lMGZtnOAznEEaNwxmCGrEiJcmYSEsWvbQ6DlFtZ1pGcy881HEEk62c2JRV0Uj+tidBJEUg41nJxdJxEEpGuYhzMHWJ5DZDakG1ElEMi6GSphVyIpeXFjAvLNSESnS2JehzTgm2SFQFcRzxUzlc/cJcKoaDuBTP/rXCb5HvviksKxdSdoPvlenwZyJDMyalJG2wmnClJGpM3kedb7HiMfpCGrcCoQwVOW+tJ0H41RZn38pIr8wMOdVrJABs7ymLFGEjZ+9gpaliFkmWRwkoSCrKK0moifvkUop6ENRKYBT0U7ZlGLSsk38BqbqcqFzREh4gAzCVlMOcYQZs2nSohwS0ZfUzGdiVSmRakNRa5wgPdpZCTaocgRfzpSpi7Vqfmbpu04kjiP3OyWfhNfiFo1xButBET8scl6XEW9IQ5lIz/Lnll1pr2ToLUgKdxNiOxFN2/KoEq7m9tuWGFALpFtS/Orkk4OaLz1/fMAXQgEThcnkAH5xpb/QWQFvq5Awisc1CyqxGyJ+hYdcu6mFchay0mo8p0DyCec0XSml2RoGdcEEowxEVKafrZV1aE2dtCRpEUoW8m/ROee2rFtnEwnuJNYgVTrkw68BAQteEEztc/Nm6OieieaJmYuFzMKWlKiKMwoKi4o4S5gMpKfm2BqnOQdynjr5Ck2qle6Q7FKRgoFRmd1xlm3imzfQBWWQrEIUDOSDovCW9+MWMa0fwQNUsICGKts01nbBXBKnLUt4hynwsGxsFP48pAGQ1jCEZYvSmRjHAyX2LrWpTCtCiWdFRdKVQ5RFEeWcuETJ1gxYpJKhBe8l7EIpL6AMbGNg1ObkYR3NHwx/50Px+urGo84w08OspMxvJzp3gm3zpVdlbVcIhoG14Sp/KxOWGFXSNYWulhGc3o+meYzB3dCrMysm3WJNzbHmUJikrOd9QzN92oZT9YkaVOxhQ2WFKXQhKYIog2d6EM3mtGPXjRUhjVbQAu6Ipf9KcomQx5hZXpbTvVpqNmF0UCDutSivhYRLb1qU+uruqhudciCGmtY1/rU1YKqn5tD5zzX2de9BnabI9lmYndys8X+9ZkbF+w9m0hzyW72Q4oUbWpDu0m107VCaOrEJXIbihv89g6f6G1ys7Dc4D63uLttbnPveNzsRje81R1uvKg63TyUN77jve/9EfHd/Nb3vP/PLUN6/1vg+Q7wvQ0ecIYv3OF7yfZzCJM/ik+84hRf1cUzThiLcxzjF++4xkHekZFv/OMeD7nHTZ7yk4sc5SR/+cph7nKRz0jlJcd5zHNO85vr3OciZnnQZ95znrec6Ec3OsbbhnShy9zpO2f6zGU89KQ3HepVp7pAWOPzolud61F/+texLvM+R3xo40VJ2tG+drW3ne1vd3vc4T53uded7ne3e97xvne9953vf/d74AE/eMEXnvCHN3zi2e4qxDe+8IxXfOQdP3m5j8zs1Lt85jW/ec6fvfOfR8nnRe/50etaaaW/XdlRv3o8WT7iNGV97GW/+dPPvsq1t322cZ//e5Kpnvek//3sd3/54Qff+McHPvJLVnzll8z3zR8a9Nsr/dE///Wup372R8987Veq+1L9/tmxz3vrh1/L4ze9+dXP+fJ3n/vrT6/6309++F+/87Cvf/61PP/s89/8/j++9rM9/NO/6QLAqDo9hhDAhVhAbVPAAvy8svM9/CPA2MsP3MGT6Gu+PpvA96rA7ese19u2DzQ7CWw9DYTABowq2FOjk1gr5+sZKvIj1agy9IPABAQKk8mOGJxB3sNBi8C9jNCJSlkP5as9n0krggiRrFGrIUQ9/MuJ3XsJsqIgG5yuPjsJJzSIwRiK/oBAFKS9LTwaSLGdEJmmNurCFSzC/y9kQJq4lCMRCpiICbNSQc47vZ9hDwDhHoNYwwAUivZwCfGDQ+phr9VLwPaAD/twQjnUQs0TCu7Bw6EJRDe0Qu2rQyrCv7Jip4zwrEepCTaKGkC8CB8jpZxIxKEQGk/RDVJEK04cwrhyjxhkw4Q4vSIsmoOACU45wDBUxIG4xTzUw10EQSJ8niQERV/UFAtMpzecRLIaGACZQ/Z7xIH5mUdsj53AjlkUximqvTJKm+fBiPrgFD/BDsG6lGocRUtRR/9go+wYRWdMR3fkGXJ6EJ0gx1lECOk6EqXpQ/sYp5qoRDtcp+xZw43YQ/cIyCdcRiJECI1go6+xwIG0jzvMQv8tPMXMe69rvEXv+RmlGcQUTMgz/MPz4A6lcRWtyRRXaZ/66Jn7SMkHKUdsYEdGORKj0Y9yHIis2Qh3TEl8pMWCsMWRWSOJzL1a5EePhMRrbMbju8MhPMh1wsWLjEDpugizSiuLCBpsbESMxEaxgg/sIMV2nMWQpKJuVJLqYQ926o+MOMv6MMN4VElEaZ6lQaOcXEmVzB5+ekj5YMttDL/34p6PXCfcc8iiJMo1TEb46Mfg08c1ekqzQoitFL3T28OCJEiT9MvUA8pn/Mr1GETB1L9LnKJMXI8lbB6U9JO1/Mae8UK4JCeZ1MLqSaf52AjsCL0yQru1ZCefxEWa0In/mtTHIaQpyWQ9ihSINxwnZtyZzBTIhxyYRtFKmuhCsjy/P6RDzKtGmdwZ0YM9hxwMR/HMNcpGNmROkqFInfEeiyDJnzgPdtIaRoKt9Sij4fnGvVzD+ByY3TyTrmFCNOTN6RvI9ojEaMwUqWS9C6yUd4xEnDDJTkQ+DwTCPPRHIkzG4jSZCI3EHITMqZRErHwURYQPiATJ+/NAOTKv6BsnE01RNlLRE62TFUVRMGxRoaDO/+seFHxR4wvCDGTRGu28HbUTH63OHGXRIpVRR4TGFjXSJS1AITXP/4RSDNS/8lQ+0TRM6nPSEZXGLI1S6qPSK9XG9bPS6uPSAvxS5zjT/y7dwDIVPp8c09ljU65U0yCNUwec0/9MU9nL09/b06msU/u70+b404ZUwEI11ENF1ERV1EVl1EZ11EeF1EiV1Eml1EpdVDm01EzV1ELF1E311Ev91FBVwE4V1VI11TftvdzJHfNa1fNKUVd10VZNr1ed1Vil1W1j1VuVVVyF1Vfd1V/V1WC11WH11WfMVWI9Vl8VVmVF1l5lVU9x1mitVWal1mS1VmnlVeoB1mad1mvtVmwF1lbd1moF12W91gck129V12xl13H11nY1Vx8DvUB100EtPQRlQ1SNvT4FVHrNR3sVP38N0yn1SX7dUunT1y8EWOkUWIVd2MmsV/nLvv+HtcSHJcGGXT+DLT2N1VMxpdgV/Nj+m1eMbdKQ5cWxNFldS9nRXNkqpdiEJVkwrT+O3Vf1u9jfg9kmHdkuDMSMCD0fc5Wx8tmgbdmY7bOgXYixYiOafdKhUdqlQdouZFpMpFHG+9fQW9qadVqsxUWP5Fr2g1oJ1IiiJVESHUlmNEklKUaCuBt1Sr2pxVPrOUlcXEov9EG5jcSB/E3vNEY+3Vr36J6nxMkD/UMhrBObOEY/BUWeqFqXgFtD3FlUPMu0TUJ3up7pAdDMpUoDjdlUlUjB3MeleccBpEgJdc51GqdfZMyufJ5uHEpF7FuITdCcBJAKnUgOvd28NV3ezNn/1gMQrwHKNzqIIeKU4+zInDzel9AJnuzcaTJKto3DLNyZmrxbJcwe4fUex+TMx21asipGbGzP0NuP2DVb+2DGtEKrp+VcQA3d7ExQtf1Pig3C6Ane3dvPu4kPXIGt/YVP72lekRxMJaxFsNJW6gXTyjTG14xHsj1BsmJIYrTNOXzMYXQPgrxK9lxO7iyIogHL7hFelC3fLbye5nlGXZTJ9bgbl/ya6Tkr/JXF/72d59VOULRcBoVf26PMo1wvp6xPpqTED+1CZpyatOFeAK3KmLhKpXmj9c22jKRGqSzMgo3c/7AU+sgeNZqU51ESfvITeWqfKjZDGE69R7TcP+TM/6FdTAtlSNB8SLl1WaBsXTdUxKBlpAP1Tdr1TQG+3bLN3deliSLmzpc1YyUUGv39ppw8iTZaSy8mJyS8lLISY/AjRodkvJp8WsCt3jzEw9z8j4ckX5llvDcMvURuxglWXCU03JYwK5QYykTmY8gMxLHiD+JMwZ312QT92aC12qHdZaKtFKs9yX9k4Pw7WqW9QF/O2kyW2qx9CWzEVz/c2nRa2q99ZgqW2n+cZmy85M37Xawdq2aOX4u1Psuj0SO9USaN5Bg+53VeGr9t53wMv+Ej59U1Z20LZPQbZtltznTuvt61QEB+5e8DaOfLZ/IT54Lm5+6dWSk2v5s16P+V3/+E7meENsCItVEsbV5/ZkAYLWcj7WglBemPFmkcJWmPLumQPumRNumVRmmWLi+VbumYhumZJmljdemUxumb1umYTsCclmmfpukeLdGdDuqfNmkKBOqkJmqlNmqQZlpdJlqohjyp5uWopmqpTrur1mqoruqp5mqs7uqs/uqt1uWwNmuwRmuytuqxZuuyTuu2Xmu3Vmuvlmu4puu4xuu71muxruu+zmu+/uuznmus3mvBtmvALuy39uvEHmzDXmzEhmzD/thbddHKpuzLtuzMxuzN1uzO5uzP9uzQBu3RFu3SJu3TNu3URu3VVu3WZu3Xdu3Yhu3Zlu3apm2Jxu3c1u3/3ebt3vbt3wbu4Bbu4Sbu4jbu40bu5Fbu5Wbu5nbu54bu6Jbu6abu6rbu68bu7Nbu7ebu7vbu7wbv8Bbv8Sbv8jbv80bv9Fbv9Wbv9nbv94bv+Jbv+abv+rbv+8bv/Nbv/ebv/vbv/wbwABfwASfwAjfwAyduWuXm0VwIzSRo48PnVW3ikJxnQW1DBPdurakJ/rNdQd2JF7QyMmTLpWTYgxhd53DFT14IFV/BtcSTJTxxdrbwCpYq3GQr5oOtfPxgDM9uNNxmhQjjIH3DDrcT55kUP07cqOy9Q0nIGNe9HR/eohnhrhVyRbyd+7hFG3GOIB+ameRx7P4ZSXEnnxEU/54d5a7pFPEd2+PcjTXniRLOCQ9MmzRfKyGclLSzRzInRHcCZ/NtFC6c6u4pyaW1RzUXzpwYjEUszAdB2ksuq7RJr709FEJESpjImphYTjh3c0LXFJW8RjcXwukRFJjQmT7/cupWo6rMYkxv86N5y7ZkZY3AD/isj1EsTZJUXbwkxEKeGjJx9bmR9ePkyJ+Iw1InkzNiwgsVGkL2I+OtxnUkpXMkk3NUW3E8RA5uyXCUdbDax31kJKDBYls3o7zSQ0U+E0aBSc9KRSVJZJ3Z4oE+dYENc1WPD8d1j22fS910SyI+YbYVLOhhz+b53TMZW/Tc4oDPX3BkS20l4uidif+WlBR1Olw1Uhou7pr2UXdzFPh9ZM0xbJ+DwN8wL8f0VVtxH8U5rGOpeWQMhuRKwRXuQSO1bfX8lJQhYqRDKUmKjvcubSN3QsVkRCtXp/a4JCO9NEOcdMtL4eJcpx6+xGKYtEuFRye9NJpExMlX78uG1A3/YKSlP2ShyXarnJ4Hafp+vB6KR6Omf84/30jMNUPP8nqq977BscumX171XPs0+qacgfedD9Qwx5Q1InvYhEgvlnVE2YgKPmRNgfuat/rQqw+clHxcUat1unhMJUMqFHiXHHxJTE78sPisbEWVF/gHMXwlNHv51BT8DZr9wA+2fEOBh03/7XravE9+bI//mP/3CJ6eQUeUtCNyv19u2PeJKxbOrFF5Towac4SUFpb247X5XxzE24RP+qx3jrzzM0l097V74xUrL89O7edIcpSJTrl5X1f+ktxN3SWr9pzEOWdzTlyr93B9V5EnUmJlUsJesXriko8rQQEIFlcCsGjViuCVVoIItmIhqCC2iBInUqxo8SLGjBo3cuzo8SPIkCJHkixp8iTJVthUspSoMmLLljBX0pRJcyZMgzVvuuQZsydOnzV1/pz5s2jPl0JtKj26sylNQUqTLqVqEKpRn0R3VoTK1CrQpyjHki1r9izatGrXsm3r9i1cbFccxq1r9y7evHr38u3r9y9ZQYIB/xMubPgw4sSKFzNu7Pgx5MiSJ1OubPky5r8st24+ynnl56tPQ5P2bHr06ZilUbNWnXro686tYc+W7bo26Ni5cYu+7Zv2b9vAhwsvvjv4ceLJjfdW3pz5auTPl1Ofbj26c+zQdbsVLNW7wkTfBYcf/128QvKC0INfb758+vPv3ceH3569evzy69O/P1+/ffn9NyB/APYnYIEE+pcggws6iOCD+0UYoIQQWlghhhRqeGCGHG5oIIgKXvihiB2GyJ96I3q44okTslhielOpJViLKtZoIowvNmhjjjeSuCOOQP7ooo86EtkjkkIayaOSRTqZ5JFNQsmkiykGGSWWVKooY/9a3yn0pZdhgjmmmGWSeaaZaaK5ppptsvmmm3HCOaecddJ5p5154rmnnn3y+aefgQI6qKCFEkpnd4YqeuiijTKKp3qPOlpnpJNKKil7l1paZqabasrfp57a2R2YlYJapqnthbqqqKy62mqc4qUKpyBekDkrrGhaKWugnb6K4ny/xooqqlekKoixlIbZKbKHcomWYIgIFgBDgg0UgLEKSVveQg7BN1BC/W0rroTjsmfud+hqWy679qkryLvxtkuuu/OqN65A69a7L739njvvv/wNxALBDsELsFQsBLCfwgjr62+6Dh+s71zYEhyuvPaB+1BBE0P8cMDtIavwFXMZ3HD/xPyGvHLKH7c3cME0jnzFDAKRZzLODyNS8sEPEWStwRwLRLDL7z57VpjXhquwt2kqzCy2uUotrKYLZTv1ohX7HC6dD93aMdV2UpsQQnsOfFDTrv6sULJrY33n2Voju3DC+Zr8pcFkopye29mSzDehiYZHcAAzCMa0dxeLZ/LPM087g69vSx52qWH6THd6F7NtM9FEP1TyyZ5zbOzFD+HKZorUzg066ImYnDjnyZZus9BAz3X1q+MqhDLJh+c7ssE4T0t63qI2nOzumE9OLN8GN973z6oHO97YXw6MnrHXqoe7nqR+B/rCDq2N7e1Wx+2QzVFTvvz6dJrcsdU/Xxy+/0MDHY7twhZj2xDO9iOrKMkQ0i2B3I5gB3mdwhpCQJ81RH9C85+w8kMwGWxMgbezH/kcaLJEuE1tFuNaB9lnJq0VrBXXsh61uhDCM5XMbVb73uoWdTSzyKwhPkvfwjaGkKjJ74NWkxIQs3QlLQ2xiEMSEM9+5jravfB+c5sW3ZBVMWNF7Qri+ZkVp0Q92imEIOIh2fyMZavjla1hZszg6o5oxBUFMAC2Gp3Fgva55xFPjXZ0DxhltrclrbFIqSNc8JQIRe2ZZy4UZNvWpHI2a93wjgaaYVmSNi3nOUSFwNtYwwQ4PBG2j5PpGlrvILhIBT7RZ9tinPhwhzhDVQw/6v/T3CbZlkPEYbB+4ePh6Tw1to11az1y2xwZZ4C2XDXOS0Tr5JwGtpD9Vcx6mfwd3kIJygA0D4a9lJTg1hO6VniBjuaLHvqw5YVJItOT64sUz1ynOmVSkXb2Q5nbmMbDHFprhcri28LAVC1e9pJ09Dyjzzi2OnAJq1O8ixrJ/Fk/cI3thXvUFHtQpp6HevKPc/PW3wJ6Ufckyzumu18WJSo79f1vUN4zoQGTBTMF2iwRhFsn01qYNnPS1FWvmxtDFEa3lVYsoRIlSAmn+B1oCopwVxtYuJDaQPKB8WJaex/xjCq17QF1ZEyN2jJzuEOTIUCporoW2HAa1nIGSKrKZOn/HrtZMK7VLSE9dSvTwgoz7u0JkoE5GK/i41Fg7VVkEstY0f4qWJW1jGUgKyxiH2alaLnHSnz1K7Di47pmecywHlvsZeXTV8w61q/SutdgA6vYnjXWOwbyKGoTW1nVjouz5ukrYC0bW9XuKrXmEY9+aniwz84ntdxy0Gy/dNKaEpesxW2T54yr3OMut7nMVVQ2nSvd595KhFIsVC6nSyjdUbe7kPKuoIarXfCOt7zkxVp2z2ve9cowUU8KIhEdqUX5wreP8+UjfYWY3/jit7/v1a9/7/tf/g64Q50t8H7tCyrxqrfB7H2wgyPcNQlTGFFtkUpmMuwRnUyGwxq+iCAg/xPix3g4wyWe0YdTnJETi1jFXbErYUTjGBZjBsYosbGLL0Pjxuw4xSOeMWR6TBkhB8YsWElJjseCYY7guC4/3siReezhJje5LCx+1pGrPBaFbNgxT1YLl7iskWfNrGRXKInXkmySKf+PIiY0lkTMrGUjT0WKbpYiS0rq5Ym8WUZyhslcvtydieCZz9ZSyedKNuc117lkFbmuRJB1GCIr+dHUXHFCLEI0bDSLJJ9Tc0k4zDZs0EUiBTFhiGO06EhGOqkugXO2YC3oSTeFBZzOdESuZmtOS2XVKRG1rTud66HKJXN28fBRKRKuVwcAMb7+iIwa+GMMc5mUESFPRBJokP9UX+XHwl3JksG9kBEL9yqUTjKwEb3rlUBE0nN59paVAueFvKQhLClIu9e9mGmfmdTyTjVESn3sOPcbIsNm97+PXW+6NITQAr+1vgFz7pN8mZy5pqb90HZmAgacfOQGqkBu/e6DfK6LW5OLnOPXb1CD+CVpTvNKqLk7iIcbLz+GyEEWHmLT0QXmipGxo99s6n7LnGPwDgmwz2xDU/f60ivBdVxE7fRm58RkgD4IYmZ9FhmF2H4X76JBCMLpfGKL1Bar9/7KHnK4mjyB4+YfvX949MpI/d7yRt/K7a0XDzNc4IGei1zQF/HEJJzU056lvFcOl5uP2OCctmW9FQ8XYAP/PNreOvNDaD3oV9/w2s32+v7iTjBSNw00ZD9z2S8ulww+nmzIGzfLx7xwREN923AGtNYXb2qWUD3SNkT00hnjYUkHHyYj47Pk29KboNd+c3GeO9KVcumGVyR7t9ufYaA/5okkxCANW73J0RZ+tTvdO2nfuNMHMnYGvl3s215Y7FesFLoEndCSh/pdkI35XSvF58JmDL8hnJs5HrgN3LUp3eCZkLxpH7QlnKQ92mb4nMQxIEZMG84R1AGNDenZGsepRPUE2guJ3daEXcnFj9jFVPjFX8vlBL7lmtJlXrb9X154W/Z0Xab9n6LNGN+h2sx530sQEP65RYnh287Zmg9y/x9eiNnTtVmsZQ+fDd5f5J6VucSSbdu1eUkBXmEEggloYCGv1dtgGN+3YZuMqeCLGVq9bQYa8kWJfZtO1FzNNUbFpSG4NUUc6h4V0qES5kQSToUb5lm0USBZTJxJSKEZRgYhGoYhZkYismEj7sUjCp8QHqKJCeLkWaIiBlmQYeLecSKhUWJm3OFjLGKNRWIfbqKKkeJJeCIoXuKQsaJmtBiJwaLCide2jcktekkufsku9tqZ9GK34eIvDqOYAKMxEqMwFiMy6uIy8uKdHKMyRmMyTiMzSqObQCM1OqM1amM2+uI2euOcYGM1dmMwjqM5vok4cuM5guM6lqM6viM7wv+jO8YjPc6jPWIjLZrbcRAFP9pGP+6jPwYkQA5kb/xjQQrkQRKkQS4kQjKkQjYkRD6kRCYkRTpkRUbkRU6kRW5kZ/CiRmIkR35kF4pkRpZkSCZkr5HkSa4kSDbk5rViZYjinoFaGS6GKebFHmaYKq4ZTFrGTTpi7O2kXwhlX/yk5l1YT9IdLaIYTS5lvDnljUHl5E2ioWGbSCBStxElFZJHfYhbfGylUTqbH96hVcYIJI6loKWH8cnkYTyZWVZlpIVhXaDlo8mld+QfXd5ZFWqlKyJlpOUTyYWE0R2OAi3aqWXewMSZIv2M5z3ch+1lQhBg98EaIu0Fv42a8UXmzp3/WfcJ3w5a5a3N3PbM5VQURLbE5ag1YS0eIKopWw6inFR+BF8inUuI3cW55ZglkO8NRL25WVh43ogRXfpd2/ehVO+p2A4+nktAhKL1m6PhpAM6Hs7hm7oB2QEK4MHZW9/ZxY89p+PdIHU+nlTy3Wb2H/5RX/bRYnem3tjh21o9Hc8QGvbNxMihTxFKWv00Rdm4XNmJWRcFm/jlWLqJJ0xcmtfsGuDpxc3xns6xm+kQqE0mnoMyXURQE/xhw3EKIVGEXGvGYMylGdaRZoHem8OF3NU5G1Wu3sp1SwOJ34EqHrhUHzXRm/qBYGHWGbVcaIuuqLQlTIZ+GLDBhHS6lWby/08nUmjODd3M2Q0UtuXQeWiudYxpSul4Uqi/xRmVvtu7iWjM2R1FwBx6FkZYcsR6rltiGpDCkJ6ZwcwPKRtzsmi3oJ/TsaBump3BmRtCYCgBPqbL5VtTgM5m3o5l7l6Xct/tqASgStlLBJ2JglsLOWqC9qWR6p3xDQ18jt5c/tjUWcTtKWDWpSjaxFk+JRCaTh9qMURP8KYNqd/unFqMXuHXNWb/ed+lpamAzl91DmD/NWlfykURWp637WliBCClLueTOeZLeqfl1VpskuetWUQYhmmMqad8Nl2mlWp7ymWUGqEJBd7wFJMgrIBcGN5ygtu6GanvoVyFCmslRhoHIv8auTlesc5gq41atN7eEurgRASb0q0cvhreeO5gtnQrlGZbbIJGvc7btinSqx1sR+Tjl9YgoFkf7c0aIqUaH0ymIqXHDZ5mZoImp/2pL4ocW/pkowUnIFJhUdZZGA7Gt32hLK5lntVhHiZhVc7sy74sl17hiA2Gnjmqw0IZtfKZRWTZ9gVFlP2m0gYF0TJtKobZ0vIEX/Bl0I5ETRatZkCt025t1C1a1ZJpiiYlIn5tqJHtW5itbKIikA6t2I5tTLLcmF4i2j7s3E6h8rWtZMQtXgalJs5iiultRozjLQ4uNxIuOBpuMCKu4hYu4x5u4ybu4y6u404u5FKu5FYu5l7/ruZGLudabudm7udurudSruCGrumOLuiiruimLuuuruuebuvCruSSSm7oo7kdpO3ebkfiLu/2Y+7ebu327u76rvAGL/Ee7/Amr/F2ZErqrvI67/JGL/ROr+5Kb/VSL/Bi7+9ub/Fqb0H64vNer/hm7/hyr+82b/l2b/oir/U67/qGL/nGr/nC7/yyLd7qK2XMptoqqt/+bdjer2e+It+S2P62q18CsBzWrWDCLeCyRQMrnwJX2t26BJFd7RluWPUisJv5IZZN2QODRBtemQdH8Ipt8Is1xQdvxJeJMAofbAh3MB0eJVvQ4KmtWL5A69Px6kUcqgyCMAnT2XJKq0oh/yrbfC0NtlnDtltk/vAF51qH4h7z3WXX/ltn8qzPMtJqLuG/Bt4FSqvEhe2W2nDyEai2kumFlmxuNiWFPqcAxlq2jXGvShoMGuyzKvFMPmuxwpm96bEOp4V5shv3/Wq+ZbGjPaAYEiobT+BLgt8S7pye7Q/HPl0UveBprun9hRy3OSEihQtXolpbqVkV+mmBEif1RarNIanTLR3P/apnuhwCMl6F/h6pcanQFSvOWSiDZqpETJ++fU7lhWYmTnCUqgQX5QtGbZxmBh4xL426jlugyaegkloyXxAP4tuymbKL1R12Lp2iMSccn+38pdqFSjPC9VwfA0Z3Nl53GuFCYf8nHpLrlbrg+mEyu64F5bmzQdQncabnIqsqBo6qrYUeFeVzs6EUOdOo80xF+SAgMpvQ1F0XQZGzmnHYy30n66BcNNMrHSepkB6qDTXqz0kouuYwq5rZOdPZvfUrvz2qFGFq145otmkaYUJojIHxBXJOwkQm5pFO1wVcZNrauyH0k8EewIXY6G3ppEoyZ560yQpprlLwnUogIbPyEwar/nInwdHxE/ZbPl+11WZ1PLOgVa/m0nnxFw6ymNqvqioSEX7a/hBNPvkzUMOVt9Ib9zWqkdJf7xnQLQLeDSdZFToaw4XzvpJtAGbe8J1rIof0wgHzZRJcwJZrpyUeZRPyTtP/hkrDc2F49cO+Wqx12v/0K5HaWUkpGqp1LKRNbLiFNqLimi/7KmpPtMhimB7LpaeeZSD3G/HZdhnbJKFh227vJRp3yW93rFeu5Skqpm7bYGD3tl+k8JeCRbSFxVd8RVZMN3YrrdEycbw1LVV499T6JtKKN/56N1IEbRlWt2867IkZLdEyMcRqcAB3GHezGgH3r4ZB9yfK977Vt0ZwdiuXN48B+Gb/L3/zM30P8B1L2d9Sq1M8uFhAuIRHOIVPuIVXOIZfuIZnOIcD34Z/eIeHuGekmoiXOIifeIeHhomvOIpbB4u/+IdXG4zPeItXuH5TsPvmuI7vOI/3uI//OJAHtbmQDzmRF7mRHzmSJ7mSLzmTN7mTPzmUR7k+HjiVV7mVXzmWZ7mWbzmXd7mXfzmYh7mYjzmZl7mZnzmap7marzmbt7mbvzmcx7mczzmd17md3zme57me7zmf97mf/zmgB7qgDzqhF7qhyzeWYW1XKDp5g3fUJvqiR3qjPzqjOzrXQvqkX3qlU7qkW7rUcnqmf7qmdzqoezqmm/qmj3qon3qpt7qqozqpv7qrizqts7qs33qtBwQAOw==\" alt=\"image\"\u003e\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"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":"gastrointestinal bleeding, CT angiography, machine learning, clinical decision support, predictive modeling","lastPublishedDoi":"10.21203/rs.3.rs-7086958/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7086958/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003ePurpose\u003c/h2\u003e\u003cp\u003eCT angiography (CTA) is valuable in evaluating acute gastrointestinal (GI) bleeding but lacks guidance for use in patients who are neither hemodynamically unstable nor clearly stable, creating a gray zone of uncertainty in imaging decisions. Our goal was to develop a risk stratifying machine learning (ML) model for hemodynamic borderline patients with GI bleeding to help mitigate testing uncertainty by predicting the probability of a positive CTA.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eWe retrospectively analyzed 11,938 patients with GI bleeding from the MIMIC-IV database. Among 890 CTA scans, 140 were eligible after applying exclusion criteria. A logistic regression model with SMOTE upsampling was trained using seven routine lab values obtained within 24 hours of CTA. Model performance was evaluated using recall, F1 score, and ROC-AUC.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThe model achieved an F1 score of 0.71, recall of 0.83, and ROC-AUC of 0.71. The features - delta hematocrit/hemoglobin and the maximum INR in the last 24 hours were influential predictors, while the feature minimum platelets in the last 24 hours was not. Logistic regression outperformed random forest and XGBoost in identifying true positives.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eA simple, interpretable ML model can assist in identifying patients most likely to benefit from CTA in GI bleeding. Its reliance on structured, readily available labs supports potential real-time integration into electronic health record workflows. With further validation, this approach could improve triage, reduce unnecessary scans, and support real-time decision-making.\u003c/p\u003e","manuscriptTitle":"A Machine Learning Model to Guide CT Angiography Use in Acute Gastrointestinal Bleeding: A Decision-Support Tool for Gray-Zone Cases","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-14 11:07:44","doi":"10.21203/rs.3.rs-7086958/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":"253cfb6a-ecde-41db-8069-2854514c32b8","owner":[],"postedDate":"July 14th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-08-11T08:38:35+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-14 11:07:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7086958","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7086958","identity":"rs-7086958","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}