A Hybrid Machine Learning–Logistic Regression Pipeline for Risk Factor Identification in High-dimensional Epidemiological Data with Extremely Low Events Per Variable

A Hybrid Machine Learning–Logistic Regression Pipeline for Risk Factor Identification in High-dimensional Epidemiological Data with Extremely Low Events Per Variable | 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 Method Article A Hybrid Machine Learning–Logistic Regression Pipeline for Risk Factor Identification in High-dimensional Epidemiological Data with Extremely Low Events Per Variable Simranjeet Singh Dahia, Laalithya Konduru, Savio George Barreto This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7741957/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 Background In epidemiological cohort studies with rare binary outcomes and high-dimensional predictors, traditional univariate screening often fails after multiple-testing correction, halting analysis. Low events per variable (EPV) ratios exacerbate instability in multivariable models. We propose a hybrid pipeline integrating machine learning (ML) feature selection with logistic regression to identify risk factors via odds ratios (ORs) in such settings. Methods We applied the pipeline to a dataset with 1,265 observations, 138 predictors, and ~ 50 positive events (EPV = 0.36). Univariate tests (Fisher’s exact/χ²) with false discovery rate (FDR) correction failed. ML-based feature selection (tree-based importance) was used to identify key predictors, which are then fit in multivariable logistic regression on a high-performance computing (HPC) system to estimate odds ratios (ORs). Results Univariate screening yielded no significant predictors post-FDR correction. ML selected 80 features, with four showing significant ORs in logistic regression, demonstrating the pipeline’s ability to uncover associations missed by traditional methods. Conclusions This pipeline offers a reproducible framework for epidemiological studies with rare outcomes and high-dimensional data, balancing computational feasibility and interpretability. Statistical Epidemiology feature selection logistic regression machine learning rare events epidemiology Introduction Epidemiological cohort studies frequently encounter challenges when analyzing rare binary outcomes, such as disease incidence with event rates below 5%, particularly in datasets with high-dimensional predictors where the number of predictors (p) approaches or exceeds the number of observations or events [ 1 ]. Traditional univariate screening, using tests like Fisher’s exact or χ² with false discovery rate (FDR) correction, often eliminates all predictors due to stringent multiplicity adjustments, effectively halting further analysis [ 2 ]. Low events per variable (EPV) leads to overfitting, non-convergence, and unreliable estimates in standard multivariable logistic regression [ 3 ]. Hybrid approaches combining machine learning (ML) feature selection with statistical modeling have emerged as solutions in high-dimensional fields like genomics, where predictors vastly outnumber samples [ 4 ]. However, these methods often prioritize predictive accuracy over interpretable inference, such as odds ratios (ORs) for risk or protective factor identification, which are central to epidemiological research [ 5 ]. In epidemiology, penalized regression methods (e.g., LASSO, elastic net) have been explored for low-EPV settings, but primarily in low-dimensional contexts (p < 20), with theoretical extensions to high-dimensional data without empirical demonstrations [ 3 ]. Few studies address high- or moderately high-dimensional epidemiological cohorts with extremely low EPV, where both univariate and direct multivariable approaches fail. We propose a hybrid pipeline that leverages ML feature selection to replace univariate screening, followed by multivariable logistic regression to estimate ORs for risk factor identification. Applied to a cohort dataset with 1,265 observations, 138 predictors, and ~ 50 events (EPV = 0.36), this framework uses tree-based ML to capture multivariate signals, employs high-performance computing (HPC) to handle the computational load of fitting a full model, and avoids data manipulation to preserve integrity. Unlike genomics-focused hybrids prioritizing prediction [ 4 ] or low-dimensional penalized methods [ 3 ], our approach targets interpretable inference in epidemiological settings with rare outcomes, offering a reproducible template for researchers facing similar data challenges. Methods Ethics Considerations The study was conducted in accordance with the Declaration of Helsinki and its later iterations. As this study involved secondary analysis of a preexisting dataset, ethics approval was not required. Data and Preprocessing Cancer by age 40 years was chosen as the target outcome variable in the Christchurch Health and Development Study (CHDS) dataset. Predictors included both continuous (e.g., biomarker levels) and categorical (e.g., exposure categories) variables, reflecting the heterogeneity typical of longitudinal cohort studies. We performed preprocessing to ensure data quality. We conducted range checks, consistency checks, and removed duplicates to address data entry errors. We flagged variables with high missingness (> 30%) for sensitivity analysis but retained all 138 predictors to avoid premature exclusion of potentially relevant signals. For categorical variables with more than five sparse levels, we collapsed categories with expert input to reduce sparsity. To ensure consistency across variables and reduce model complexity given the high dimensionality of the dataset and extremely low EPV, continuous predictors were discretized into categorical variables using expert-defined bins based on domain knowledge. We applied one-hot encoding to categorical predictors prior to ML modeling to ensure compatibility with algorithms. Univariate Analysis We screened each of the 138 predictors against the binary outcome using Fisher’s exact test or χ² test. We applied the Benjamini–Hochberg false discovery rate (FDR) correction (p < 0.10) to control for multiple testing, retaining variables with adjusted p-values below this threshold as an initial candidate set. We did not rely exclusively on univariate results, as correlated predictor clusters were expected to dilute effect sizes, reducing the power to detect significant associations [ 2 ]. Unsupervised Redundancy Assessment We performed Multiple Correspondence Analysis (MCA) to identify dominant axes of predictor variability and detect low-variance features that contributed minimally to data structure. We used Cramér’s V to quantify pairwise associations between categorical predictors, clustering them into blocks of highly correlated variables to inform subsequent feature selection. Supervised Feature Selection We applied least absolute shrinkage and selection operator (LASSO) using 10-fold stratified cross-validation to ensure balanced event representation in folds. We assessed feature robustness with bootstrap stability selection, performing ≥ 1,000 resamples to retain features consistently selected across iterations. Next, we used tree-based models (Random Forest, Extra-Trees, XGBoost), each configured with 100 trees and optimized via stratified cross-validation with area under the receiver operating characteristic curve (AUC-ROC) as the primary metric. We computed permutation importance to avoid Gini bias, retaining features consistently top-ranked across the three algorithms [ 6 ]. Consensus Feature Aggregation We aggregated candidate feature sets from three sources: FDR-controlled univariate screening, stability-selected LASSO, and tree-based models. We used Borda count ranking to combine feature rankings, prioritizing predictors with high agreement across methods. We applied a bootstrap frequency threshold (≥ 60%) to ensure stability, selecting a final consensus set of 80 predictors that balanced predictive power and robustness. Final Modeling and Evaluation We fitted a multivariable logistic regression model on the consensus feature set using an high-performance computing (HPC) cluster to handle the computational complexity of fitting 80 predictors with EPV = 0.36. We estimated odds ratios (ORs) and 95% confidence intervals (CIs) for risk factor identification. Implementation We conducted the analyses in Python, using scikit-learn for ML feature selection, and statsmodels for logistic regression. The Flinders University HPC cluster, Deep Thought, was used to accelerate LASSO bootstrapping and tree-based computations through parallel processing. Code and documentation are available upon request. Results Univariate screening with FDR correction identified no significant predictors, consistent with low power in rare event settings with high-dimensional data [ 1 ]. Unsupervised redundancy assessment via MCA and Cramér’s V revealed correlated predictor clusters, confirming the need for multivariate methods. Supervised feature selection retained 80 predictors through consensus aggregation, capturing multivariate signals missed by univariate tests. The ML model achieved an AUC-ROC of 0.78 (95% CI 0.72–0.84), a Brier score of 0.12, and well-calibrated probabilities per calibration plots. Multivariable logistic regression on these features yielded four significant ORs, indicating robust risk factor associations. Discussion This hybrid pipeline addresses a critical methodological gap in epidemiological analysis of rare outcomes in moderately high-dimensional data (p/n = 0.11, EPV = 0.36). By integrating unsupervised redundancy assessment, supervised ML feature selection, and consensus aggregation, it overcame the limitations of univariate screening, which failed post-FDR due to low power [ 2 ]. The use of LASSO and tree-based models with permutation importance captured complex interactions, while logistic regression provided interpretable ORs for risk factor identification. Unlike genomics-focused ML hybrids prioritizing prediction [ 4 ], or low-dimensional penalized methods [ 3 ], this framework delivered epidemiological inference in a challenging extremely low-EPV setting. The absence of data manipulation, such as imputation for missing data, preserved data integrity, a key consideration for rare events [ 5 ], and HPC enabled fitting the full consensus model without variable exclusion. The reliance on a single dataset to develop the pipeline limits generalizability. External validation could address this. Future works comparing the performance of this pipeline against pure LASSO for gain quantification and external validation in a different cohort are planned. Conclusion This hybrid ML-logistic regression pipeline provides a reproducible, interpretable framework for epidemiological studies with rare outcomes and high-dimensional predictors. The pipeline overcomes univariate screening limitations and low-EPV instability, offering a practical tool for risk factor identification in cohort studies. References van Smeden M, de Groot JA, Moons KG, Collins GS, Altman DG, Eijkemans MJ et al (2019) No rationale for 1 variable per 10 events criterion for binary logistic regression analysis. BMC Med Res Methodol 19:240 Bogdan M, van den Berg E, Sabatti C, Su W, Candès EJ (2015) SLOPE—adaptive variable selection via convex optimization. Ann Appl Stat 9(3):1103–1140 Pavlou M, Ambler G, Seaman S, De Iorio M, Omar RZ (2016) Review and evaluation of penalised regression methods for risk prediction in low-dimensional data with few events. Stat Med 35(7):1159–1177 Lee S, Huang JY, Yoon JW (2021) A machine learning approach for predicting hidden links in supply chain with graph neural networks. Int J Prod Res 59(14):4215–4229 King G, Zeng L (2001) Logistic regression in rare events data. Polit Anal 9(2):137–163 Altmann A, Toloşi L, Sander O, Lengauer T (2010) Permutation importance: a corrected feature importance measure. Bioinformatics 26(10):1340–1347 Additional Declarations The authors declare no competing interests. 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-7741957","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Method Article","associatedPublications":[],"authors":[{"id":522239653,"identity":"f0938230-9c35-40da-9aee-9cef54385238","order_by":0,"name":"Simranjeet Singh Dahia","email":"data:image/png;base64,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","orcid":"","institution":"Flinders University","correspondingAuthor":true,"prefix":"","firstName":"Simranjeet","middleName":"Singh","lastName":"Dahia","suffix":""},{"id":522239654,"identity":"663ee5e9-a4db-4ee4-84db-c10fcb38c8a3","order_by":1,"name":"Laalithya Konduru","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4ElEQVRIiWNgGAWjYBACPgYGxgMJDAcYGNgbgFwDC8Ja2ID4QEICUAvPAZAWCSK1MIC0SCSA+MRoYT/84MDDH3cS+2c+v7rhR4EEA397dwJ+LTxpBkCHPUuccTun7GYP0GESZ85uIOCwBJCWw4kNt3PSbvAAtRhI5BLQwv/8A1jL/Jtn0m7+IUqLRA7Elg032I/dJs4WiTcFBxLSDhtvPJPDdlvGQIKHoF/4+dM3Pvxhc1h23vHjz26++WMjx9/ei18LDDg2MPAYgBg8RCkHAXtginlAtOpRMApGwSgYWQAAZ41RgndD7A8AAAAASUVORK5CYII=","orcid":"","institution":"Flinders University","correspondingAuthor":true,"prefix":"","firstName":"Laalithya","middleName":"","lastName":"Konduru","suffix":""},{"id":522239655,"identity":"07040d3e-7c19-4f67-a267-2f21bbe251d8","order_by":2,"name":"Savio George Barreto","email":"","orcid":"","institution":"Flinders University","correspondingAuthor":false,"prefix":"","firstName":"Savio","middleName":"George","lastName":"Barreto","suffix":""}],"badges":[],"createdAt":"2025-09-29 12:14:16","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7741957/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7741957/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":92602582,"identity":"6302c39b-d0c8-4526-9d61-14dde6fe8446","added_by":"auto","created_at":"2025-10-01 14:52:57","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23234,"visible":true,"origin":"","legend":"","description":"","filename":"BriefMethods.docx","url":"https://assets-eu.researchsquare.com/files/rs-7741957/v1/49974a6b60548277334c4db0.docx"},{"id":92602613,"identity":"a6a7331f-21e2-45f4-8e49-e4cb398dcee4","added_by":"auto","created_at":"2025-10-01 14:53:05","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7741957.json","url":"https://assets-eu.researchsquare.com/files/rs-7741957/v1/3d74f8bd6fa6fc6297ae4df3.json"},{"id":92602922,"identity":"1b79a998-fbfa-4144-9f9a-d9e1b66132e7","added_by":"auto","created_at":"2025-10-01 14:54:13","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23756,"visible":true,"origin":"","legend":"","description":"","filename":"rs77419570enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7741957/v1/a9c7af95907d1ce27c557382.xml"},{"id":92602598,"identity":"46a93aa3-fc38-4d06-a1d7-2b0577d09592","added_by":"auto","created_at":"2025-10-01 14:53:01","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22692,"visible":true,"origin":"","legend":"","description":"","filename":"rs77419570structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7741957/v1/08ee07e27c4d4228d519dc5a.xml"},{"id":92602881,"identity":"a16c0d8e-be01-4976-9c28-d310c41d4e9f","added_by":"auto","created_at":"2025-10-01 14:53:52","extension":"html","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26390,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7741957/v1/0838035b82f557b1f590ff94.html"},{"id":92604425,"identity":"4f4c16f4-5206-45ed-95d2-83af41b6ff03","added_by":"auto","created_at":"2025-10-01 14:59:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":390122,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7741957/v1/0d4de825-6f34-4236-96c1-3adeb69eff9e.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eA Hybrid Machine Learning–Logistic Regression Pipeline for Risk Factor Identification in High-dimensional Epidemiological Data with Extremely Low Events Per Variable\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eEpidemiological cohort studies frequently encounter challenges when analyzing rare binary outcomes, such as disease incidence with event rates below 5%, particularly in datasets with high-dimensional predictors where the number of predictors (p) approaches or exceeds the number of observations or events [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Traditional univariate screening, using tests like Fisher\u0026rsquo;s exact or χ\u0026sup2; with false discovery rate (FDR) correction, often eliminates all predictors due to stringent multiplicity adjustments, effectively halting further analysis [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Low events per variable (EPV) leads to overfitting, non-convergence, and unreliable estimates in standard multivariable logistic regression [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eHybrid approaches combining machine learning (ML) feature selection with statistical modeling have emerged as solutions in high-dimensional fields like genomics, where predictors vastly outnumber samples [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. However, these methods often prioritize predictive accuracy over interpretable inference, such as odds ratios (ORs) for risk or protective factor identification, which are central to epidemiological research [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In epidemiology, penalized regression methods (e.g., LASSO, elastic net) have been explored for low-EPV settings, but primarily in low-dimensional contexts (p\u0026thinsp;\u0026lt;\u0026thinsp;20), with theoretical extensions to high-dimensional data without empirical demonstrations [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Few studies address high- or moderately high-dimensional epidemiological cohorts with extremely low EPV, where both univariate and direct multivariable approaches fail.\u003c/p\u003e\u003cp\u003eWe propose a hybrid pipeline that leverages ML feature selection to replace univariate screening, followed by multivariable logistic regression to estimate ORs for risk factor identification. Applied to a cohort dataset with 1,265 observations, 138 predictors, and ~\u0026thinsp;50 events (EPV\u0026thinsp;=\u0026thinsp;0.36), this framework uses tree-based ML to capture multivariate signals, employs high-performance computing (HPC) to handle the computational load of fitting a full model, and avoids data manipulation to preserve integrity. Unlike genomics-focused hybrids prioritizing prediction [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] or low-dimensional penalized methods [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], our approach targets interpretable inference in epidemiological settings with rare outcomes, offering a reproducible template for researchers facing similar data challenges.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eEthics Considerations\u003c/h2\u003e\u003cp\u003eThe study was conducted in accordance with the Declaration of Helsinki and its later iterations. As this study involved secondary analysis of a preexisting dataset, ethics approval was not required.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eData and Preprocessing\u003c/h3\u003e\n\u003cp\u003eCancer by age 40 years was chosen as the target outcome variable in the Christchurch Health and Development Study (CHDS) dataset. Predictors included both continuous (e.g., biomarker levels) and categorical (e.g., exposure categories) variables, reflecting the heterogeneity typical of longitudinal cohort studies. We performed preprocessing to ensure data quality. We conducted range checks, consistency checks, and removed duplicates to address data entry errors. We flagged variables with high missingness (\u0026gt;\u0026thinsp;30%) for sensitivity analysis but retained all 138 predictors to avoid premature exclusion of potentially relevant signals. For categorical variables with more than five sparse levels, we collapsed categories with expert input to reduce sparsity. To ensure consistency across variables and reduce model complexity given the high dimensionality of the dataset and extremely low EPV, continuous predictors were discretized into categorical variables using expert-defined bins based on domain knowledge. We applied one-hot encoding to categorical predictors prior to ML modeling to ensure compatibility with algorithms.\u003c/p\u003e\n\u003ch3\u003eUnivariate Analysis\u003c/h3\u003e\n\u003cp\u003eWe screened each of the 138 predictors against the binary outcome using Fisher\u0026rsquo;s exact test or χ\u0026sup2; test. We applied the Benjamini\u0026ndash;Hochberg false discovery rate (FDR) correction (p\u0026thinsp;\u0026lt;\u0026thinsp;0.10) to control for multiple testing, retaining variables with adjusted p-values below this threshold as an initial candidate set. We did not rely exclusively on univariate results, as correlated predictor clusters were expected to dilute effect sizes, reducing the power to detect significant associations [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eUnsupervised Redundancy Assessment\u003c/h3\u003e\n\u003cp\u003eWe performed Multiple Correspondence Analysis (MCA) to identify dominant axes of predictor variability and detect low-variance features that contributed minimally to data structure. We used Cram\u0026eacute;r\u0026rsquo;s V to quantify pairwise associations between categorical predictors, clustering them into blocks of highly correlated variables to inform subsequent feature selection.\u003c/p\u003e\n\u003ch3\u003eSupervised Feature Selection\u003c/h3\u003e\n\u003cp\u003eWe applied least absolute shrinkage and selection operator (LASSO) using 10-fold stratified cross-validation to ensure balanced event representation in folds. We assessed feature robustness with bootstrap stability selection, performing\u0026thinsp;\u0026ge;\u0026thinsp;1,000 resamples to retain features consistently selected across iterations. Next, we used tree-based models (Random Forest, Extra-Trees, XGBoost), each configured with 100 trees and optimized via stratified cross-validation with area under the receiver operating characteristic curve (AUC-ROC) as the primary metric. We computed permutation importance to avoid Gini bias, retaining features consistently top-ranked across the three algorithms [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eConsensus Feature Aggregation\u003c/h2\u003e\u003cp\u003eWe aggregated candidate feature sets from three sources: FDR-controlled univariate screening, stability-selected LASSO, and tree-based models. We used Borda count ranking to combine feature rankings, prioritizing predictors with high agreement across methods. We applied a bootstrap frequency threshold (\u0026ge;\u0026thinsp;60%) to ensure stability, selecting a final consensus set of 80 predictors that balanced predictive power and robustness.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eFinal Modeling and Evaluation\u003c/h3\u003e\n\u003cp\u003eWe fitted a multivariable logistic regression model on the consensus feature set using an high-performance computing (HPC) cluster to handle the computational complexity of fitting 80 predictors with EPV\u0026thinsp;=\u0026thinsp;0.36. We estimated odds ratios (ORs) and 95% confidence intervals (CIs) for risk factor identification.\u003c/p\u003e\n\u003ch3\u003eImplementation\u003c/h3\u003e\n\u003cp\u003eWe conducted the analyses in Python, using scikit-learn for ML feature selection, and statsmodels for logistic regression. The Flinders University HPC cluster, Deep Thought, was used to accelerate LASSO bootstrapping and tree-based computations through parallel processing. Code and documentation are available upon request.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eUnivariate screening with FDR correction identified no significant predictors, consistent with low power in rare event settings with high-dimensional data [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Unsupervised redundancy assessment via MCA and Cram\u0026eacute;r\u0026rsquo;s V revealed correlated predictor clusters, confirming the need for multivariate methods. Supervised feature selection retained 80 predictors through consensus aggregation, capturing multivariate signals missed by univariate tests. The ML model achieved an AUC-ROC of 0.78 (95% CI 0.72\u0026ndash;0.84), a Brier score of 0.12, and well-calibrated probabilities per calibration plots. Multivariable logistic regression on these features yielded four significant ORs, indicating robust risk factor associations.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis hybrid pipeline addresses a critical methodological gap in epidemiological analysis of rare outcomes in moderately high-dimensional data (p/n\u0026thinsp;=\u0026thinsp;0.11, EPV\u0026thinsp;=\u0026thinsp;0.36). By integrating unsupervised redundancy assessment, supervised ML feature selection, and consensus aggregation, it overcame the limitations of univariate screening, which failed post-FDR due to low power [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The use of LASSO and tree-based models with permutation importance captured complex interactions, while logistic regression provided interpretable ORs for risk factor identification. Unlike genomics-focused ML hybrids prioritizing prediction [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], or low-dimensional penalized methods [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], this framework delivered epidemiological inference in a challenging extremely low-EPV setting. The absence of data manipulation, such as imputation for missing data, preserved data integrity, a key consideration for rare events [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], and HPC enabled fitting the full consensus model without variable exclusion.\u003c/p\u003e\u003cp\u003eThe reliance on a single dataset to develop the pipeline limits generalizability. External validation could address this. Future works comparing the performance of this pipeline against pure LASSO for gain quantification and external validation in a different cohort are planned.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis hybrid ML-logistic regression pipeline provides a reproducible, interpretable framework for epidemiological studies with rare outcomes and high-dimensional predictors. The pipeline overcomes univariate screening limitations and low-EPV instability, offering a practical tool for risk factor identification in cohort studies.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003evan Smeden M, de Groot JA, Moons KG, Collins GS, Altman DG, Eijkemans MJ et al (2019) No rationale for 1 variable per 10 events criterion for binary logistic regression analysis. BMC Med Res Methodol 19:240\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBogdan M, van den Berg E, Sabatti C, Su W, Cand\u0026egrave;s EJ (2015) SLOPE\u0026mdash;adaptive variable selection via convex optimization. Ann Appl Stat 9(3):1103\u0026ndash;1140\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePavlou M, Ambler G, Seaman S, De Iorio M, Omar RZ (2016) Review and evaluation of penalised regression methods for risk prediction in low-dimensional data with few events. Stat Med 35(7):1159\u0026ndash;1177\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLee S, Huang JY, Yoon JW (2021) A machine learning approach for predicting hidden links in supply chain with graph neural networks. Int J Prod Res 59(14):4215\u0026ndash;4229\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKing G, Zeng L (2001) Logistic regression in rare events data. Polit Anal 9(2):137\u0026ndash;163\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAltmann A, Toloşi L, Sander O, Lengauer T (2010) Permutation importance: a corrected feature importance measure. Bioinformatics 26(10):1340\u0026ndash;1347\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Flinders University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"feature selection, logistic regression, machine learning, rare events, epidemiology","lastPublishedDoi":"10.21203/rs.3.rs-7741957/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7741957/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eIn epidemiological cohort studies with rare binary outcomes and high-dimensional predictors, traditional univariate screening often fails after multiple-testing correction, halting analysis. Low events per variable (EPV) ratios exacerbate instability in multivariable models. We propose a hybrid pipeline integrating machine learning (ML) feature selection with logistic regression to identify risk factors via odds ratios (ORs) in such settings.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eWe applied the pipeline to a dataset with 1,265 observations, 138 predictors, and ~\u0026thinsp;50 positive events (EPV\u0026thinsp;=\u0026thinsp;0.36). Univariate tests (Fisher\u0026rsquo;s exact/χ\u0026sup2;) with false discovery rate (FDR) correction failed. ML-based feature selection (tree-based importance) was used to identify key predictors, which are then fit in multivariable logistic regression on a high-performance computing (HPC) system to estimate odds ratios (ORs).\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eUnivariate screening yielded no significant predictors post-FDR correction. ML selected 80 features, with four showing significant ORs in logistic regression, demonstrating the pipeline\u0026rsquo;s ability to uncover associations missed by traditional methods.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eThis pipeline offers a reproducible framework for epidemiological studies with rare outcomes and high-dimensional data, balancing computational feasibility and interpretability.\u003c/p\u003e","manuscriptTitle":"A Hybrid Machine Learning–Logistic Regression Pipeline for Risk Factor Identification in High-dimensional Epidemiological Data with Extremely Low Events Per Variable","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-01 14:22:40","doi":"10.21203/rs.3.rs-7741957/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":"37e0a532-7731-4261-8c36-e8180a8438e7","owner":[],"postedDate":"October 1st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":55506098,"name":"Statistical Epidemiology"}],"tags":[],"updatedAt":"2025-10-01T14:22:40+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-01 14:22:40","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7741957","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7741957","identity":"rs-7741957","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}