Implicit Bias in ICU Electronic Health Record Data Measurement Frequencies and Missingness Rates of Clinical Variables

Implicit Bias in ICU Electronic Health Record Data Measurement Frequencies and Missingness Rates of Clinical Variables | 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 Implicit Bias in ICU Electronic Health Record Data Measurement Frequencies and Missingness Rates of Clinical Variables Junming Seraphina Shi, Alan E. Hubbard, Nicholas Fong, Romain Pirracchio This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5362869/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 01 Jul, 2025 Read the published version in BMC Medical Informatics and Decision Making → Version 1 posted 9 You are reading this latest preprint version Abstract Background: Disparities in data collection within electronic health records (EHRs), especially in Intensive Care Units (ICUs), can reveal underlying biases that may affect patient outcomes. Identifying and mitigating these biases is critical for ensuring equitable healthcare. This study aims to develop an analytical framework for measurement patterns, including missingness rates and measurement frequencies, evaluate the association between them and demographic factors, and assess their impact on in-hospital mortality prediction. Methods: We conducted a retrospective cohort study using the Medical Information Mart for Intensive Care III (MIMIC-III) database, which includes data on over 40,000 ICU patients from Beth Israel Deaconess Medical Center (2001–2012). Adult patients with ICU stays longer than 24 hours were included. Measurement patterns, such as missingnessrates and measurement frequencies, were derived from EHR data and analyzed. Targeted Machine Learning (TML) methods were used to assess potential biases in measurement patterns across demographic factors (age, gender, race/ethnicity) while controlling for confounders such as other demographics and disease severity. The predictive power of measurement patterns on in-hospital mortality was evaluated. Results: Among 23,426 patients, significant demographic disparities were observed in the first 24 hours of ICU stays. Elderly patients (≥ 65 years) had more frequent temperature measurements compared to younger patients, while males had slightly fewer missing temperature measurements than females. Racial disparities were notable: White patients had more frequent blood pressure and oxygen saturation (SpO2) measurements compared to Black and Hispanic patients. Measurement patterns were associated with ICU mortality, with models based solely on these patterns achieving an area under the receiver operating characteristic curve (AUC) of 0.76 (95% CI: 0.74–0.77). Conclusions: This study underscores the significance of measurement patterns in ICU EHR data, which are associated with patient demographics and ICU mortality. Analyzing patterns of missing data and measurement frequencies provides valuable insights into patient monitoring practices and potential biases in healthcare delivery. Understanding these disparities is critical for improving the fairness of healthcare delivery and developing more accurate predictive models in critical care settings. Bias Detection Critical Care Data Completeness Electronic Health Records (EHRs) Health Equity Healthcare Applications Implicit Bias Measurement Frequency MIMIC-III Dataset Missing Data Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction The digitization of health records through Electronic Health Records (EHRs) has supplanted traditional paper-based systems, thereby centralizing patient-specific information in an electronic medium. Over the past decade, several de-identified EHR datasets were made available to the public, mainly for research purposes. Notable examples include the MIMIC Database [ 15 ], PCORnet [ 7 ], I2B2 Data [ 30 ], and the COVID-19 Research Database. The advent of large-scale, accessible EHR databases has led to a surge in studies to improve healthcare delivery by identifying patient phenotypes [ 26 ], developing risk prediction models [ 13 ], or enriching our understanding of risk factors in relation to health outcomes [ 1 ]. These real-world datasets not only reflect how care is delivered but also offer valuable insights into measurement patterns, specifically how often variables are measured or missing. Previous research [ 2 – 4 , 17 , 34 ] has emphasized the necessity of addressing the issue of missing values and has recognized that incomplete data is often related to clinical practice, rather than missing at random. Common methods to handle missing data include imputation techniques [ 25 ], Inverse Probability Weighting (IPW) [ 24 ], and ignoring missing observations through complete case analysis (CCA) or available case analysis (ACA). While these methods are widely used, they have drawbacks, including the risk of compromised generalizability and the potential for increased bias due to their implicit modeling assumptions. Furthermore, they overlook the informative potential of missing data. This becomes especially problematic when data are not missing at random, as has been observed in Intensive Care Unit (ICU) settings [ 18 , 35 ]. Similarly, the concept of measurement frequency is often overlooked in studies using real-world EHR data, despite its critical importance, especially in the ICU [ 14 ]. Measurement frequency refers to how often data points, such as vital signs or laboratory values, are collected. The frequency of vitals collected can vary, and incorporating this variability can improve the performance of outcome prediction models [ 36 ]. However, this measurement frequency is rarely explicitly explored and considered in EHR-based studies. In this study, we hypothesize that, for both missingness rates and measurement frequencies, the underlying mechanism may not be random and potentially consistent with implicit bias. We introduce an analytical framework designed to analyze the occurrences of missing data and the frequency of sampling in Electronic Health Records (EHRs). Specifically, we propose (i.) to investigate the association between demographic variables, such as age, gender, and race/ethnicity, and both missing data and variability in measurement patterns of key clinical and biological parameters in critical care patients; (ii.) to study how variations in measurement patterns are predictive of in-hospital mortality. Methods Data Source The data used in this study were sourced from the Medical Information Mart for Intensive Care III (MIMIC-III) [ 15 ], the most frequently used ICU EHR dataset and one of the very few publicly available datasets where protected attributes, such as demographic variables, are collected and documented [ 8 ]. MIMIC-III is a cohort of more than 40,000 de-identified patients who were admitted to ICUs at Beth Israel Deaconess Medical Center in Boston, Massachusetts, from 2001 to 2012, and is publicly accessible through PhysioNet [ 12 ]. MIMIC-III contains patient-level healthcare information, including demographics, hospital mortality, diagnostic data, laboratory tests, prescriptions, and medical procedures. Physiological signals (e.g., heart rate, blood pressure, oxygen saturation) are captured directly from bedside monitors and integrated into the EHR system, reducing variability in data collection. We included the first ICU admission of patients who met the following criteria: (1) 18 years or older; (2) no ICU admissions related to pregnancy, childbirth, or postpartum; (3) no live discharge within the initial 24-hour period following admission; and (4) presence of at least one arterial line during the stay, ensuring continuous availability of vital signs [ 27 ]. Variables and Data Structure We extracted data on 11 vital signs and 35 laboratory tests, selected from the top 80% of the most commonly performed tests [ 9 ] (Appendix A, Table 3), as well as baseline characteristics and severity scores for each ICU stay [ 16 ]. Data was extracted from the first five days of ICU stays, and segmented into consecutive 12-hour intervals. Further details on data structure are available in Appendix B. Since some vital signs are often measured together (e.g., systolic, diastolic, and mean blood pressures), we grouped them and consolidated their missingness rate and measurement frequency into average rates. Similarly, for laboratory tests that are ordered together (e.g., complete blood count), we grouped them and consolidated their measurement rates into single variables that reflect group averages, as detailed in Appendix C, Table 5. Statistical Analysis Measurement patterns and Demographic variables Two distinct types of measurement pattern variables were used: measurement frequency (total number of observations per variable during the first 24 hours or subsequent 12-hour blocks) and missingness rate (hours without an observation per variable quantifying the frequency of missing observations). Since vital signs are measured regularly throughout the ICU stay, unlike lab tests, which are ordered as needed after ICU admission, the missingness rates are only applied to vital signs. For the first 24-hour block, this rate can range from 0 to 24, and for the 12-hour blocks, from 0 to 12. Detailed methods for calculating measurement patterns are described in Appendix C. The demographic variables considered in our study were: age groups, gender, race/ethnicity, insurance status, language, and religion. Association of demographic variables and measurement rates To estimate the association between demographic variables and measurement patterns, we used the double robust Targeted Maximum Likelihood Estimator (TMLE) [ 32 ]. This nonparametric method provides efficient and robust estimates of causal parameters, while rigorously addressing high-dimensional intermediating and confounding covariates. Given the unknown data distribution inherent in observational studies, we utilized TMLE in combination with a Super Learner (SL) algorithm. The SL algorithm applies cross-validation to select the best-performing model from an ensemble of machine learning algorithms, maximizing the ability to account for associations between outcomes and covariates, thus isolating the independent effect of demographic factors. To account for intra-patient correlation stemming from multiple ICU admissions, we implemented stratified cross-validation. In the TMLE analysis, the measurement pattern variable served as the outcome. We assessed the association between each demographic variable and the outcome, while adjusting for other demographic factors and clinical severity scores, such as the Sequential Organ Failure Assessment (SOFA) score. For example, we estimated the number of heart rate measurements taken over the initial 24 hours, comparing groups like Hispanic and White, while accounting for potential confounding from severity scores and other demographic variables. Furthermore, we conducted several sensitivity analyses utilizing more traditional regression methods, including generalized linear models (GLMs), to assess the robustness of our findings. These analyses examined both unadjusted and adjusted coefficients for the association between demographic variables and measurement pattern variables across different model specifications. Predictive power analysis We investigated the relationship between measurement patterns and patient outcomes by predicting ICU mortality within the next 12 hours and evaluating their predictive capacities. Using the SL algorithm [ 31 ], we compared the predictive performance of three models: (1) based solely on clinical values (e.g., baseline characteristics, vital sign values, and laboratory test results), (2) based solely on measurement patterns including measurement frequencies and missingness rates, and (3) a combined model using both clinical values and measurement patterns. We implemented 10-fold cross-validation, stratified by patient ID to account for repeated measures and avoid overlap between training and validation sets. The SL library included models such as Generalized Linear Models (GLM), Bayesian GLM, Generalized Additive Models (GAM), Ridge Regression, ElasticNet, Lasso Regression, Random Forests, Gradient Boosting Machines, and Bayesian Additive Regression Trees. In models (1) and (3), we imputed missing values using the median for continuous variables and the mode for categorical variables. Predictive performance was assessed using the area under the receiver operating characteristic curve (AUC), comparing the effectiveness of each model in predicting ICU mortality. All analyses were performed using the R software version 4.3.1 (2023-06-16), utilizing the sl3 (SuperLearner) and tmle3 (TMLE) packages [ 5 ]. Results Patient Demographics Of 46,520 patients in the MIMIC-III database, 23,426 met our inclusion criteria for the study. Of these, 464 patients (2.02%) experienced mortality within the first 5 days of their ICU stay. Patients’ characteristics are summarized in Table 1 . Table 1 Demographic breakdown from the datasets used in the MIMIC-III. Demographics No ICU Mortality (N = 22962) ICU Mortality (N = 464) Overall (N = 23426) Age Mean (SD) 64.8 (17.2) 67.0 (18.2) 64.8 (17.2) Median (IQR) 66.0 (23.0) 69.0 (25.0) 66.0 (23.0) Gender Female 9531 (41.5%) 210 (45.3%) 9741 (41.6%) Male 13431 (58.5%) 254 (54.7%) 13685 (58.4%) Religion Christian 11671 (50.8%) 198 (42.7%) 11869 (50.7%) Not Christian 3509 (15.3%) 56 (12.1%) 3565 (15.2%) Missing 7782 (33.9%) 210 (45.3%) 7992 (34.1%) Partner Partner 11576 (50.4%) 208 (44.8%) 11784 (50.3%) No Partner 9869 (43.0%) 180 (38.8%) 10049 (42.9%) Missing 1517 (6.6%) 76 (16.4%) 1593 (6.8%) Language English 11798 (51.4%) 206 (44.4%) 12004 (51.2%) Not English 1879 (8.2%) 46 (9.9%) 1925 (8.2%) Missing 9285 (40.4%) 212 (45.7%) 9497 (40.5%) Ethnicity White 16300 (71.0%) 306 (65.9%) 16606 (70.9%) Black 1487 (6.5%) 25 (5.4%) 1512 (6.5%) Hispanic 674 (2.9%) 17 (3.7%) 691 (2.9%) Asian 497 (2.2%) 14 (3.0%) 511 (2.2%) Other 653 (2.8%) 13 (2.8%) 666 (2.8%) Missing 3351 (14.6%) 89 (19.2%) 3440 (14.7%) Insurance Yes 22716 (98.9%) 449 (96.8%) 23165 (98.9%) No 246 (1.1%) 15 (3.2%) 261 (1.1%) Table 2 Performance Metrics of Random Forest Models Across Different Variable Combinations. This table summarizes the average Negative Log-Likelihood (NLL), Area Under the ROC Curve (AUC), and Area Under the Precision-Recall Curve (AUCPR) for Random Forest models trained with different variable sets, based on 10-fold cross-validation. NLL AUC AUCPR Values variables 0.17 (0.16, 0.19) 0.85 (0.83, 0.87) 0.18 (0.16, 0.20) Pattern variables 0.21 (0.18, 0.23) 0.76 (0.74, 0.77) 0.10 (0.09, 0.11) Value and pattern variables 0.17 (0.15, 0.19) 0.86 (0.84, 0.87) 0.19 (0.17, 0.22) Differential Measurement Pattern Our analysis first focused on the initial 24 hours of ICU data, comprising 24,517 ICU stays. After adjusting for severity scores and demographic variables, statistically significant differences in measurement patterns were observed across different age groups (Fig. 1 ). Elderly patients, particularly those aged 46–65 and over 65, received more frequent vital sign measurements and had fewer missing measurements compared to younger patients, except for the Glasgow Coma Scale (GCS). For example, the adjusted average number of temperature measurements in patients aged 65 and older was 12.2 (95% CI: 12.0 to 12.4), with 14.72 hours (95% CI: 14.61 to 14.83) of adjusted average missingness rate. In contrast, patients aged 18–30 had 9.7 adjusted average temperature measurements (95% CI: 9.4 to 10.1) and 15.45 adjusted average hours of missing measurements (95% CI: 15.21 to 15.69) (Fig. 1 ). However, laboratory test frequencies decreased linearly with increasing age, a trend consistent across all tests, even after adjusting for SOFA scores and other covariates (Appendix D.1). From gender-based analysis, males had approximately 0.2 more hours of missing vital sign measurements, although no significant differences in measurement frequencies were noted. An exception was temperature monitoring, where males received marginally more checks (12.22, 95% CI: 12.03 to 12.40) and had 0.4 fewer hours of missing data (14.69 hours, 95% CI: 14.55 to 14.85) compared to females, who had on average 11.26 measurements (95% CI: 11.05 to 11.50) and 15.13 hours (95% CI: 15.02 to 15.25) of missing measurements (Fig. 2 ). While there were some differences in laboratory test frequencies by gender, these were inconsistent across different test groups (Appendix D.2). Racial and ethnic differences were also observed. Black and Hispanic patients had fewer vital sign measurements compared to White patients, although no significant differences were seen in the rate of missing data. For instance, the adjusted average number of blood pressure measurements within the first 24 hours was 34.22 (95% CI: 34.00 to 34.44) for White patients, compared to 32.79 (95% CI: 32.11 to 33.46) for Black patients and 32.39 (95% CI: 31.82 to 32.98) for Hispanic patients (p-values < 0.001 for the comparisons between White and Black, and White and Hispanic) (Fig. 3 ). Notable differences were also observed in SpO2 and temperature, both in measurement frequencies and missingness rates. For SpO2, the adjusted average number of measurements for White patients was 30.64 (95% CI: 30.48 to 30.81) with 2.60 hours (95% CI: 2.55 to 2.65) of adjusted average missingness rates, compared to 28.83 measurements (95% CI: 28.32 to 29.34) and 3.09 hours (95% CI: 2.85 to 3.32) for Black patients. Similar trends were seen for temperature measurements: White patients had an average of 11.70 measurements (95% CI: 11.54 to 11.86) and 14.90 hours (95% CI: 14.79 to 15.01) of adjusted average missingness rates, compared to 9.65 measurements (95% CI: 9.21 to 10.08) and 15.82 hours (95% CI: 14.54 to 16.09) for Black patients. Laboratory test patterns did not show consistent disparities across racial/ethnic groups, although hematocrit tests revealed small but statistically significant differences (2.75 [95% CI: 2.73 to 2.78] for White patients vs. 2.49 [95% CI: 2.40 to 2.57] for Black patients, p < 0.001) (Fig. 3 ). Other demographic factors, including insurance status, language, marital status, and religion, did not exhibit consistent patterns in measurement frequency (Appendix D). Sensitivity analyses confirmed that the patterns observed in the primary analysis were robust, with no notable differences or opposing trends (Appendix F). Measurement patterns predictive of mortality The 12-hour mortality prediction model trained on both count variables (measurement frequency and missingness rate) and original value variables demonstrated the highest predictive accuracy, with an AUC of 0.86 (95% CI: 0.84 to 0.87) (Fig. 4 , Appendix E). This performance was comparable to the model trained solely on original value variables, which had an AUC of 0.85 (95% CI: 0.83 to 0.87). No significant difference was observed between these two models. The model trained exclusively on measurement frequency and missingness rate variables, while still showing notable predictive accuracy, registered a significantly lower AUC of 0.76 (0.74, 0.77) compared to the other two models. Discussion This study offers a novel perspective on measurement patterns within EHR for ICU patients. Even after aggressively adjusting for baseline clinical health measures, we found statistically significant disparities in the frequency of vital sign and laboratory test measurements and missing data rates across different demographic groups during the first 24 hours of ICU admission. Older patients received more frequent measurements with fewer gaps in data, except for GCS measurements. Males experienced slightly more frequent temperature checks than females. Racial disparities were present, with Black and Hispanic patients having significantly fewer vital sign measurements compared to White patients. Furthermore, laboratory test frequency decreased linearly with age, with some differences in specific tests like hematocrit, though no consistent disparities were found across other demographic categories. Additional factors such as insurance status, language, marital status, religion, and care unit were explored, but no strong patterns were observed. Our analysis also demonstrated that measurement patterns including measurement frequencies and missingness rates are strongly predictive of patient outcomes. We contribute to the growing body of literature highlighting biases in healthcare data analytics. Landmark studies, such as [ 20 ] and [ 29 ], showed significant biases in medical algorithms and healthcare data, emphasizing the need for equitable data representation. Our findings align with previous research indicating the possible presence of biases in EHR data collection [ 21 , 33 ]. Our focus on patterns of missingness rates and measurement frequencies, their associations with demographics, and their predictive power offers deeper insights into the impact of measurement variability, emphasizing the importance of considering these factors when analyzing and modeling data from critical care patients. Previous studies [ 22 , 13 ] have explored the implications of missing data and sampling biases in predictive modeling, highlighting the potential for these factors to skew results and perpetuate inequalities. Our research builds on these findings by specifically analyzing the demographic disparities in missingness rates and measurement frequencies. We found statistically significant disparities in these patterns between demographic groups and demonstrated the potential to improve the accuracy of the predictive model by including these disparities in the training data. Based on our results, we caution against relying solely on advanced imputation techniques without first understanding the underlying measurement patterns. Imputation methods should be applied with careful consideration of demographic characteristics to ensure fair representation. Specifically, advanced imputation techniques such as multiple imputation by chained equations (MICE) and deep learning-based imputation methods can be adapted to account for the demographic factors that influence data collection [ 23 , 28 ]. We also recommend incorporating measurement patterns into statistical and machine learning models to mitigate the impact of varied measurement patterns. This can be achieved by weighting the samples and observations based on their measurement frequencies, modeling the data generation process as a multilevel model where the measurement pattern is treated as a latent variable, or employing a Bayesian model to estimate the impact of different measurement frequencies on model outcomes, thus providing a probabilistic framework to handle uncertainty [ 19 , 11 , 10 , 22 ]. Finally, promoting transparency in data collection processes and conducting regular audits of EHR systems to detect bias can help create a more equitable healthcare data landscape. Our comparison with previously published studies underscores the continued need for methodological advancements to tackle these complex challenges and ensure that healthcare analytics are fair and accurate [ 20 , 29 ]. One of the primary strengths of our study is its innovative approach to analyzing EHR data, particularly in identifying and interpreting biases in missing data and measurement frequencies. This approach not only offers a framework for future research but also reveals important disparities in healthcare practices. However, limitations include the potential limited generalizability of our findings, as our study was confined to ICU data coming from the MIMI-III database. Additionally, the study's retrospective nature may limit the ability to infer causality between data collection practices and patient outcomes. Conclusions In conclusion, this study emphasizes the significance of measurement patterns in ICU EHR data, which are linked to patient demographics and ICU mortality. Analyzing missing data patterns and measurement frequencies provides valuable insights into patient monitoring practices and healthcare equity. Abbreviations ICU Intensive Care Unit MIMIC-III Medical Information Mart for Intensive Care III EHR Electronic Health Record GCS Glasgow Coma Scale SOFA Sequential Organ Failure Assessment TMLE Targeted Maximum Likelihood Estimator SL Super Learner GLM Generalized Linear Models AUC area under the receiver operating characteristic curve MICE multiple imputation by chained equations Declarations Ethics approval and consent to participate: Not applicable Clinical trial number: Not applicable Consent for publication: Not applicable Availability of data and materials: The data that support the findings of this study are derived from the MIMIC-III database [15], which is publicly available but subject to specific access requirements. Access to MIMIC-III data requires users to complete a data use agreement and undergo a certification process, which includes completing the required training in human subjects research (CITI Program). While the data used in this study were obtained under this agreement and are not publicly available due to licensing restrictions, they may be made available by the authors upon reasonable request, contingent on approval from the MIMIC-III data custodians and adherence to the relevant data use agreements Competing interests: Not applicable Funding: Not applicable Authors' contributions: JS developed the analytical framework for assessing measurement patterns in the ICU, analyzed the measurement patterns and their relationships with demographic variables, assessed the predictive power of these patterns, interpreted the results, and was the primary contributor to writing the manuscript. NF assisted with data extraction and preprocessing. AH and RP provided oversight and guidance throughout the project. All authors read, revised, and approved the final manuscript. Acknowledgements: Not applicable References Adler NE, Stead WW. Patients in context—EHR capture of social and behavioral determinants of health. Obstet Gynecol Surv. 2015;70(6):388–90. Beaulieu-Jones BK, Lavage DR, Snyder JW, Moore JH, Pendergrass SA, Bauer CR. Characterizing and managing missing structured data in electronic health records: data analysis. JMIR Med Inf. 2018;6(1):e8960. Beaulieu-Jones BK, Moore JH, PIVOTAL-ACT CONSORTIUM. Missing data imputation in the electronic health record using deeply learned autoencoders. In: Pacific Symposium on Biocomputing 2017. Singapore: World Scientific; 2017. pp. 207 – 18. Botsis T, Hartvigsen G, Chen F, Weng C. Secondary use of EHR: data quality issues and informatics opportunities. Summit Transl Bioinform. 2010;2010:1. Coyle JR. tmle3: The extensible TMLE framework [Internet]. 2021. https://github.com/tlverse/tmle3 . Accessed 2021. Ezzie ME, Aberegg SK, O’Brien JM Jr.. Laboratory testing in the intensive care unit. Crit Care Clin. 2007;23(3):435–65. Fleurence RL, Curtis LH, Califf RM, Platt R, Selby JV, Brown JS. Launching PCORnet, a national patient-centered clinical research network. J Am Med Inf Assoc. 2014;21(4):578–82. Fong N, Langnas E, Law T, Reddy M, Lipnick M, Pirracchio R. Availability of information needed to evaluate algorithmic fairness—a systematic review of publicly accessible critical care databases. Anaesth Crit Care Pain Med. 2023;101248. Frassica JJ. Frequency of laboratory test utilization in the intensive care unit and its implications for large-scale data collection efforts. J Am Med Inf Assoc. 2005;12(2):229–33. Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian data analysis. Boca Raton: Chapman and Hall/CRC; 1995. Gelman A, Hill J. Data analysis using regression and multilevel/hierarchical models. Cambridge: Cambridge University Press; 2006. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. Physiobank, Physiotoolkit, and Physionet: components of a new research resource for complex physiologic signals. Circulation. 2000;101(23):e215–20. Goldstein BA, Navar AM, Pencina MJ, Ioannidis JP. Opportunities and challenges in developing risk prediction models with electronic health records data: a systematic review. J Am Med Inf Assoc. 2017;24(1):198–208. Hyland SL, Faltys M, Hüser M, Lyu X, Gumbsch T, Esteban C, et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nat Med. 2020;26(3):364–73. Johnson AE, Pollard TJ, Shen L, Lehman LW, Feng M, Ghassemi M, et al. MIMIC-III, a freely accessible critical care database. Sci Data. 2016;3(1):1–9. Johnson AE, Stone DJ, Celi LA, Pollard TJ. The MIMIC code repository: enabling reproducibility in critical care research. J Am Med Inf Assoc. 2018;25(1):32–9. Kharrazi H, Wang C, Scharfstein D. Prospective EHR-based clinical trials: the challenge of missing data. J Am Med Inf Assoc. 2014. Khope SR, Elias S. Strategies of predictive schemes and clinical diagnosis for prognosis using MIMIC-III: a systematic review. Healthc. 2023;11(710). Little RJ, Rubin DB. Statistical analysis with missing data. 3rd ed. Hoboken: Wiley; 2019. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447–53. Pivovarov R, Albers DJ, Sepulveda JL, Elhadad N. Identifying and mitigating biases in EHR laboratory tests. J Biomed Inf. 2014;51:24–34. Rusanov A, Weiskopf NG, Wang S, Weng C. Hidden in plain sight: bias towards sick patients when sampling patients with sufficient electronic health record data for research. BMC Med Inf Decis Mak. 2014;14:1–9. Samad MD, Abrar S, Diawara N. Missing value estimation using clustering and deep learning within multiple imputation framework. Knowl Based Syst. 2022;249:108968. Seaman SR, White IR. Review of inverse probability weighting for dealing with missing data. Stat Methods Med Res. 2013;22(3):278–95. Sharafoddini A, Dubin JA, Maslove DM, Lee J. A new insight into missing data in intensive care unit patient profiles: observational study. JMIR Med Inf. 2019;7(1):e11605. Shivade C, Raghavan P, Fosler-Lussier E, Embi PJ, Elhadad N, Johnson SB, et al. A review of approaches to identifying patient phenotype cohorts using electronic health records. J Am Med Inf Assoc. 2014;21(2):221–30. Shiver S, Blaivas M, Lyon M. A prospective comparison of ultrasound-guided and blindly placed radial arterial catheters. Acad Emerg Med. 2006;13(12):1275–9. Sun Y, Li J, Xu Y, Zhang T, Wang X. Deep learning versus conventional methods for missing data imputation: a review and comparative study. Expert Syst Appl. 2023;120201. Suresh H, Guttag JV. A framework for understanding unintended consequences of machine learning. arXiv preprint arXiv:1901 10002. 2019;2(8):73. Uzuner Ö, South BR, Shen S, DuVall SL. 2010 i2b2/VA challenge on concepts, assertions, and relations in clinical text. J Am Med Inform Assoc. 2011;18(5):552-6. n der Laan MJ, Polley EC, Hubbard AE. Super learner. Stat Appl Genet Mol Biol. 2007;6(1):1–29. n der Laan MJ, Rose S. Targeted learning in data science. New York: Springer; 2018. Verheij RA, Curcin V, Delaney BC, McGilchrist MM. Possible sources of bias in primary care electronic health record data use and reuse. J Med Internet Res. 2018;20(5):e185. Wells BJ, Chagin KM, Nowacki AS, Kattan MW. Strategies for handling missing data in electronic health record derived data. EGEMS (Wash DC). 2013;1(3):1–8. Zhang D, Yin C, Zeng J, Yuan X, Zhang P. Combining structured and unstructured data for predictive models: a deep learning approach. BMC Med Inf Decis Mak. 2020;20(1):1–11. Zhang L, Chen X, Chen T, Wang Z, Mortazavi BJ. Dynehr: dynamic adaptation of models with data heterogeneity in electronic health records. In: 2021 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI). 2021. pp. 1–4. Additional Declarations No competing interests reported. Supplementary Files Appendix.pdf Cite Share Download PDF Status: Published Journal Publication published 01 Jul, 2025 Read the published version in BMC Medical Informatics and Decision Making → Version 1 posted Editorial decision: Revision requested 08 Jan, 2025 Reviews received at journal 26 Dec, 2024 Reviews received at journal 10 Dec, 2024 Reviewers agreed at journal 07 Dec, 2024 Reviewers agreed at journal 06 Dec, 2024 Reviewers invited by journal 05 Nov, 2024 Editor assigned by journal 05 Nov, 2024 Submission checks completed at journal 02 Nov, 2024 First submitted to journal 30 Oct, 2024 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-5362869","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":375057775,"identity":"a70dc301-b757-48c9-a0c9-6c44091a053e","order_by":0,"name":"Junming Seraphina Shi","email":"","orcid":"","institution":"University of California Berkeley","correspondingAuthor":false,"prefix":"","firstName":"Junming","middleName":"Seraphina","lastName":"Shi","suffix":""},{"id":375057780,"identity":"f69e19b1-6d8e-48f9-9412-3e5c993e8d67","order_by":1,"name":"Alan E. Hubbard","email":"","orcid":"","institution":"University of California Berkeley","correspondingAuthor":false,"prefix":"","firstName":"Alan","middleName":"E.","lastName":"Hubbard","suffix":""},{"id":375057781,"identity":"3aab1e53-6a41-4e96-b777-11022d746a84","order_by":2,"name":"Nicholas Fong","email":"","orcid":"","institution":"University of California, San Francisco","correspondingAuthor":false,"prefix":"","firstName":"Nicholas","middleName":"","lastName":"Fong","suffix":""},{"id":375057783,"identity":"6be9097f-a4c2-4933-bf4d-902d02ec30c9","order_by":3,"name":"Romain Pirracchio","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABD0lEQVRIiWNgGAWjYBAC+wYgkcDAUM8G4R/gYWAHCdmA2Ni1GByAaElsg2vhAQmlEdACBIkNUC0MDBIJBLQc70788HAHQ2Kf9NmHH37U3JHhl3xj+PFHAoMc340E7H7pObtZIvEM0C986caSPcee8UjOzjGW5klgMJbEocVOIneDBNAjCWw8bAzSjA2HeQxu55gxM/5gSNyAQ4uxRO7mH1AtzL/BWm6eMWMEOqwelxbDGbnbYLawQWy5wWPGAHRYggEOLQZnzm6zSGyTAGux7Dl2mEeyJ60Y6BcJw5lnHuAIsd7NN3+22STI97Ax3/hRc9ien/3wRmCI2cjzHcduCxRIECEyCkbBKBgFo4B4AAADx15XJnjL8wAAAABJRU5ErkJggg==","orcid":"","institution":"University of California, San Francisco","correspondingAuthor":true,"prefix":"","firstName":"Romain","middleName":"","lastName":"Pirracchio","suffix":""}],"badges":[],"createdAt":"2024-10-30 17:53:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5362869/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5362869/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12911-025-03058-9","type":"published","date":"2025-07-01T15:57:10+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":70044145,"identity":"a93566ab-d994-4f9e-9b58-bdd736922247","added_by":"auto","created_at":"2024-11-27 18:48:48","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":777439,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTMLE Estimated Measurement Patterns by Age Groups.\u003c/strong\u003e This figure illustrates TMLE-estimated marginal average measurement frequencies across four age groups, considering baseline characteristics and SOFA scores. It includes three vital signs and three laboratory tests, showcasing implicit biases in monitoring. Full details and additional plots are available in Appendix D.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-5362869/v1/ec86bc0b4690ffc92f7ee8ec.png"},{"id":70043677,"identity":"9f72702a-ae49-42f2-a867-e11bd0c95ef6","added_by":"auto","created_at":"2024-11-27 18:40:48","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1110479,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTMLE Estimated Measurement Patterns by Gender.\u003c/strong\u003e This figure displays TMLE-estimated measurement patterns for female and male patients, incorporating baseline characteristics and SOFA scores. It focuses on three vital signs, highlighting gender-based differences in patient monitoring.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-5362869/v1/0f1c72d216c71174e21e6e08.png"},{"id":70043676,"identity":"241c8535-6b7b-4736-bd4c-fd56177bd50d","added_by":"auto","created_at":"2024-11-27 18:40:48","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1309607,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTMLE Estimated Measurement Patterns by Race/Ethnicity.\u003c/strong\u003e Displaying TMLE-estimated marginal average measurement frequencies for different racial and ethnic groups. The top plot shows five vital signs and the bottom plot shows three laboratory tests, revealing racial and ethnic disparities in monitoring within healthcare settings.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-5362869/v1/be51c3a46d75c7c1bf4e31e0.png"},{"id":70043678,"identity":"a4ab92ac-5cfa-4d1c-818c-aacb3c76af0c","added_by":"auto","created_at":"2024-11-27 18:40:48","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":280727,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAUC Metrics in Random Forest Models Trained with Different Variable Combinations.\u003c/strong\u003e This figure compares AUC scores for three Random Forest models, each trained with a distinct set of variables: (1) ’Values solely’, including baseline characteristics, vital sign measurement values, and laboratory test values (LT); (2) ’Patterns solely’, covering counts of baseline information, vital sign measurement frequencies, missingness rates for vital signs, and laboratory test frequencies; and (3) ‘Values and Patterns’, including variables of both clinical values and measurement patterns. Models with missing values handled by imputing medians for numeric and modes for categorical variables are colored in blue; models with missing values imputed by random values within the observed range are colored in yellow.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-5362869/v1/7c32e3281eed941970471a37.png"},{"id":86178921,"identity":"bf7df117-7d3c-4190-8f8f-4e576814b72d","added_by":"auto","created_at":"2025-07-07 16:11:21","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3274974,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5362869/v1/3482b442-4f9f-4c35-9929-b5d3a8407f72.pdf"},{"id":70043680,"identity":"521e967a-5a82-4d01-9a2d-42a053f53559","added_by":"auto","created_at":"2024-11-27 18:40:48","extension":"pdf","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":2345453,"visible":true,"origin":"","legend":"","description":"","filename":"Appendix.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5362869/v1/a27e6c3136957641a95fc1a6.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Implicit Bias in ICU Electronic Health Record Data Measurement Frequencies and Missingness Rates of Clinical Variables","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe digitization of health records through Electronic Health Records (EHRs) has supplanted traditional paper-based systems, thereby centralizing patient-specific information in an electronic medium. Over the past decade, several de-identified EHR datasets were made available to the public, mainly for research purposes. Notable examples include the MIMIC Database [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], PCORnet [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], I2B2 Data [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], and the COVID-19 Research Database.\u003c/p\u003e \u003cp\u003eThe advent of large-scale, accessible EHR databases has led to a surge in studies to improve healthcare delivery by identifying patient phenotypes [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], developing risk prediction models [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], or enriching our understanding of risk factors in relation to health outcomes [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. These real-world datasets not only reflect how care is delivered but also offer valuable insights into measurement patterns, specifically how often variables are measured or missing.\u003c/p\u003e \u003cp\u003ePrevious research [\u003cspan additionalcitationids=\"CR3\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e–\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] has emphasized the necessity of addressing the issue of \u003cem\u003emissing values\u003c/em\u003e and has recognized that incomplete data is often related to clinical practice, rather than missing at random. Common methods to handle missing data include imputation techniques [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], Inverse Probability Weighting (IPW) [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e], and ignoring missing observations through complete case analysis (CCA) or available case analysis (ACA). While these methods are widely used, they have drawbacks, including the risk of compromised generalizability and the potential for increased bias due to their implicit modeling assumptions. Furthermore, they overlook the informative potential of missing data. This becomes especially problematic when data are not missing at random, as has been observed in Intensive Care Unit (ICU) settings [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSimilarly, the concept of \u003cem\u003emeasurement frequency\u003c/em\u003e is often overlooked in studies using real-world EHR data, despite its critical importance, especially in the ICU [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Measurement frequency refers to how often data points, such as vital signs or laboratory values, are collected. The frequency of vitals collected can vary, and incorporating this variability can improve the performance of outcome prediction models [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. However, this measurement frequency is rarely explicitly explored and considered in EHR-based studies.\u003c/p\u003e \u003cp\u003eIn this study, we hypothesize that, for both missingness rates and measurement frequencies, the underlying mechanism may not be random and potentially consistent with implicit bias. We introduce an analytical framework designed to analyze the occurrences of missing data and the frequency of sampling in Electronic Health Records (EHRs). Specifically, we propose (i.) to investigate the association between demographic variables, such as age, gender, and race/ethnicity, and both missing data and variability in measurement patterns of key clinical and biological parameters in critical care patients; (ii.) to study how variations in measurement patterns are predictive of in-hospital mortality.\u003c/p\u003e "},{"header":"Methods","content":"\u003cp\u003eData Source\u003c/p\u003e\u003cp\u003eThe data used in this study were sourced from the Medical Information Mart for Intensive Care III (MIMIC-III) [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], the most frequently used ICU EHR dataset and one of the very few publicly available datasets where protected attributes, such as demographic variables, are collected and documented [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMIMIC-III is a cohort of more than 40,000 de-identified patients who were admitted to ICUs at Beth Israel Deaconess Medical Center in Boston, Massachusetts, from 2001 to 2012, and is publicly accessible through PhysioNet [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. MIMIC-III contains patient-level healthcare information, including demographics, hospital mortality, diagnostic data, laboratory tests, prescriptions, and medical procedures. Physiological signals (e.g., heart rate, blood pressure, oxygen saturation) are captured directly from bedside monitors and integrated into the EHR system, reducing variability in data collection.\u003c/p\u003e\u003cp\u003eWe included the first ICU admission of patients who met the following criteria: (1) 18 years or older; (2) no ICU admissions related to pregnancy, childbirth, or postpartum; (3) no live discharge within the initial 24-hour period following admission; and (4) presence of at least one arterial line during the stay, ensuring continuous availability of vital signs [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eVariables and Data Structure\u003c/p\u003e\u003cp\u003eWe extracted data on 11 vital signs and 35 laboratory tests, selected from the top 80% of the most commonly performed tests [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] (Appendix A, Table\u0026nbsp;3), as well as baseline characteristics and severity scores for each ICU stay [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Data was extracted from the first five days of ICU stays, and segmented into consecutive 12-hour intervals. Further details on data structure are available in Appendix B.\u003c/p\u003e\u003cp\u003eSince some vital signs are often measured together (e.g., systolic, diastolic, and mean blood pressures), we grouped them and consolidated their missingness rate and measurement frequency into average rates. Similarly, for laboratory tests that are ordered together (e.g., complete blood count), we grouped them and consolidated their measurement rates into single variables that reflect group averages, as detailed in Appendix C, Table\u0026nbsp;5.\u003c/p\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eMeasurement patterns and Demographic variables\u003c/p\u003e\u003cp\u003eTwo distinct types of measurement pattern variables were used: \u003cem\u003emeasurement frequency\u003c/em\u003e (total number of observations per variable during the first 24 hours or subsequent 12-hour blocks) and \u003cem\u003emissingness rate\u003c/em\u003e (hours without an observation per variable quantifying the frequency of missing observations). Since vital signs are measured regularly throughout the ICU stay, unlike lab tests, which are ordered as needed after ICU admission, the missingness rates are only applied to vital signs. For the first 24-hour block, this rate can range from 0 to 24, and for the 12-hour blocks, from 0 to 12. Detailed methods for calculating measurement patterns are described in Appendix C.\u003c/p\u003e\u003cp\u003eThe demographic variables considered in our study were: age groups, gender, race/ethnicity, insurance status, language, and religion.\u003c/p\u003e\u003cp\u003eAssociation of demographic variables and measurement rates\u003c/p\u003e\u003cp\u003eTo estimate the association between demographic variables and measurement patterns, we used the double robust Targeted Maximum Likelihood Estimator (TMLE) [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. This nonparametric method provides efficient and robust estimates of causal parameters, while rigorously addressing high-dimensional intermediating and confounding covariates. Given the unknown data distribution inherent in observational studies, we utilized TMLE in combination with a Super Learner (SL) algorithm. The SL algorithm applies cross-validation to select the best-performing model from an ensemble of machine learning algorithms, maximizing the ability to account for associations between outcomes and covariates, thus isolating the independent effect of demographic factors. To account for intra-patient correlation stemming from multiple ICU admissions, we implemented stratified cross-validation.\u003c/p\u003e\u003cp\u003eIn the TMLE analysis, the measurement pattern variable served as the outcome. We assessed the association between each demographic variable and the outcome, while adjusting for other demographic factors and clinical severity scores, such as the Sequential Organ Failure Assessment (SOFA) score. For example, we estimated the number of heart rate measurements taken over the initial 24 hours, comparing groups like Hispanic and White, while accounting for potential confounding from severity scores and other demographic variables.\u003c/p\u003e\u003cp\u003eFurthermore, we conducted several sensitivity analyses utilizing more traditional regression methods, including generalized linear models (GLMs), to assess the robustness of our findings. These analyses examined both unadjusted and adjusted coefficients for the association between demographic variables and measurement pattern variables across different model specifications.\u003c/p\u003e\u003cp\u003ePredictive power analysis\u003c/p\u003e\u003cp\u003eWe investigated the relationship between measurement patterns and patient outcomes by predicting ICU mortality within the next 12 hours and evaluating their predictive capacities. Using the SL algorithm [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], we compared the predictive performance of three models: (1) based solely on clinical values (e.g., baseline characteristics, vital sign values, and laboratory test results), (2) based solely on measurement patterns including measurement frequencies and missingness rates, and (3) a combined model using both clinical values and measurement patterns. We implemented 10-fold cross-validation, stratified by patient ID to account for repeated measures and avoid overlap between training and validation sets. The SL library included models such as Generalized Linear Models (GLM), Bayesian GLM, Generalized Additive Models (GAM), Ridge Regression, ElasticNet, Lasso Regression, Random Forests, Gradient Boosting Machines, and Bayesian Additive Regression Trees.\u003c/p\u003e\u003cp\u003eIn models (1) and (3), we imputed missing values using the median for continuous variables and the mode for categorical variables. Predictive performance was assessed using the area under the receiver operating characteristic curve (AUC), comparing the effectiveness of each model in predicting ICU mortality.\u003c/p\u003e\u003cp\u003eAll analyses were performed using the R software version 4.3.1 (2023-06-16), utilizing the sl3 (SuperLearner) and tmle3 (TMLE) packages [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003ePatient Demographics\u003c/p\u003e \u003cp\u003eOf 46,520 patients in the MIMIC-III database, 23,426 met our inclusion criteria for the study. Of these, 464 patients (2.02%) experienced mortality within the first 5 days of their ICU stay. Patients\u0026rsquo; characteristics are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDemographic breakdown from the datasets used in the MIMIC-III.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e Demographics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo ICU Mortality\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;22962)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eICU Mortality\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;464)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOverall\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;23426)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAge\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e64.8 (17.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e67.0 (18.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e64.8 (17.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMedian (IQR)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e66.0 (23.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e69.0 (25.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e66.0 (23.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eGender\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9531 (41.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e210 (45.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9741 (41.6%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e13431 (58.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e254 (54.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13685 (58.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eReligion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eChristian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11671 (50.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e198 (42.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11869 (50.7%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNot Christian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3509 (15.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e56 (12.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3565 (15.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMissing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e7782 (33.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e210 (45.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e7992 (34.1%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003ePartner\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePartner\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11576 (50.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e208 (44.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11784 (50.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNo Partner\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9869 (43.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e180 (38.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e10049 (42.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMissing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1517 (6.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e76 (16.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1593 (6.8%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eLanguage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEnglish\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11798 (51.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e206 (44.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12004 (51.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNot English\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1879 (8.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e46 (9.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1925 (8.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMissing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9285 (40.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e212 (45.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9497 (40.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eEthnicity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWhite\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16300 (71.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e306 (65.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16606 (70.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBlack\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1487 (6.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e25 (5.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1512 (6.5%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHispanic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e674 (2.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e17 (3.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e691 (2.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAsian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e497 (2.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e14 (3.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e511 (2.2%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOther\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e653 (2.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e13 (2.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e666 (2.8%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMissing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3351 (14.6%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89 (19.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3440 (14.7%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eInsurance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e22716 (98.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e449 (96.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e23165 (98.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e246 (1.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e15 (3.2%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e261 (1.1%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003ePerformance Metrics of Random Forest Models Across Different Variable Combinations.\u003c/b\u003e This table summarizes the average Negative Log-Likelihood (NLL), Area Under the ROC Curve (AUC), and Area Under the Precision-Recall Curve (AUCPR) for Random Forest models trained with different variable sets, based on 10-fold cross-validation.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNLL\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUCPR\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValues variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.17 (0.16, 0.19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.85 (0.83, 0.87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.18 (0.16, 0.20)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePattern variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.21 (0.18, 0.23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.76 (0.74, 0.77)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.10 (0.09, 0.11)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValue and pattern variables\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.17 (0.15, 0.19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.86 (0.84, 0.87)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.19 (0.17, 0.22)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eDifferential Measurement Pattern\u003c/p\u003e \u003cp\u003eOur analysis first focused on the initial 24 hours of ICU data, comprising 24,517 ICU stays. After adjusting for severity scores and demographic variables, statistically significant differences in measurement patterns were observed across different age groups (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Elderly patients, particularly those aged 46\u0026ndash;65 and over 65, received more frequent vital sign measurements and had fewer missing measurements compared to younger patients, except for the Glasgow Coma Scale (GCS). For example, the adjusted average number of temperature measurements in patients aged 65 and older was 12.2 (95% CI: 12.0 to 12.4), with 14.72 hours (95% CI: 14.61 to 14.83) of adjusted average missingness rate. In contrast, patients aged 18\u0026ndash;30 had 9.7 adjusted average temperature measurements (95% CI: 9.4 to 10.1) and 15.45 adjusted average hours of missing measurements (95% CI: 15.21 to 15.69) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). However, laboratory test frequencies decreased linearly with increasing age, a trend consistent across all tests, even after adjusting for SOFA scores and other covariates (Appendix D.1).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFrom gender-based analysis, males had approximately 0.2 more hours of missing vital sign measurements, although no significant differences in measurement frequencies were noted. An exception was temperature monitoring, where males received marginally more checks (12.22, 95% CI: 12.03 to 12.40) and had 0.4 fewer hours of missing data (14.69 hours, 95% CI: 14.55 to 14.85) compared to females, who had on average 11.26 measurements (95% CI: 11.05 to 11.50) and 15.13 hours (95% CI: 15.02 to 15.25) of missing measurements (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). While there were some differences in laboratory test frequencies by gender, these were inconsistent across different test groups (Appendix D.2).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eRacial and ethnic differences were also observed. Black and Hispanic patients had fewer vital sign measurements compared to White patients, although no significant differences were seen in the rate of missing data. For instance, the adjusted average number of blood pressure measurements within the first 24 hours was 34.22 (95% CI: 34.00 to 34.44) for White patients, compared to 32.79 (95% CI: 32.11 to 33.46) for Black patients and 32.39 (95% CI: 31.82 to 32.98) for Hispanic patients (p-values\u0026thinsp;\u0026lt;\u0026thinsp;0.001 for the comparisons between White and Black, and White and Hispanic) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Notable differences were also observed in SpO2 and temperature, both in measurement frequencies and missingness rates. For SpO2, the adjusted average number of measurements for White patients was 30.64 (95% CI: 30.48 to 30.81) with 2.60 hours (95% CI: 2.55 to 2.65) of adjusted average missingness rates, compared to 28.83 measurements (95% CI: 28.32 to 29.34) and 3.09 hours (95% CI: 2.85 to 3.32) for Black patients. Similar trends were seen for temperature measurements: White patients had an average of 11.70 measurements (95% CI: 11.54 to 11.86) and 14.90 hours (95% CI: 14.79 to 15.01) of adjusted average missingness rates, compared to 9.65 measurements (95% CI: 9.21 to 10.08) and 15.82 hours (95% CI: 14.54 to 16.09) for Black patients. Laboratory test patterns did not show consistent disparities across racial/ethnic groups, although hematocrit tests revealed small but statistically significant differences (2.75 [95% CI: 2.73 to 2.78] for White patients vs. 2.49 [95% CI: 2.40 to 2.57] for Black patients, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eOther demographic factors, including insurance status, language, marital status, and religion, did not exhibit consistent patterns in measurement frequency (Appendix D). Sensitivity analyses confirmed that the patterns observed in the primary analysis were robust, with no notable differences or opposing trends (Appendix F).\u003c/p\u003e \u003cp\u003eMeasurement patterns predictive of mortality\u003c/p\u003e \u003cp\u003eThe 12-hour mortality prediction model trained on both count variables (measurement frequency and missingness rate) and original value variables demonstrated the highest predictive accuracy, with an AUC of 0.86 (95% CI: 0.84 to 0.87) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, Appendix E). This performance was comparable to the model trained solely on original value variables, which had an AUC of 0.85 (95% CI: 0.83 to 0.87). No significant difference was observed between these two models. The model trained exclusively on measurement frequency and missingness rate variables, while still showing notable predictive accuracy, registered a significantly lower AUC of 0.76 (0.74, 0.77) compared to the other two models.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study offers a novel perspective on measurement patterns within EHR for ICU patients. Even after aggressively adjusting for baseline clinical health measures, we found statistically significant disparities in the frequency of vital sign and laboratory test measurements and missing data rates across different demographic groups during the first 24 hours of ICU admission. Older patients received more frequent measurements with fewer gaps in data, except for GCS measurements. Males experienced slightly more frequent temperature checks than females. Racial disparities were present, with Black and Hispanic patients having significantly fewer vital sign measurements compared to White patients. Furthermore, laboratory test frequency decreased linearly with age, with some differences in specific tests like hematocrit, though no consistent disparities were found across other demographic categories. Additional factors such as insurance status, language, marital status, religion, and care unit were explored, but no strong patterns were observed.\u003c/p\u003e \u003cp\u003eOur analysis also demonstrated that measurement patterns including measurement frequencies and missingness rates are strongly predictive of patient outcomes. We contribute to the growing body of literature highlighting biases in healthcare data analytics. Landmark studies, such as [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], showed significant biases in medical algorithms and healthcare data, emphasizing the need for equitable data representation. Our findings align with previous research indicating the possible presence of biases in EHR data collection [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Our focus on patterns of missingness rates and measurement frequencies, their associations with demographics, and their predictive power offers deeper insights into the impact of measurement variability, emphasizing the importance of considering these factors when analyzing and modeling data from critical care patients. Previous studies [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] have explored the implications of missing data and sampling biases in predictive modeling, highlighting the potential for these factors to skew results and perpetuate inequalities. Our research builds on these findings by specifically analyzing the demographic disparities in missingness rates and measurement frequencies. We found statistically significant disparities in these patterns between demographic groups and demonstrated the potential to improve the accuracy of the predictive model by including these disparities in the training data.\u003c/p\u003e \u003cp\u003eBased on our results, we caution against relying solely on advanced imputation techniques without first understanding the underlying measurement patterns. Imputation methods should be applied with careful consideration of demographic characteristics to ensure fair representation. Specifically, advanced imputation techniques such as multiple imputation by chained equations (MICE) and deep learning-based imputation methods can be adapted to account for the demographic factors that influence data collection [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. We also recommend incorporating measurement patterns into statistical and machine learning models to mitigate the impact of varied measurement patterns. This can be achieved by weighting the samples and observations based on their measurement frequencies, modeling the data generation process as a multilevel model where the measurement pattern is treated as a latent variable, or employing a Bayesian model to estimate the impact of different measurement frequencies on model outcomes, thus providing a probabilistic framework to handle uncertainty [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Finally, promoting transparency in data collection processes and conducting regular audits of EHR systems to detect bias can help create a more equitable healthcare data landscape. Our comparison with previously published studies underscores the continued need for methodological advancements to tackle these complex challenges and ensure that healthcare analytics are fair and accurate [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOne of the primary strengths of our study is its innovative approach to analyzing EHR data, particularly in identifying and interpreting biases in missing data and measurement frequencies. This approach not only offers a framework for future research but also reveals important disparities in healthcare practices. However, limitations include the potential limited generalizability of our findings, as our study was confined to ICU data coming from the MIMI-III database. Additionally, the study's retrospective nature may limit the ability to infer causality between data collection practices and patient outcomes.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eIn conclusion, this study emphasizes the significance of measurement patterns in ICU EHR data, which are linked to patient demographics and ICU mortality. Analyzing missing data patterns and measurement frequencies provides valuable insights into patient monitoring practices and healthcare equity.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eICU\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntensive Care Unit\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMIMIC-III\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMedical Information Mart for Intensive Care III\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eEHR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eElectronic Health Record\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eGCS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGlasgow Coma Scale\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSOFA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSequential Organ Failure Assessment\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTMLE\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTargeted Maximum Likelihood Estimator\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSuper Learner\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eGLM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGeneralized Linear Models\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAUC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003earea under the receiver operating characteristic curve\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMICE\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emultiple imputation by chained equations\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eEthics approval and consent to participate:\u003c/h2\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003ch2\u003eClinical trial number:\u003c/h2\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003ch2\u003eConsent for publication:\u003c/h2\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003ch2\u003eAvailability of data and materials:\u003c/h2\u003e\n\u003cp\u003eThe data that support the findings of this study are derived from the MIMIC-III database [15], which is publicly available but subject to specific access requirements. Access to MIMIC-III data requires users to complete a data use agreement and undergo a certification process, which includes completing the required training in human subjects research (CITI Program). While the data used in this study were obtained under this agreement and are not publicly available due to licensing restrictions, they may be made available by the authors upon reasonable request, contingent on approval from the MIMIC-III data custodians and adherence to the relevant data use agreements\u003c/p\u003e\n\u003ch2\u003eCompeting interests:\u003c/h2\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003ch2\u003eFunding:\u003c/h2\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003ch2\u003eAuthors\u0026apos; contributions:\u003c/h2\u003e\n\u003cp\u003eJS developed the analytical framework for assessing measurement patterns in the ICU, analyzed the measurement patterns and their relationships with demographic variables, assessed the predictive power of these patterns, interpreted the results, and was the primary contributor to writing the manuscript. NF assisted with data extraction and preprocessing. AH and RP provided oversight and guidance throughout the project. All authors read, revised, and approved the final manuscript.\u003c/p\u003e\n\u003ch2\u003eAcknowledgements:\u003c/h2\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAdler NE, Stead WW. Patients in context\u0026mdash;EHR capture of social and behavioral determinants of health. Obstet Gynecol Surv. 2015;70(6):388\u0026ndash;90.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeaulieu-Jones BK, Lavage DR, Snyder JW, Moore JH, Pendergrass SA, Bauer CR. Characterizing and managing missing structured data in electronic health records: data analysis. JMIR Med Inf. 2018;6(1):e8960.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeaulieu-Jones BK, Moore JH, PIVOTAL-ACT CONSORTIUM. Missing data imputation in the electronic health record using deeply learned autoencoders. In: Pacific Symposium on Biocomputing 2017. Singapore: World Scientific; 2017. pp. 207\u0026thinsp;\u0026ndash;\u0026thinsp;18.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBotsis T, Hartvigsen G, Chen F, Weng C. Secondary use of EHR: data quality issues and informatics opportunities. Summit Transl Bioinform. 2010;2010:1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCoyle JR. tmle3: The extensible TMLE framework [Internet]. 2021. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/tlverse/tmle3\u003c/span\u003e\u003cspan address=\"https://github.com/tlverse/tmle3\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEzzie ME, Aberegg SK, O\u0026rsquo;Brien JM Jr.. Laboratory testing in the intensive care unit. Crit Care Clin. 2007;23(3):435\u0026ndash;65.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFleurence RL, Curtis LH, Califf RM, Platt R, Selby JV, Brown JS. Launching PCORnet, a national patient-centered clinical research network. J Am Med Inf Assoc. 2014;21(4):578\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFong N, Langnas E, Law T, Reddy M, Lipnick M, Pirracchio R. Availability of information needed to evaluate algorithmic fairness\u0026mdash;a systematic review of publicly accessible critical care databases. Anaesth Crit Care Pain Med. 2023;101248.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrassica JJ. Frequency of laboratory test utilization in the intensive care unit and its implications for large-scale data collection efforts. J Am Med Inf Assoc. 2005;12(2):229\u0026ndash;33.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGelman A, Carlin JB, Stern HS, Rubin DB. Bayesian data analysis. Boca Raton: Chapman and Hall/CRC; 1995.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGelman A, Hill J. Data analysis using regression and multilevel/hierarchical models. Cambridge: Cambridge University Press; 2006.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. Physiobank, Physiotoolkit, and Physionet: components of a new research resource for complex physiologic signals. Circulation. 2000;101(23):e215\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoldstein BA, Navar AM, Pencina MJ, Ioannidis JP. Opportunities and challenges in developing risk prediction models with electronic health records data: a systematic review. J Am Med Inf Assoc. 2017;24(1):198\u0026ndash;208.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHyland SL, Faltys M, H\u0026uuml;ser M, Lyu X, Gumbsch T, Esteban C, et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nat Med. 2020;26(3):364\u0026ndash;73.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson AE, Pollard TJ, Shen L, Lehman LW, Feng M, Ghassemi M, et al. MIMIC-III, a freely accessible critical care database. Sci Data. 2016;3(1):1\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson AE, Stone DJ, Celi LA, Pollard TJ. The MIMIC code repository: enabling reproducibility in critical care research. J Am Med Inf Assoc. 2018;25(1):32\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKharrazi H, Wang C, Scharfstein D. Prospective EHR-based clinical trials: the challenge of missing data. J Am Med Inf Assoc. 2014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhope SR, Elias S. Strategies of predictive schemes and clinical diagnosis for prognosis using MIMIC-III: a systematic review. Healthc. 2023;11(710).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLittle RJ, Rubin DB. Statistical analysis with missing data. 3rd ed. Hoboken: Wiley; 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eObermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447\u0026ndash;53.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePivovarov R, Albers DJ, Sepulveda JL, Elhadad N. Identifying and mitigating biases in EHR laboratory tests. J Biomed Inf. 2014;51:24\u0026ndash;34.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRusanov A, Weiskopf NG, Wang S, Weng C. Hidden in plain sight: bias towards sick patients when sampling patients with sufficient electronic health record data for research. BMC Med Inf Decis Mak. 2014;14:1\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSamad MD, Abrar S, Diawara N. Missing value estimation using clustering and deep learning within multiple imputation framework. Knowl Based Syst. 2022;249:108968.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeaman SR, White IR. Review of inverse probability weighting for dealing with missing data. Stat Methods Med Res. 2013;22(3):278\u0026ndash;95.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSharafoddini A, Dubin JA, Maslove DM, Lee J. A new insight into missing data in intensive care unit patient profiles: observational study. JMIR Med Inf. 2019;7(1):e11605.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShivade C, Raghavan P, Fosler-Lussier E, Embi PJ, Elhadad N, Johnson SB, et al. A review of approaches to identifying patient phenotype cohorts using electronic health records. J Am Med Inf Assoc. 2014;21(2):221\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShiver S, Blaivas M, Lyon M. A prospective comparison of ultrasound-guided and blindly placed radial arterial catheters. Acad Emerg Med. 2006;13(12):1275\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun Y, Li J, Xu Y, Zhang T, Wang X. Deep learning versus conventional methods for missing data imputation: a review and comparative study. Expert Syst Appl. 2023;120201.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSuresh H, Guttag JV. A framework for understanding unintended consequences of machine learning. arXiv preprint arXiv:1901 10002. 2019;2(8):73.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUzuner \u0026Ouml;, South BR, Shen S, DuVall SL. 2010 i2b2/VA challenge on concepts, assertions, and relations in clinical text. J Am Med Inform Assoc. 2011;18(5):552-6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003en der Laan MJ, Polley EC, Hubbard AE. Super learner. Stat Appl Genet Mol Biol. 2007;6(1):1\u0026ndash;29.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003en der Laan MJ, Rose S. Targeted learning in data science. New York: Springer; 2018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerheij RA, Curcin V, Delaney BC, McGilchrist MM. Possible sources of bias in primary care electronic health record data use and reuse. J Med Internet Res. 2018;20(5):e185.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWells BJ, Chagin KM, Nowacki AS, Kattan MW. Strategies for handling missing data in electronic health record derived data. EGEMS (Wash DC). 2013;1(3):1\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang D, Yin C, Zeng J, Yuan X, Zhang P. Combining structured and unstructured data for predictive models: a deep learning approach. BMC Med Inf Decis Mak. 2020;20(1):1\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang L, Chen X, Chen T, Wang Z, Mortazavi BJ. Dynehr: dynamic adaptation of models with data heterogeneity in electronic health records. In: 2021 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI). 2021. pp. 1\u0026ndash;4.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Bias Detection, Critical Care, Data Completeness, Electronic Health Records (EHRs), Health Equity, Healthcare Applications, Implicit Bias, Measurement Frequency, MIMIC-III Dataset, Missing Data","lastPublishedDoi":"10.21203/rs.3.rs-5362869/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5362869/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground:\u003c/h2\u003e \u003cp\u003eDisparities in data collection within electronic health records (EHRs), especially in Intensive Care Units (ICUs), can reveal underlying biases that may affect patient outcomes. Identifying and mitigating these biases is critical for ensuring equitable healthcare. This study aims to develop an analytical framework for measurement patterns, including missingness rates and measurement frequencies, evaluate the association between them and demographic factors, and assess their impact on in-hospital mortality prediction.\u003c/p\u003e\u003ch2\u003eMethods:\u003c/h2\u003e \u003cp\u003e We conducted a retrospective cohort study using the Medical Information Mart for Intensive Care III (MIMIC-III) database, which includes data on over 40,000 ICU patients from Beth Israel Deaconess Medical Center (2001\u0026ndash;2012). Adult patients with ICU stays longer than 24 hours were included. Measurement patterns, such as missingnessrates and measurement frequencies, were derived from EHR data and analyzed. Targeted Machine Learning (TML) methods were used to assess potential biases in measurement patterns across demographic factors (age, gender, race/ethnicity) while controlling for confounders such as other demographics and disease severity. The predictive power of measurement patterns on in-hospital mortality was evaluated.\u003c/p\u003e\u003ch2\u003eResults:\u003c/h2\u003e \u003cp\u003eAmong 23,426 patients, significant demographic disparities were observed in the first 24 hours of ICU stays. Elderly patients (\u0026ge;\u0026thinsp;65 years) had more frequent temperature measurements compared to younger patients, while males had slightly fewer missing temperature measurements than females. Racial disparities were notable: White patients had more frequent blood pressure and oxygen saturation (SpO2) measurements compared to Black and Hispanic patients. Measurement patterns were associated with ICU mortality, with models based solely on these patterns achieving an area under the receiver operating characteristic curve (AUC) of 0.76 (95% CI: 0.74\u0026ndash;0.77).\u003c/p\u003e\u003ch2\u003eConclusions:\u003c/h2\u003e \u003cp\u003eThis study underscores the significance of measurement patterns in ICU EHR data, which are associated with patient demographics and ICU mortality. Analyzing patterns of missing data and measurement frequencies provides valuable insights into patient monitoring practices and potential biases in healthcare delivery. Understanding these disparities is critical for improving the fairness of healthcare delivery and developing more accurate predictive models in critical care settings.\u003c/p\u003e","manuscriptTitle":"Implicit Bias in ICU Electronic Health Record Data Measurement Frequencies and Missingness Rates of Clinical Variables","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-11-27 18:40:43","doi":"10.21203/rs.3.rs-5362869/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-01-08T05:44:42+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-12-27T04:06:20+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-12-10T15:51:05+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"309249114688808112226667735827023632232","date":"2024-12-08T00:36:51+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"219933135096686643731198972023173823335","date":"2024-12-06T18:58:31+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-11-05T13:49:04+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-11-05T07:24:08+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-11-02T07:32:25+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2024-10-30T17:46:04+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"4dbd664f-cd30-4f8b-81d4-8897c5ba84b5","owner":[],"postedDate":"November 27th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-07-07T16:00:24+00:00","versionOfRecord":{"articleIdentity":"rs-5362869","link":"https://doi.org/10.1186/s12911-025-03058-9","journal":{"identity":"bmc-medical-informatics-and-decision-making","isVorOnly":false,"title":"BMC Medical Informatics and Decision Making"},"publishedOn":"2025-07-01 15:57:10","publishedOnDateReadable":"July 1st, 2025"},"versionCreatedAt":"2024-11-27 18:40:43","video":"","vorDoi":"10.1186/s12911-025-03058-9","vorDoiUrl":"https://doi.org/10.1186/s12911-025-03058-9","workflowStages":[]},"version":"v1","identity":"rs-5362869","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5362869","identity":"rs-5362869","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}