Use of a Machine Learning Model to Predict Iatrogenic Hypoglycemia (2024)

Key Points

Question Can the risk of iatrogenic hypoglycemia resulting from insulin or insulin secretagogues be predicted continuously throughout hospitalization without the use of continuous glucose monitors?

Findings In this cohort study of 54 978 admissions in a large US health care system, a stochastic gradient boosting machine learning model using 43 static and time-varying clinical predictors available in the electronic medical record accurately predicted the risk of iatrogenic hypoglycemia in a prediction horizon of 24 hours from the time of each point-of-care and serum glucose measurement throughout a patient’s hospital admission.

Meaning These findings suggest that translating this machine learning prediction model into a real-time informatics alert embedded in the electronic medical record has the potential to reduce the incidence of iatrogenic hypoglycemia, a serious adverse event.


Importance Accurate clinical decision support tools are needed to identify patients at risk for iatrogenic hypoglycemia, a potentially serious adverse event, throughout hospitalization.

Objective To predict the risk of iatrogenic hypoglycemia within 24 hours after each blood glucose (BG) measurement during hospitalization using a machine learning model.

Design, Setting, and Participants This retrospective cohort study, conducted at 5 hospitals within the Johns Hopkins Health System, included 54 978 admissions of 35 147 inpatients who had at least 4 BG measurements and received at least 1 U of insulin during hospitalization between December 1, 2014, and July 31, 2018. Data from the largest hospital were split into a 70% training set and 30% test set. A stochastic gradient boosting machine learning model was developed using the training set and validated on internal and external validation.

Exposures A total of 43 clinical predictors of iatrogenic hypoglycemia were extracted from the electronic medical record, including demographic characteristics, diagnoses, procedures, laboratory data, medications, orders, anthropomorphometric data, and vital signs.

Main Outcomes and Measures Iatrogenic hypoglycemia was defined as a BG measurement less than or equal to 70 mg/dL occurring within the pharmacologic duration of action of administered insulin, sulfonylurea, or meglitinide.

Results This cohort study included 54 978 admissions (35 147 inpatients; median [interquartile range] age, 66.0 [56.0-75.0] years; 27 781 [50.5%] male; 30 429 [55.3%] White) from 5 hospitals. Of 1 612 425 index BG measurements, 50 354 (3.1%) were followed by iatrogenic hypoglycemia in the subsequent 24 hours. On internal validation, the model achieved a C statistic of 0.90 (95% CI, 0.89-0.90), a positive predictive value of 0.09 (95% CI, 0.08-0.09), a positive likelihood ratio of 4.67 (95% CI, 4.59-4.74), a negative predictive value of 1.00 (95% CI, 1.00-1.00), and a negative likelihood ratio of 0.22 (95% CI, 0.21-0.23). On external validation, the model achieved C statistics ranging from 0.86 to 0.88, positive predictive values ranging from 0.12 to 0.13, negative predictive values of 0.99, positive likelihood ratios ranging from 3.09 to 3.89, and negative likelihood ratios ranging from 0.23 to 0.25. Basal insulin dose, coefficient of variation of BG, and previous hypoglycemic episodes were the strongest predictors.

Conclusions and Relevance These findings suggest that iatrogenic hypoglycemia can be predicted in a short-term prediction horizon after each BG measurement during hospitalization. Further studies are needed to translate this model into a real-time informatics alert and evaluate its effectiveness in reducing the incidence of inpatient iatrogenic hypoglycemia.


Inpatient hypoglycemia is a prevalent and often preventable adverse event associated with increased morbidity and mortality, length of stay, readmissions, and health care expenditures.1-6 Considering that most patients with diabetes are hospitalized for reasons other than glucose management,7,8 situational unawareness of the near-term risk of iatrogenic hypoglycemia may result from competing medical priorities, lack of sufficient training in glycemic pattern recognition,9-11 or other system factors.12 Studies13-15 have demonstrated practitioner inertia in adjusting glucose-lowering medications before hypoglycemic events or in response to antecedent hypoglycemia.

The hospital setting poses unique challenges for the identification and prevention of hypoglycemia that results from insulin or other antihyperglycemic therapies.7 Many hospitalized patients have 1 or more risk factors for hypoglycemia, including advanced age, decreased renal function, liver disease, poor appetite, or nil per os (nothing by mouth) status, and some of these risk factors may change throughout a patient’s hospital stay.1 Changes in dextrose-containing fluids, tapering corticosteroid doses, disruption in parenteral nutrition or continuous tube feedings, mismatched timing of point-of-care blood glucose (BG) testing with insulin delivery and meal consumption, correctional scale overuse, and insulin stacking are just a few examples of the various factors that can influence the risk of iatrogenic hypoglycemia.1

One approach to improving situational awareness in the face of multiple dynamic data elements is to leverage the power of electronic medical record (EMR) data for real-time prediction and alerting. Previously published inpatient hypoglycemia prediction models14,16-19 differ with respect to cohort sample size, outcome definition, number of predictor variables, prediction horizon, statistical modeling approach, validation methods, and performance. Recently published models using logistic regression19 or machine learning20 have achieved high predictive accuracy, but the prediction horizon (during several days or entire admission) may be too long to be clinically useful in real time to guide therapeutic changes and prevent this outcome.

Because insulin dose adjustments are typically made daily for hospitalized patients based on a review of glycemic trends, an inpatient hypoglycemia prediction model that is intended to be usable in real time should predict the outcome of interest within a relatively narrow prediction horizon (ie, 24 hours) using rolling data from the EMR to account for the dynamic changes that occur throughout a patient’s hospitalization. In this study, we sought to develop a real-time prediction model for the outcome of iatrogenic hypoglycemia within a 24-hour rolling window from the time of each individual BG measurement throughout hospitalization. Moreover, given the lack of externally validated models for the outcome of iatrogenic hypoglycemia in the inpatient setting, we sought to externally validate our model by comparing its performance within different hospitals in a large health care system.


Study Design

The study investigators worked with Epic Clarity–certified clinical analysts to extract the data from EpicCare, the EMR system for all study hospitals. A schematic representation of the data flow from Epic Clarity (source data) to the final analytic data set is shown in eFigure 1 in the Supplement. This study followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) reporting guideline.21 The study protocol was approved by the institutional review board of the Johns Hopkins School of Medicine with a waiver of informed consent because use of the large EMR data set made it impossible. Deidentified data were used during the analysis.

This retrospective cohort study consisted of EMR data obtained during routine clinical care from 5 (2 academic and 3 community) hospitals within the Johns Hopkins Health System located in Maryland and the District of Columbia. Figure 1 shows the study flowchart. The cohort consisted of hospitalized adult patients who had at least 4 BG measurements obtained during admission, received at least 1 U of subcutaneous insulin at any time during admission, and were admitted between December 1, 2014, and July 31, 2018. Admissions of patients with a length of stay of less than 24 hours, missing weight information, and treatment with intravenous insulin or insulin pumps were excluded.

The focus of this study was on noncritically ill hospitalized patients in general medical or surgical wards; therefore, BG measurements obtained at any time during admission in the intensive care unit setting were excluded. Duplicate BG measurements (ie, same value occurring at the same time) were excluded. The unit of measurement was each BG value. Because hypoglycemia as an outcome cannot be predicted for the first BG value, these values were excluded as index BG measurements. Similarly, the last BG measurement of admission was excluded because there was no future outcome to predict. After these exclusions, there were approximately 1.6 million index BG measurements of approximately 55 000 patient admissions from approximately 35 000 patients.


The outcome of interest was iatrogenic hypoglycemia occurring within a rolling prediction horizon of 24 hours after each index BG measurement throughout the patient’s hospital stay. Iatrogenic hypoglycemia was defined as a serum or point-of-care BG level less than or equal to 70 mg/dL (to convert to millimoles per liter, multiply by 0.0555) within the pharmacologic duration of action of administered insulins or within 24 hours of administration of a sulfonylurea or meglitinide.


We extracted candidate predictors from the EMR based on clinical knowledge and prior studies.14,16,18-20 All predictor variables were collected from the index admission, although some variables were from prior admissions. eTable 1 in the Supplement gives the definitions of each candidate predictor variable. Predictor variables were processed to capture accruing information in varying time frames between the time of admission and the index BG value.

Missing Data

For most predictor variables in the data set, there were no missing observations. For missing laboratory and vital sign values, we first replaced missing observations with the median of all prior results of the variable for the patient during the admission up to the time of the index BG value. If no prior results were available, we replaced the missing result with the hospital-specific median value for all patient admissions. The counts of missing observations for each variable by hospital and the imputation methods are provided in eTable 2 in the Supplement.

Machine Learning Models

We evaluated the performance of 4 different classification algorithms for prediction of the outcome: multivariable logistic regression, random forest classification, naive Bayes, and stochastic gradient boosting (SGB). Consistent with findings from a recent study,20 SGB had the highest area under the receiver operating curve on internal validation (70/30 split) using data from hospital 1; therefore, this approach was used for model development and validation.

Model Development and Validation

Although SGB was used for development of the final prediction model, we relied on various machine learning methods for selection of our predictor variables, including univariate analysis on simple logistic regression, stepwise logistic regression and Akaike information constant, variable importance plots from random forest models, and relative influence plots in the SGB model. In selecting predictor variables, we sought to achieve the most parsimonious model possible to minimize the amount of preprocessing and computing demands that would ultimately be required if such a model were integrated into an EMR for real-time prediction.

Data from hospital 1 (largest hospital) were split into a 70% training set and 30% test set after sorting the data in chronological order by patient admission and the time of each index BG observation. The rationale for sorting data chronologically for validation was to ensure that the performance measures of the model were as conservative as possible by mimicking the real world, in which secular trends associated with unmeasured variables could influence rates of the outcome.

We developed the SGB model on the training data set. The model was constructed to predict the probability of the outcome for each index BG value. Hyperparameter tuning was performed via grid search with 10-fold cross-validation in the training set with the aim of optimizing the F1 score (harmonic mean of precision and recall). Internal validation was performed by using the training set from hospital 1 to predict the observations on the test set from hospital 1. External validation was performed by using the training set from hospital 1 to predict observations in the full data sets from hospitals 2, 3, 4, and 5.

All statistical analyses were performed using R statistical software, version 3.4.1 (R Foundation for Statistical Computing) and Stata software, version 15.1 (StataCorp LLC). The Caret R package was used for parameter optimizing, model training, and evaluation.22 Our prediction model included multiple correlated covariates at the individual level. Because our goal was to predict events at the individual level, our sampling criteria and subsequent modeling process were designed to exploit this within-patient correlation information to better predict patient-specific outcomes. Thus, our analysis did not use any clustering techniques to account for nonindependence of covariates at the patient level.

Because of the imbalance in class (positive or negative) of our outcome and the limitations of using the area under the curve alone as a comprehensive indicator of performance,23 we report the sensitivity (recall), positive predictive value (PPV; precision), and positive likelihood ratio (+LR; allows inferences independent of disease prevalence) together with their inverse measures (eg, specificity, negative predictive value [NPV], and negative likelihood ratio [−LR]), false-positive rate, and false-negative rate. Performance measures are reported at the probability threshold perceived by the investigators to achieve most favorable combination of sensitivity and specificity. A 2-sided P < .05 was considered statistically significant.


Cohort Characteristics

This study included 54 978 admissions (35 147 inpatients; median [interquartile range] age, 66.0 [56.0-75.0] years; 27 781 [50.5%] male; 30 429 [55.3%] White) from 5 hospitals, of which 33 301 (61.6%) were academic hospital admissions and 21 677 (39.4%) were community hospital admissions. Table 1 gives the characteristics at the admission level. Despite the inclusion criterion of at least 1 U of insulin administration during hospitalization, 31 203 patients (56.8%) had no coded diagnosis of diabetes present on admission; type 2 diabetes was present in 21 660 admissions (39.4%) and type 1 diabetes in 1321 admissions (2.4%). Insulin was used at home in 13 764 admissions (25.0%) and insulin secretagogues in 7650 admissions (13.9%). There were notable differences with respect to age, race/ethnicity, weight, body mass index, length of stay, diabetes type, use of home insulin, and use of insulin secretagogues by hospital.

At least 1 hypoglycemic episode occurred in 8765 admissions (15.9%). At the index BG observation level, 50 345 of 1 612 425 index BG measurements (3.1%) were followed by an iatrogenic hypoglycemic episode in the next 24 hours. eTable 3 in the Supplement gives the characteristics of the cohort at the index BG level. eFigures 2 and 3 in the Supplement show the frequency distribution of iatrogenic hypoglycemia at the admission and patient levels, respectively.

Model Specification

The final prediction model included 43 predictor variables, of which 13 were static and 30 were time varying (eTables 1 and 4 in the Supplement). Figure 2 shows the relative importance of the top 30 predictor variables. The most important predictors were pharmacologically active basal insulin at the time of the index BG measurement, coefficient of variation of BG, any previous hypoglycemic episodes, hospital day number, nadir BG value during admission, index BG value, weight, and mean BG value in the past 24 hours. Figure 3 illustrates the dynamic nature of the prediction model using a sample patient from the cohort.

Model Performance

Table 2 gives the model performance on internal and external validation. Our model achieved strong discrimination, with a C statistic of 0.90 (95% CI, 0.89-0.90) on internal validation and ranging from 0.86 to 0.88 on external validation; however, the high discrimination can in large part be attributed to the high prevalence of correctly predicted nonevents. On internal validation, the model achieved a PPV of 0.09 (95% CI, 0.08-0.09), NPV of 1.00 (95% CI, 1.00-1.00), +LR of 4.67 (95% CI, 4.59-4.74), −LR of 0.22 (95% CI, 0.21-0.23), false-negative rate of 18%, and false-positive rate of 18%. On external validation, the PPV ranged from 0.12 to 0.13, NPV was 0.99, +LR ranged from 3.09 to 3.89, −LR ranged from 0.23 to 0.25, false-negative rate was 18%, and false-positive rate ranged from 21% to 27%. The +LR, which is an indicator of disease likelihood that is independent of disease prevalence, was approximately 5 on internal validation, which corresponds to an approximate 30% increase in the probability of the outcome with a positive prediction.24 We conducted a sensitivity analysis in which we excluded observations in which any variable had a missing result and found that the model performed similarly (eTable 5 in the Supplement).


In this cohort study, an SGB machine learning algorithm predicted the potentially serious outcome of iatrogenic hypoglycemia within a narrow prediction horizon of 24 hours after each BG measurement throughout hospitalization. To our knowledge, this prediction model is the first to use such a near-term prediction horizon without reliance on continuous BG monitoring, increasing the generalizability of the model for use in a large number of hospitalized patients. Unlike other machine learning models that have been developed for prediction of inpatient hypoglycemia,14,19,20 the narrow prediction horizon of the model in the present study allows for the possibility of short-term, continuous prediction throughout the patient’s hospital stay. Furthermore, unlike previous models that have undergone only internal validation,20 these findings appear to be generalizable across hospitals with heterogeneous patient populations.

Different prediction horizons can influence not only outcome prevalence (and therefore model performance) but also the ability to translate a model into a real-time alerting system. Models that use a prediction horizon of the entire admission span20 would be expected to have a higher outcome prevalence and potentially higher discrimination than those that use more narrow prediction horizons; although hospital-level prediction models could potentially be used to flag patients at risk for the outcome at the time of admission, they would not be useful for real-time alerting throughout a patient’s entire hospitalization. Furthermore, many inpatients experience repeated episodes of hypoglycemia during an admission; a 1-time prediction of hypoglycemia at admission would have diminished the ability to prevent repeated episodes in such patients.

Most practitioners who treat inpatients review glycemic data and adjust insulin doses at most daily; because basal insulin is usually administered once or twice daily, practitioners must be alerted to the possibility of iatrogenic hypoglycemia as soon as the risk is detected to allow them sufficient time to proactively adjust insulin doses. A prediction horizon of 24 hours after each BG measurement should generally allow sufficient lead time for practitioners to adjust antihyperglycemic therapies. Despite the low prevalence of iatrogenic hypoglycemia with this prediction horizon (approximately 3%), the prediction model nonetheless achieved +LRs of 4.67 on internal validation and 3.09 to 3.89 on external validation. As a reference, a +LR of 2 indicates small (approximately 15%), a +LR of 5 indicates moderate (approximately 30%), and a +LR of 10 indicates large (approximately 45%) increases in the probability of an outcome with a positive result.24 The present model had a nearly perfect NPV (given low disease prevalence), with −LRs of 0.22 to 0.25, corresponding to an approximately 30% decrease in the probability of hypoglycemia with a negative result. Because practitioners currently have no existing clinical decision support tools to assess a patient’s risk of near-term hypoglycemia on an ongoing basis in the admission, translating this prediction model into an EMR-based decision support tool could facilitate hypoglycemia risk prediction.

A challenge in developing prediction models using EMR data is the extensive amount of data processing required to create a relational database that presents observations in the same way that a practitioner would when assessing a patient’s risk of a near-term outcome through manual review of the EMR. For example, a practitioner might review insulin doses, corticosteroid doses, glycemic trends, diet orders, and laboratory data throughout the admission and from the previous 24 hours when considering the need for therapeutic changes. In the present model, extensive data processing from the EMR was required to capture pharmacologic doses of basal insulin and corticosteroids, 2 key medications that influence glucose levels. Not surprisingly, the active basal insulin dose and glycemic summary measures were the strongest predictors of near-term hypoglycemia. It was surprising to find that corticosteroid doses and kidney function were only modestly predictive. Interestingly, this machine learning approach identified some predictors (laboratory test results and vital signs) that are not traditionally known to be risk factors for hypoglycemia, and it is more likely that these predictors are associated with patient severity of illness than directly with risk of insulin or sulfonylurea-related hypoglycemia.

In recent years, EMR data mining and machine learning have been used increasingly to develop prediction models for a wide range of clinical outcomes in the inpatient setting,25,26 including hypoglycemia.20 Big data studies18,19,27 have achieved significantly higher discrimination (C statistics of 0.80-0.99) for prediction of iatrogenic hypoglycemia compared with previous studies14,16,17 that used smaller cohorts (C statistics of 0.68-0.73). In a UK sample of 32 758 admissions, Ruan et al20 achieved a C statistic of 0.96 using an XGBoost machine learning model with 42 predictor variables. Notably, however, their prediction horizon was the entire patient admission and achieved a sensitivity of 0.70. Furthermore, in their study, 10-fold cross-validation in a single data set was used to estimate model performance, which could overestimate the performance in a real-world setting, where secular trends cannot be accounted for. The findings of the present study suggest that the narrower prediction horizon and external validation of the model advance the work of other groups, making the model more useful at the point of care and more generalizable across hospitals. The present study used a chronological 70/30 split for validation rather than 10-fold cross-validation, which may provide more conservative estimates of model performance and better account for temporal trends than methods that randomly partition data for training and validation.

The high discrimination in this model was in large part the result of the high number of correctly predicted true negative results. Translation of this model into a real-time informatics alert would require adjustment of the probability threshold to further increase specificity to reduce alert fatigue, even at the expense of reduced sensitivity. Consider the following 2 scenarios. First, when increasing the sensitivity to 87%, specificity decreases to 77%, with a corresponding increase in the false-positive rate from 18% to 23% and a decrease in the false-negative rate from 18% to 13%. Second, conversely, reducing sensitivity to 75% results in a specificity increase to 87%, with a corresponding decrease in the false-positive rate from 18% to 13% and an increase in the false-negative rate from 18% to 25%. Consider a practitioner who is treating 10 patients with diabetes and is expected to receive 30 BG readings from these patients in a single day. An increase in the false-positive rate of 5% could mean an additional 1 or 2 unnecessary alerts per day, which could easily contribute to practitioner alert fatigue when compounded during multiple shifts, whereas an increase in the false-negative rate could mean an additional 1 or 2 missed opportunities for hypoglycemia prevention. Thus, selecting an appropriate probability threshold would need to carefully balance perceived benefit of increased sensitivity (efficacy) against reduced specificity (alert fatigue).


This study has limitations. It was not possible to extract information about dextrose doses because of the various numbers of medications that contain dextrose as an additive. In addition, it was not possible to capture insulin doses as additives in continuous parenteral nutrition formulations. Also, external validation was conducted at hospitals that are within the overall health system and region, and results could differ when validated in different sociodemographic populations. Finally, the model could not account for erroneous finger-stick glucose readings or distinguish between symptomatic and asymptomatic hypoglycemia. Reliance exclusively on EMR data cannot capture all the clinical information that a practitioner uses on a daily basis when making insulin dosing decisions.


This study demonstrates for the first time, to our knowledge, that a machine learning algorithm has been used to predict the short-term risk of iatrogenic hypoglycemia continuously throughout hospitalization; it achieved a modest degree of accuracy without reliance on data from continuous glucose monitors. Our next step will be to embed the prediction model in our EMR system. We have recently completed a qualitative study of practitioner stakeholders to identify the optimal features and format of a clinical decision support tool based on our prediction model. Additional studies will be needed to test the real-time effectiveness of an informatics alert derived from this prediction model in reducing the incidence of this potentially serious adverse event.

Back to top

Article Information

Accepted for Publication: November 1, 2020.

Published: January 8, 2021. doi:10.1001/jamanetworkopen.2020.30913

Open Access: This is an open access article distributed under the terms of the CC-BY License. © 2021 Mathioudakis NN et al. JAMA Network Open.

Corresponding Author: Nestoras N. Mathioudakis, MD, MHS, Division of Endocrinology, Diabetes & Metabolism, Department of Medicine, Johns Hopkins University School of Medicine, 1830 E Monument St, Ste 333, Baltimore, MD 21287 (

Author Contributions: Drs Mathioudakis and Abusamaan contributed significantly to this article and should be considered co–first authors. They had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Concept and design: Mathioudakis, Abusamaan, Fayzullin, McGready, Zilbermint, Golden.

Acquisition, analysis, or interpretation of data: Mathioudakis, Abusamaan, Shakarchi, Sokolinsky, Fayzullin, McGready, Saria.

Drafting of the manuscript: Mathioudakis, Abusamaan, Shakarchi, Sokolinsky, McGready.

Critical revision of the manuscript for important intellectual content: Mathioudakis, Abusamaan, Shakarchi, Fayzullin, Zilbermint, Saria, Golden.

Statistical analysis: Mathioudakis, Abusamaan, Shakarchi, Fayzullin, McGready.

Obtained funding: Mathioudakis.

Administrative, technical, or material support: Mathioudakis, Abusamaan, Sokolinsky.

Supervision: Mathioudakis, Abusamaan, Saria, Golden.

Conflict of Interest Disclosures: Dr McGready reported receiving grants from Johns Hopkins University during the conduct of the study. Dr Zilbermint reported receiving consulting fees from Guidepoint, G.L.G., and Sacramento HealthCare Investors LLC Investor outside the submitted work. Dr Saria reported being a founder of and holding significant equity in Bayesian Health; serving as a member of the scientific advisory board for PatientPing, Child Health Imprints, Halcyon, and Duality Technologies; receiving honoraria for speaking engagements by Sanofi, Abbvie, and Novartis; and receiving funding from Defense Advanced Research Projects Agency, the US Food and Drug Administration, American Heart Association, National Institutes of Health, National Science Foundation, and the Gordon Betty Moore Foundation. Dr Golden reported receiving grants from Merck and Co Inc outside the submitted work. No other disclosures were reported.

Funding/Support: This study was supported by grant K23DK111986 from the National Institute for Diabetes and Digestive and Kidney Diseases (Drs Mathioudakis, Abusamaan, and McGready).

Role of the Funder/Sponsor: The funding source had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Disclaimer: The opinions expressed in this article are the authors’ own and do not reflect the view of the National Institutes of Health, the US Department of Health and Human Services, or the US government.



Cruz P. Inpatient hypoglycemia: the challenge remains. J Diabetes Sci Technol. 2020;14(3):560-566. doi:10.1177/1932296820918540 PubMedGoogle ScholarCrossref


Garg R, Hurwitz S, Turchin A, Trivedi A. Hypoglycemia, with or without insulin therapy, is associated with increased mortality among hospitalized patients. Diabetes Care. 2013;36(5):1107-1110. doi:10.2337/dc12-1296 PubMedGoogle ScholarCrossref


Turchin A, Matheny ME, Shubina M, Scanlon JV, Greenwood B, Pendergrass ML. Hypoglycemia and clinical outcomes in patients with diabetes hospitalized in the general ward. Diabetes Care. 2009;32(7):1153-1157. doi:10.2337/dc08-2127 PubMedGoogle ScholarCrossref


Boucai L, Southern WN, Zonszein J. Hypoglycemia-associated mortality is not drug-associated but linked to comorbidities. Am J Med. 2011;124(11):1028-1035. doi:10.1016/j.amjmed.2011.07.011 PubMedGoogle ScholarCrossref


Brutsaert E, Carey M, Zonszein J. The clinical impact of inpatient hypoglycemia. J Diabetes Complications. 2014;28(4):565-572. doi:10.1016/j.jdiacomp.2014.03.002PubMedGoogle ScholarCrossref


Brodovicz KG, Mehta V, Zhang Q, et al. Association between hypoglycemia and inpatient mortality and length of hospital stay in hospitalized, insulin-treated patients. Curr Med Res Opin. 2013;29(2):101-107. doi:10.1185/03007995.2012.754744 PubMedGoogle ScholarCrossref


Hulkower RD, Pollack RM, Zonszein J. Understanding hypoglycemia in hospitalized patients. Diabetes Manag (Lond). 2014;4(2):165-176. doi:10.2217/dmt.13.73 PubMedGoogle ScholarCrossref


Roman SH, Chassin MR. Windows of opportunity to improve diabetes care when patients with diabetes are hospitalized for other conditions. Diabetes Care. 2001;24(8):1371-1376. doi:10.2337/diacare.24.8.1371 PubMedGoogle ScholarCrossref


Cook CB, Jameson KA, Hartsell ZC, et al. Beliefs about hospital diabetes and perceived barriers to glucose management among inpatient midlevel practitioners. Diabetes Educ. 2008;34(1):75-83. doi:10.1177/0145721707311957PubMedGoogle ScholarCrossref


Latta S, Alhosaini MN, Al-Solaiman Y, et al. Management of inpatient hyperglycemia: assessing knowledge and barriers to better care among residents. Am J Ther. 2011;18(5):355-365. doi:10.1097/MJT.0b013e3181d1d847 PubMedGoogle ScholarCrossref


Cook CB, McNaughton DA, Braddy CM, et al. Management of inpatient hyperglycemia: assessing perceptions and barriers to care among resident physicians. Endocr Pract. 2007;13(2):117-124. doi:10.4158/EP.13.2.117PubMedGoogle ScholarCrossref


Horton WB, Law S, Darji M, et al. A multicenter study evaluating perceptions and knowledge of inpatient glycemic control among resident physicians: analyzing themes to inform and improve care. Endocr Pract. 2019;25(12):1295-1303. doi:10.4158/EP-2019-0299 PubMedGoogle ScholarCrossref


Mathioudakis N, Everett E, Golden SH. Prevention and management of insulin-associated hypoglycemia in hospitalized patients. Endocr Pract. 2016;22(8):959-969. doi:10.4158/EP151119.OR PubMedGoogle ScholarCrossref


Elliott MB, Schafers SJ, McGill JB, Tobin GS. Prediction and prevention of treatment-related inpatient hypoglycemia. J Diabetes Sci Technol. 2012;6(2):302-309. doi:10.1177/193229681200600213 PubMedGoogle ScholarCrossref


Cook CB, Castro JC, Schmidt RE, et al. Diabetes care in hospitalized noncritically ill patients: more evidence for clinical inertia and negative therapeutic momentum. J Hosp Med. 2007;2(4):203-211. doi:10.1002/jhm.188 PubMedGoogle ScholarCrossref


Stuart K, Adderley NJ, Marshall T, et al. Predicting inpatient hypoglycaemia in hospitalized patients with diabetes: a retrospective analysis of 9584 admissions with diabetes. Diabet Med. 2017;34(10):1385-1391. doi:10.1111/dme.13409PubMedGoogle ScholarCrossref


Ena J, Gaviria AZ, Romero-Sánchez M, et al; Diabetes and Obesity Working Group of the Spanish Society of Internal Medicine. Derivation and validation model for hospital hypoglycemia. Eur J Intern Med. 2018;47:43-48. doi:10.1016/j.ejim.2017.08.024 PubMedGoogle ScholarCrossref


Mathioudakis NN, Everett E, Routh S, et al. Development and validation of a prediction model for insulin-associated hypoglycemia in non-critically ill hospitalized adults. BMJ Open Diabetes Res Care. 2018;6(1):e000499. doi:10.1136/bmjdrc-2017-000499 PubMedGoogle Scholar


Winterstein AG, Jeon N, Staley B, Xu D, Henriksen C, Lipori GP. Development and validation of an automated algorithm for identifying patients at high risk for drug-induced hypoglycemia. Am J Health Syst Pharm. 2018;75(21):1714-1728. doi:10.2146/ajhp180071 PubMedGoogle ScholarCrossref


Ruan Y, Bellot A, Moysova Z, et al. Predicting the risk of inpatient hypoglycemia with machine learning using electronic health records. Diabetes Care. 2020;43(7):1504-1511. doi:10.2337/dc19-1743 PubMedGoogle ScholarCrossref


Moons KG, Altman DG, Reitsma JB, Collins GS; Transparent Reporting of a Multivariate Prediction Model for Individual Prognosis or Development Initiative. New guideline for the reporting of studies developing, validating, or updating a multivariable clinical prediction model: the TRIPOD statement. Adv Anat Pathol. 2015;22(5):303-305. doi:10.1097/PAP.0000000000000072 PubMedGoogle ScholarCrossref


Kuhn M. Building predictive models in R using the caret package. J Stat Softw. 2008;28(5):1–26. doi:10.1371/journal.pone.0118432 Google ScholarCrossref


Saito T, Rehmsmeier M. The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS One. 2015;10(3):e0118432. doi:10.1371/journal.pone.0118432 PubMedGoogle Scholar


McGee S. Simplifying likelihood ratios. J Gen Intern Med. 2002;17(8):646-649. doi:10.1046/j.1525-1497.2002.10750.x PubMedGoogle ScholarCrossref


Park C, Took CC, Seong JK. Machine learning in biomedical engineering. Biomed Eng Lett. 2018;8(1):1-3. doi:10.1007/s13534-018-0058-3 PubMedGoogle ScholarCrossref


Bates DW, Saria S, Ohno-Machado L, Shah A, Escobar G. Big data in health care: using analytics to identify and manage high-risk and high-cost patients. Health Aff (Millwood). 2014;33(7):1123-1131. doi:10.1377/hlthaff.2014.0041 PubMedGoogle ScholarCrossref


Ruan Y, Tan GD, Lumb A, Rea RD. Importance of inpatient hypoglycaemia: impact, prediction and prevention. Diabet Med. 2019;36(4):434-443. doi:10.1111/dme.13897PubMedGoogle ScholarCrossref

Use of a Machine Learning Model to Predict Iatrogenic Hypoglycemia (2024)


Which machine learning model is best for diabetes prediction? ›

The random forest model emerged as the top performer, achieving an accuracy of 0.77 on the test dataset. We also gained valuable insights into feature importance, with glucose levels being the most influential predictor of diabetes in this dataset.

What is an iatrogenic cause of hypoglycemia? ›

Iatrogenic hypoglycemia is low blood sugar that usually happens when you have diabetes and have too much insulin in your body. This can happen when you accidentally take too much insulin or your body is more sensitive to insulin. It's most common among people with type 1 diabetes.

What is the problem statement for diabetes prediction using machine learning? ›

The problem statement is to determine which algorithm is more accurate in predicting diabetes. The methodology involves implementing the SVM and decision tree algorithms on the dataset and evaluating their performance using metrics such as accuracy, precision, and recall.

What is the classic test to determine a diagnosis of hypoglycemia? ›

If you have signs or symptoms of low blood sugar, check your blood sugar level with a blood glucose meter — a small device that measures and displays your blood sugar level. You have hypoglycemia when your blood sugar level drops below 70 milligrams per deciliter (mg/dL) (3.9 millimoles per liter (mmol/L)).

Which machine learning algorithm is best for disease prediction? ›

Random Forest Classifier: Random Forest is an ensemble learning-based supervised machine learning classification algorithm that internally uses multiple decision trees to make the classification.

What is the best machine learning model for prediction? ›

The most widely used predictive models are:
  • Decision trees: Decision trees are a simple, but powerful form of multiple variable analysis. ...
  • Regression (linear and logistic) Regression is one of the most popular methods in statistics. ...
  • Neural networks.

What is the most common cause of hypoglycemia is --- iatrogenic diabetes mellitus endocrine disorder? ›

Most hypoglycemia is caused by medications used to treat diabetes mellitus (including surreptitious use); insulin-secreting tumors are rare causes. If the etiology of hypoglycemia is unclear, do a 48- or 72-hour fast with measurement of plasma glucose at regular intervals and whenever symptoms occur.

What is the most common cause of hypoglycemia in diabetic patients? ›

Common causes of diabetic hypoglycemia include:
  • Taking too much insulin or diabetes medication.
  • Not eating enough.
  • Postponing or skipping a meal or snack.
  • Increasing exercise or physical activity without eating more or adjusting your medications.
  • Drinking alcohol.
May 6, 2022

What is most likely the cause of hypoglycemia? ›

Hypoglycemia can occur with malnutrition and starvation when you don't get enough food, and the glycogen stores your body needs to create glucose are used up. An eating disorder called anorexia nervosa is one example of a condition that can cause hypoglycemia and result in long-term starvation.

How to use machine learning for prediction? ›

Multiple samples of your training data are taken then models are constructed for each data sample. When you need to make a prediction for new data, each model makes a prediction and the predictions are averaged to give a better estimate of the true output value.

What machine learning algorithms are used to detect diabetes? ›

[21] used ML algorithms and neural networks (NNs) for diabetes prediction. The algorithms used included DT, random forest (RF), LR, Naïve Bayes (NB), KNN, and SVM, and they found that the LR and SVM algorithms obtained better performances in terms of accuracy for diabetes prediction. The article by Ref.

What is the most common issue faced by machine learning models? ›

The number one problem facing Machine Learning is the lack of good data. While enhancing algorithms often consumes most of the time of developers in AI, data quality is essential for the algorithms to function as intended.

What 5 things should you look for to identify hypoglycemia? ›

Signs and symptoms of hypoglycaemia can include:
  • feeling hungry.
  • sweating.
  • dizziness.
  • tiredness (fatigue)
  • blurred vision.
  • trembling or shakiness.
  • going pale.
  • fast pulse or palpitations.
Feb 20, 2023

Is there a way to test for hypoglycemia? ›

The most common way to do so is by using a blood glucose meter. If you have hypoglycemia unawareness or have low blood glucose often, a continuous glucose monitor (CGM) may be a good option. The CGM measures your blood glucose level at regular times and can sound an alarm if it drops below your target range.

Can I have hypoglycemia without diabetes? ›

Hypoglycemia is frequently observed in patients with diabetes mellitus but is uncommon in patients without diabetes. Hypoglycemia in patients without diabetes may be due to a variety of causes. Therefore, it is important to take a thorough clinical history and examination.

What algorithm predicts diabetes? ›

The naive Bayes classifier is the most widely used classification algorithm, with an accuracy of 93.27%. SVM has the highest accuracy rate of 96.54%. LightGBM has an accuracy of only 88.46%. This demonstrates that SVM is the best classification algorithm for diabetes prediction.

What is the best predictor of diabetes? ›

HbA1c exhibited the strongest predictive power for type 2 diabetes, followed by BMI, waist circumference, blood glucose levels, the number of first-degree relatives with diabetes, GGT, waist-hip ratio, HDL cholesterol, age, and urate levels.

Which type of machine learning should you use to predict? ›

The most common type of machine learning is to learn the mapping Y = f(X) to make predictions of Y for new X. This is called predictive modeling or predictive analytics and our goal is to make the most accurate predictions possible.

Top Articles
Latest Posts
Article information

Author: Frankie Dare

Last Updated:

Views: 6088

Rating: 4.2 / 5 (53 voted)

Reviews: 84% of readers found this page helpful

Author information

Name: Frankie Dare

Birthday: 2000-01-27

Address: Suite 313 45115 Caridad Freeway, Port Barabaraville, MS 66713

Phone: +3769542039359

Job: Sales Manager

Hobby: Baton twirling, Stand-up comedy, Leather crafting, Rugby, tabletop games, Jigsaw puzzles, Air sports

Introduction: My name is Frankie Dare, I am a funny, beautiful, proud, fair, pleasant, cheerful, enthusiastic person who loves writing and wants to share my knowledge and understanding with you.