Cargando…

MON-611 Using Machine Learning on Electronic Health Records to Predict Inpatient Glucose Levels and Physicians’ Insulin Dosing

The current optimal inpatient diabetes management schema involves administration of basal, prandial, and correctional insulin to maintain blood glucose (BG) within a target range. Nonetheless, practical management often fails to reach the ideal in both insulin dosing regimens and patients’ BG outcom...

Descripción completa

Detalles Bibliográficos
Autores principales: Jankovic, Ivana, Liu, Xiran, Chen, Jonathan H
Formato: Online Artículo Texto
Lenguaje:English
Publicado: Oxford University Press 2020
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7207564/
http://dx.doi.org/10.1210/jendso/bvaa046.044
_version_ 1783530634690428928
author Jankovic, Ivana
Liu, Xiran
Chen, Jonathan H
author_facet Jankovic, Ivana
Liu, Xiran
Chen, Jonathan H
author_sort Jankovic, Ivana
collection PubMed
description The current optimal inpatient diabetes management schema involves administration of basal, prandial, and correctional insulin to maintain blood glucose (BG) within a target range. Nonetheless, practical management often fails to reach the ideal in both insulin dosing regimens and patients’ BG outcomes. Given the challenges of achieving adequate BG control for hospitalized patients using guidelines and expert knowledge alone, we attempted to use machine learning methods to predict (1) individual BGs, (2) average daily BGs, and (3) physician-ordered insulin doses based on data in an electronic health record-based repository between January 2014 and December 2018. We considered inpatients on subcutaneous insulin having a BG ≥ 200 mg/dL or ≤ 70 mg/dL or with an A1c percentage ≥ 6.5%. We excluded those missing critical data (such as weight), with fewer than five BG checks in 72 hours, or those on hemodialysis, resulting in a cohort of 3,461 patients with 175,934 BG checks among them. In this cohort, the average age was 61.4 years, the average A1c was 7.1%, and the average BG was 171.6 mg/dL, with approximately 25% of BGs ≥ 200 mg/dL and 1.7% of BGs < 70 mg/dL. Using linear regression, we identified features that contributed most to prediction of each of the outcomes. For all three outcomes, the average glucose in the past 24 hours was the most important feature. For prediction of glucose levels, previous BG, BG at the same time the previous day, A1c, BG variance, recent long-acting insulin dose, and glucocorticoid dose were all in the top 10 features. Similar features were important for predicting physician-ordered insulin doses. Surprisingly, neither weight nor creatinine were identified as top features for any outcome. Using these features in our predictive model, we found that individual BGs were highly erratic and could not be predicted precisely (R(2) 0.24). Similarly, and perhaps unsurprisingly, how physicians would order insulin for patients was also difficult to predict (R(2) 0.25). However, average daily glucose levels were predicted more reliably (R(2) 0.36), as was prediction of frank hyperglycemia (BG ≥ 200 mg/dL) in the next day (sensitivity 0.73, specificity 0.79). Given the typical practice pattern of a clinician evaluating the previous day’s insulin regimen performance and adjusting it by anticipating BGs for the next day, prediction of hyperglycemia in the next 24 hours can support decision-making for inpatient BG management.
format Online
Article
Text
id pubmed-7207564
institution National Center for Biotechnology Information
language English
publishDate 2020
publisher Oxford University Press
record_format MEDLINE/PubMed
spelling pubmed-72075642020-05-13 MON-611 Using Machine Learning on Electronic Health Records to Predict Inpatient Glucose Levels and Physicians’ Insulin Dosing Jankovic, Ivana Liu, Xiran Chen, Jonathan H J Endocr Soc Diabetes Mellitus and Glucose Metabolism The current optimal inpatient diabetes management schema involves administration of basal, prandial, and correctional insulin to maintain blood glucose (BG) within a target range. Nonetheless, practical management often fails to reach the ideal in both insulin dosing regimens and patients’ BG outcomes. Given the challenges of achieving adequate BG control for hospitalized patients using guidelines and expert knowledge alone, we attempted to use machine learning methods to predict (1) individual BGs, (2) average daily BGs, and (3) physician-ordered insulin doses based on data in an electronic health record-based repository between January 2014 and December 2018. We considered inpatients on subcutaneous insulin having a BG ≥ 200 mg/dL or ≤ 70 mg/dL or with an A1c percentage ≥ 6.5%. We excluded those missing critical data (such as weight), with fewer than five BG checks in 72 hours, or those on hemodialysis, resulting in a cohort of 3,461 patients with 175,934 BG checks among them. In this cohort, the average age was 61.4 years, the average A1c was 7.1%, and the average BG was 171.6 mg/dL, with approximately 25% of BGs ≥ 200 mg/dL and 1.7% of BGs < 70 mg/dL. Using linear regression, we identified features that contributed most to prediction of each of the outcomes. For all three outcomes, the average glucose in the past 24 hours was the most important feature. For prediction of glucose levels, previous BG, BG at the same time the previous day, A1c, BG variance, recent long-acting insulin dose, and glucocorticoid dose were all in the top 10 features. Similar features were important for predicting physician-ordered insulin doses. Surprisingly, neither weight nor creatinine were identified as top features for any outcome. Using these features in our predictive model, we found that individual BGs were highly erratic and could not be predicted precisely (R(2) 0.24). Similarly, and perhaps unsurprisingly, how physicians would order insulin for patients was also difficult to predict (R(2) 0.25). However, average daily glucose levels were predicted more reliably (R(2) 0.36), as was prediction of frank hyperglycemia (BG ≥ 200 mg/dL) in the next day (sensitivity 0.73, specificity 0.79). Given the typical practice pattern of a clinician evaluating the previous day’s insulin regimen performance and adjusting it by anticipating BGs for the next day, prediction of hyperglycemia in the next 24 hours can support decision-making for inpatient BG management. Oxford University Press 2020-05-08 /pmc/articles/PMC7207564/ http://dx.doi.org/10.1210/jendso/bvaa046.044 Text en © Endocrine Society 2020. http://creativecommons.org/licenses/by-nc-nd/4.0/ This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs licence (http://creativecommons.org/licenses/by-nc-nd/4.0/), which permits non-commercial reproduction and distribution of the work, in any medium, provided the original work is not altered or transformed in any way, and that the work is properly cited. For commercial re-use, please contact journals.permissions@oup.com
spellingShingle Diabetes Mellitus and Glucose Metabolism
Jankovic, Ivana
Liu, Xiran
Chen, Jonathan H
MON-611 Using Machine Learning on Electronic Health Records to Predict Inpatient Glucose Levels and Physicians’ Insulin Dosing
title MON-611 Using Machine Learning on Electronic Health Records to Predict Inpatient Glucose Levels and Physicians’ Insulin Dosing
title_full MON-611 Using Machine Learning on Electronic Health Records to Predict Inpatient Glucose Levels and Physicians’ Insulin Dosing
title_fullStr MON-611 Using Machine Learning on Electronic Health Records to Predict Inpatient Glucose Levels and Physicians’ Insulin Dosing
title_full_unstemmed MON-611 Using Machine Learning on Electronic Health Records to Predict Inpatient Glucose Levels and Physicians’ Insulin Dosing
title_short MON-611 Using Machine Learning on Electronic Health Records to Predict Inpatient Glucose Levels and Physicians’ Insulin Dosing
title_sort mon-611 using machine learning on electronic health records to predict inpatient glucose levels and physicians’ insulin dosing
topic Diabetes Mellitus and Glucose Metabolism
url https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7207564/
http://dx.doi.org/10.1210/jendso/bvaa046.044
work_keys_str_mv AT jankovicivana mon611usingmachinelearningonelectronichealthrecordstopredictinpatientglucoselevelsandphysiciansinsulindosing
AT liuxiran mon611usingmachinelearningonelectronichealthrecordstopredictinpatientglucoselevelsandphysiciansinsulindosing
AT chenjonathanh mon611usingmachinelearningonelectronichealthrecordstopredictinpatientglucoselevelsandphysiciansinsulindosing