Predicting the Risk of Inpatient Hypoglycemia With Machine Learning Using Electronic Health Records
Article Figures & Tables
Figures
- Figure 1
ROC curves for logistic regression, XGBoost, and decision tree model when predicting biochemical hypoglycemia.
Tables
- Table 1
Potential predictors and how they are represented in the EPR and in the prediction models
Category Predictor Data in EPR Data in models Unit Completeness (%) Data sets Demographics Age Year of birth Computed based on the year of admission years 100 IH Sex Male/female Binary variable (1/0) NA 100 IH Ethnicity Ethnicity (categorical value) Categorical variable (white British, African, etc.) NA 100 IH Weight Weight measured at time of admission Weight value kg 71 IH Height Height measured at time of admission Height value cm 59 IH Type of diabetes Type of diabetes (categorical value) Categorical variable (T1D/T2D/other) NA 100 IH Vital signs Diastolic blood pressure Multiple measurements Average value throughout the admission mmHg 73 IH Systolic blood pressure Multiple measurements Average value throughout the admission mmHg 73 IH Heart rate Multiple measurements Average value throughout the admission /min 71 IH Oxygen saturation Multiple measurements Average value throughout the admission % 73 IH Temperature Multiple measurements Average value throughout the admission Celsius 72 IH Laboratory tests Albumin Multiple measurements Average value throughout the admission g/L 81 IH Amylase Multiple measurements Average value throughout the admission IU/L 15 IH C-peptide Multiple measurements Average value throughout the admission pmol/L 17 IH Cortisol Multiple measurements Average value throughout the admission nmol/L 26 IH Creatinine Multiple measurements Average value throughout the admission μmol/L 80 IH C-reactive protein Multiple measurements Average value throughout the admission mg/L 78 IH eGFR Multiple measurements Average value throughout the admission mL/min/1.73 m2 80 IH Hemoglobin Multiple measurements Average value throughout the admission g/L 80 IH HbA1c Multiple measurements Average value throughout the admission % 42 IH Potassium Multiple measurements Average value throughout the admission mmol/L 80 IH Sodium Multiple measurements Average value throughout the admission mmol/L 80 IH White cells Multiple measurements Average value throughout the admission × 109/L 79 IH Medications Sulfonylurea Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH DPP-4 Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH GLP-1 Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Metformin Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Morphine Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Pioglitazone Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Bisoprolol Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Amitriptyline Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Pregabalin Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Dexamethasone Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Prednisolone Drug dose and time Binary variable (1 for on drug and 0 for not) NA 100 IH Intravenous insulin Multiple rates of insulin infusion Binary variable (1 for on i.v. insulin and 0 for not) NA 100 IH+ Insulin (rapid-acting analog) Multiple doses of different amount Average total daily insulin dose unit 100 IH+ Insulin (mixed analog) Multiple doses of different amount Average total daily insulin dose unit 100 IH+ Insulin (long-acting analog) Multiple doses of different amount Average total daily insulin dose unit 100 IH+ Insulin (short-acting human) Multiple doses of different amount Average total daily insulin dose unit 100 IH+ Insulin (mixed human) Multiple doses of different amount Average total daily insulin dose unit 100 IH+ Insulin (intermediate-acting human) Multiple doses of different amount Average total daily insulin dose unit 100 IH+ Procedures Procedure indication Procedure name and time Binary variable (1 for had at least one procedure during the admission and 0 for not) NA 100 IH+ Previous hypoglycemia Previous biochemical hypoglycemia Blood glucose measurements Binary variable (1 for had at least one blood glucose <4 mmol/L) NA 63 PH Previous clinically significant hypoglycemia Blood glucose measurements Binary variable (1 for had at least one blood glucose <3 mmol/L) NA 63 PH DPP-4, dipeptidyl peptidase 4; GLP-1, glucagon-like peptide 1; NA, not applicable; T1D, type 1 diabetes; T2D, type 2 diabetes.
Insulin (rapid-acting analog): “Insulin aspart,” “Insulin lispro,” “Insulin glulisine,” “Insulin faster acting aspart.”
Insulin (mixed analog): “Insulin aspart biphasic (Novomix 30),” “Insulin lispro biphasic (Humalog Mix 25 and Humalog Mix 50).”
Insulin (long-acting analog): “Insulin glargine,” “Insulin detemir,” “Insulin degludec.”
Insulin (short-acting human): “Insulin Actrapid,” “Insulin Humulin S.”
Insulin (mixed human): “Insulin Humulin M3.”
Insulin (intermediate-acting human): “Insulin Insulatard,” “Insulin Humulin I.”
- Table 2
Baseline characteristics and glycemic outcomes of the inpatients cohorts
Predictors Inpatients with diabetes (N = 17,658) Inpatient hospital admissions (n = 32,758) Sex, N (%) Female 8,381 (47) Male 9,277 (53) Age, mean (SD) 66 (18) Ethnicity, N (%) White British 12,511 (70.8) African 116 (0.7) Pakistani 331 (1.9) Chinese 53 (0.3) Indian 254 (1.4) Not stated 2,869 (16.2) Other 1,524 (8.6) Type of diabetes, N (%) Insulin-dependent diabetes 1,696 (9.6) Non–insulin-dependent diabetes 14,006 (79.3) Other forms 1,956 (11.1) Systolic blood pressure, mean (SD) 132.5 (18.2) eGFR, mean (SD) 29.8 (6.4) Medication use Sulfonylurea, n (%) 6,435 (19.6) DPP-4, n (%) 1,415 (4.3) GLP-1, n (%) 349 (1.1) Metformin, n (%) 10,756 (32.8) Insulin, n (%) Intravenous insulin 4,678 (14.3) Rapid-acting analog 3,954 (12.1) Mixed-acting analog 1,553 (4.7) Long-acting analog 5,118 (15.6) Short-acting human 3,561 (10.9) Mixed-acting human 1,388 (4.2) Intermediate-acting human 2,394 (7.3) Procedures, n (%) 22,931 (70.0) Glycemic outcomes Hypoglycemia, n (%) Biochemical hypoglycemia 7,030 (21.5) Clinically significant hypoglycemia 3,154 (9.6) BG level, mean (SD) 10.1 (4.7) N (%), number of patients and percentage over the total number of patients; n (%), number of admissions and percentage over the total number of admissions. DPP-4, dipeptidyl peptidase 4; GLP-1, glucagon-like peptide 1. Insulin (rapid-acting analog): “Insulin aspart, ” “Insulin lispro, ” “Insulin glulisine, ” “Insulin faster acting aspart. ” Insulin (mixed analog): “Insulin aspart biphasic (Novomix 30), ” “Insulin lispro biphasic (Humalog Mix 25 and Humalog Mix 50). ” Insulin (long-acting analog): “Insulin glargine, ” “Insulin detemir, ” “Insulin degludec. ” Insulin (short-acting human): “Insulin Actrapid, ” “Insulin Humulin S. ” Insulin (mixed human): “Insulin Humulin M3. ” Insulin (intermediate-acting human): “Insulin Insulatard, ” “Insulin Humulin I. ”
- Table 3
Performance metrics of the machine learning models based on the PH data set
Machine learning algorithm Biochemical hypoglycemia (BG <4 mmol/L) Clinically significant hypoglycemia (BG <3 mmol/L) AUROC Precision Recall AUROC Precision Recall Logistic regression 0.73 0.48 0.10 0.75 0.39 0.10 SGD 0.74 0.12 0.10 0.77 0.10 0.10 kNN 0.62 0.40 0.18 0.62 0.30 0.15 Decision tree 0.81 0.70 0.71 0.84 0.68 0.73 Gaussian-naive Bayes 0.81 0.47 0.68 0.86 0.33 0.81 Bernoulli-naive Bayes 0.82 0.60 0.60 0.86 0.47 0.67 Multinomial-naive Bayes 0.75 0.10 0.10 0.79 0.10 0.10 SVM 0.79 0.73 0.10 0.83 0.41 0.10 QDA 0.77 0.23 0.96 0.89 0.15 0.97 Random forest 0.94 0.86 0.67 0.93 0.96 0.66 Extra trees 0.93 0.85 0.68 0.93 0.94 0.66 LDA 0.88 0.69 0.75 0.90 0.72 0.72 Passive aggressive 0.76 0.46 0.25 0.77 0.33 0.10 AdaBoost 0.89 0.68 0.60 0.93 0.63 0.46 Bagging 0.93 0.84 0.70 0.92 0.93 0.67 Gradient boosting 0.96 0.87 0.70 0.96 0.96 0.67 XGBoost 0.96 0.88 0.70 0.96 0.97 0.67 MLP 0.74 0.57 0.17 0.78 0.47 0.14 Mean (SD) 0.82 (0.10) 0.59 (0.25) 0.49 (0.29) 0.85 (0.10) 0.55 (0.31) 0.48 (0.31) kNN, k-nearest neighbor; LDA, linear discriminant analysis; MLP, multilayer perceptron (artificial neural network); QDA, quadratic discriminant analysis; SGD, stochastic gradient descent; SVM, support vector machine.
- Table 4
Most significant predictors from the logistic regression model
Predictors Biochemical hypoglycemia (BG <4 mmol/L) Clinically significant hypoglycemia (BG <3 mmol/L) Coefficient P value z score Coefficient P value z score PrevLowGlucose3 (+) 3.842 <0.001 28.42 4.021 <0.001 20.39 Albumin level (−) −0.078 <0.001 −27.22 −0.074 <0.001 −19.51 Intravenous insulin (+) 0.639 <0.001 15.43 0.501 <0.001 9.82 Procedure indication (+) 0.485 <0.001 14.87 0.339 <0.001 6.81 Sulfonylurea (+) 0.572 <0.001 14.24 0.311 <0.001 5.35 Type 2 diabetes (−) −0.820 <0.001 −13.68 −0.656 <0.001 −7.88 Weight (−) −0.010 <0.001 −7.42 −0.012 <0.001 −6.38 Oxygen saturation (+) 0.059 <0.001 6.31 0.067 <0.001 5.30 Metformin (−) −0.212 <0.001 −6.02 −0.258 <0.001 −5.02 Long-acting human insulin (+) 0.011 <0.00 4.77 0.010 <0.001 5.35 Rapid-acting human insulin (+) 0.023 <0.001 4.11 NS Mixed insulin analog (+) 0.007 <0.001 3.67 NS Factors with a positive coefficient value increase the risk of hypoglycemia, and factors with a negative coefficient value decrease the risk of hypoglycemia. The factors are listed in order of effect size on the logistic regression model, e.g., an increase in albumin value reduces the risk of hypoglycemia, people with type 2 diabetes have an decreased risk of hypoglycemia. A “+” or “−” sign is given to each of the factors to indicate the effect direction. NS, not significant.
In this Issue
