Indicators for predicting continuous insulin infusion therapy-related hypokalemia: multicenter retrospective cohort study

Abstract

Hypokalemia is a common and potentially life-threatening complication of continuous intravenous insulin infusion (CII) in patients with hyperglycemic crises. However, no simple quantitative indicator can estimate the risk of hypokalemia at treatment initiation. This multicenter retrospective cohort study enrolled patients hospitalized for hyperglycemic crises who received CII. Clinical data available at CII initiation were collected. Machine-learning techniques were primarily employed for feature selection and model comparison to develop a simple, clinically interpretable prediction model. A logistic regression–based indicator was constructed using the most contributory variables and validated internally and externally. Hypokalemia was defined as a serum potassium level <3.5 mmol/L. The model-building and external validation cohorts included 99 and 55 patients, respectively. Among multiple candidate models, a lightweight logistic regression model using only two variables, serum potassium level and insulin infusion rate per body weight, was selected for clinical applicability. In internal validation, the model demonstrated good discriminative performance (receiver operating characteristic-area under the curve [ROC-AUC] 0.821). When applied to the external validation cohort, the ROC-AUC was 0.655, and accuracy decreased slightly. A lower threshold may increase sensitivity in screening, whereas a higher threshold may improve specificity. In conclusion, this study presents a simple and clinically applicable indicator for predicting CII-related hypokalemia using only two routinely available variables at treatment initiation. This model supports individualized electrolyte monitoring during the acute management of hyperglycemic crises. Additionally, it may facilitate safer and more efficient clinical decision-making without reliance on complex algorithms.

Graphical Abstract

Introduction

Hyperglycemic crises include diabetic ketoacidosis (DKA) and hyperosmolar hyperglycemic state (HHS), which are acute metabolic disturbances characterized by hyperglycemia and absolute or relative insulin deficiency [1]. They are among the most serious complications of diabetes. Treatment for hyperglycemic crises involves continuous intravenous insulin therapy (CII) and extracellular fluid infusion [2]. CII is an essential treatment for improving the diabetic crisis; however, improper treatment has various treatment-related complications [3]. Hypokalemia is one of the most critical complications requiring vigilance after initiating CII [4]. After high-dose insulin administration, potassium is taken up into cells along with glucose. To ensure adequate insulin action without complications, such as cardiac arrest or muscle weakness caused by hypokalemia, clinicians must implement electrolyte monitoring [5]. Among pediatric patients admitted in the intensive care unit (ICU) receiving DKA treatment, 81% experienced hypokalemia, with risk increasing with higher insulin doses and longer duration of administration [6]. Furthermore, the prevalence of hypokalemia during insulin therapy is higher among patients with predisposing factors, such as the use of loop and thiazide diuretics, increased sodium loading to the distal tubule, or increased mineralocorticoid activity [7, 8]. However, no simple indicator can predict CII-related hypokalemia before the treatment, particularly in adults.

We attempted to identify predictive indicators for CII-related hypokalemia using medical record data in patients hospitalized at our institution for hyperglycemic crisis. The results suggested that the insulin infusion rate per body weight and serum potassium levels at treatment initiation could predict hypokalemia [9]. However, owing to the single-center retrospective setting of the study, the predictive performance of the model and generalizability of the results were not adequately evaluated. Thus, in the present study, we aimed to identify predictive indicators for CII-related hypokalemia and assess their predictive accuracy using external data. The purpose of this study was not to develop a sophisticated machine-learning algorithm but to establish a simple, clinically applicable indicator for individualized electrolyte monitoring during CII.

Materials and methods

Study participants

The model-building cohort initially comprised 103 patients who received CII during hospitalization at Kawasaki Medical School Hospital between April 1, 2010, and October 31, 2024. Patients with normal serum osmolality and no acidosis (n = 4) were excluded. Ultimately, 99 patients with hyperglycemic crisis who received CII were included in the model-building cohort. Furthermore, 55 patients with hyperglycemic crisis who received CII during admission to Kawasaki Medical School General Medical Center between April 1, 2010, and October 31, 2024, formed the external validation cohort.

Study design

This multicenter, retrospective cohort study analyzed data collected from medical records. Data collection methods followed our previously published protocol [10]. Age, height, weight, body mass index (BMI), and vital signs were recorded on admission. Comorbidities, duration of diabetes mellitus, and history of smoking and alcohol consumption were obtained during interviews. Blood tests were performed before CII initiation. The rate, type, and dose of insulin infusion during CII were at the discretion of the attending physician. Blood glucose levels at each point were measured using arterial blood gas analysis, blood tests, or Glucotest Mint (Sanwa Chemical Laboratory Co., Ltd., Japan). The timing of rechecking serum potassium levels was determined by the attending physician. The lowest potassium level was assessed from CII initiation until resuming oral intake or discontinuing CII.

Construction of machine-learning models

The primary objective of this study was to develop a classification task model that can predict CII-related hypokalemia as a target for clinical models. Machine learning was primarily used for feature selection and model comparison, with the ultimate goal of developing a clinically interpretable, easily applicable model. In this study, hypokalemia was defined as serum potassium <3.5 mmol/L. The dataset used contained no missing values. The model-building cohort was randomly split into 80% for training (n = 79) and 20% for internal validation (n = 20). To construct a predictive model for CII-related hypokalemia, features were initially selected using machine-learning algorithms. The following clinical information available at CII initiation were collected: age, sex, height, weight, BMI, serum potassium, plasma osmolarity, alanine aminotransferase, aspartate aminotransferase, creatinine, blood urea nitrogen (BUN), estimated glomerular filtration rate (eGFR), C-reactive protein (CRP), blood glucose, insulin flow rate per body weight, and drip flow rate. With these variables as features, a clinical model predicting CII-related hypokalemia was constructed using the random forest method. Based on SHapley Additive exPlanations (SHAP) values [11], the top seven features contributing to prediction, namely, insulin flow rate per body weight, sex, height, eGFR, serum potassium, BMI, and CRP, were selected as features for clinical model construction (Supplementary Fig. 1). A clinical model was subsequently constructed to predict CII-related hypokalemia using four methods: decision trees, logistic regression, random forests, and gradient boosting decision trees (GBDT), employing the above seven features. Using the technique with the highest accuracy, a lightweight model utilizing the most important features was then built. Model performance was evaluated by accuracy, sensitivity, specificity, precision, recall, and F1 score. During model construction, class imbalance was corrected, and oversampling was performed using the Synthetic Minority Over-sampling Technique (SMOTE). To avoid data leakage, SMOTE was applied only to the training dataset after splitting the data. The model was applied to the data in the external validation cohort to assess its performance on external data. Python version 3.12.13 and scikit-learn version 1.6.1 were used.

Statistical analysis

Data are presented as median (interquartile range). Differences in the clinical characteristics between the model-building cohort and the external validation cohort were analyzed using the Mann–Whitney U and chi-square tests. Significance was defined as p < 0.05.

Ethical statement

The study protocol, including opt-out informed consent, was approved by the Institutional Review Board of Kawasaki Medical School (No. 6352-01). This study was conducted in accordance with the Declaration of Helsinki. Owing to the retrospective design, informed consent was not obtained from each patient; instead, study details were disclosed on each institution’s hospital website.

Results

Clinical characteristics

Table 1 presents the clinical characteristics of the model-building and external validation cohorts. No differences in sex or glycated hemoglobin (HbA1c) were noted between the two groups (p = 0.482, p = 0.588). Furthermore, CII-related hypokalemia occurred in 31.3% of the patients in the model-building cohort and 29.1% in the external validation cohort, showing no difference (p = 0.774). The insulin flow rate per body weight at treatment initiation was 0.09 (0.03–0.18) units/kg/h in the model-building cohort and 0.09 (0.04–0.13) units/kg/h in the external validation cohort, showing no difference (p = 0.891 and p = 0.605, respectively). Compared with the model-building cohort, the external validation cohort had significantly higher BUN (p = 0.007) and substantially lower eGFR (p = 0.012). The external validation cohort also had a longer duration of diabetes (p = 0.031).

Table 1. Clinical characteristics of the subjects in this study.

Parameters	Model-building cohort (n = 99)	External validation cohort (n = 55)	p value
Male/female	58/41	29/26	0.482
Age (years)	61 (42–73)	67 (54–76)	0.079
Duration of diabetes (years)	5 (0–15)	11 (4–20)	0.031
Type of diabetes (type 1/type 2, n)	30/69	15/40	0.692
Type of diabetic crisis (DKA/HHS, n)	71/28	36/19	0.379
Incidence of CII-related hypokalemia (n (%))	31 (31.3)	16 (29.1)	0.774
Length of hospital day (days)	16 (13–22)	18 (8–30)	0.750
Insulin therapy on admission (n (%))	25 (25.3)	20 (36.4)	0.146
Insulin therapy at discharge (n (%))	65 (65.6)	44 (80.0)	0.359
Height (cm)	162 (153–169)	160 (154–167)	0.639
Body weight (kg)	56 (45–71)	52 (46–65)	0.525
Body mass index (kg/m2)	22.4 (18.3–25.5)	21.1 (19.0–23.9)	0.418
Systolic blood pressure (mmHg)	134 (110–149)	128 (110–158)	0.970
Diastolic blood pressure (mmHg)	79 (66–91)	80 (60–95)	0.757
Pulse rate (beats per minute)	104 (88–118)	100 (90–120)	0.798
Body temperature (°C)	36.6 (36.2–37.1)	36.7 (36.1–37.3)	0.510
Blood glucose level at start of CII (mg/dL)	702 (500–867)	798 (577–1,003)	0.049
HbA1c (%, NGSP)	11.4 (9.6–13.2)	11.3 (9.2–13.3)	0.588
Albumin (g/dL)	3.9 (3.5–4.4)	3.9 (3.4–4.2)	0.825
AST (U/L)	23 (15–34)	27 (19–34)	0.171
ALT (U/L)	22 (16–39)	21 (15–42)	0.908
Blood urea nitrogen (mg/dL)	38 (19–54)	45 (20–75)	0.007
Creatinine (mg/dL)	1.2 (0.8–1.8)	1.5 (0.8–2.1)	0.052
eGFR (mL/min/1.73 m2)	47 (29–47)	34 (20–56)	0.012
Uric acid (mg/dL)	9.2 (6.3–12.5)	9.1 (6.5–13.0)	0.764
C-reactive protein (mg/dL)	1.02 (0.32–3.00)	0.96 (0.28–3.64)	0.989
Serum sodium (mmol/L)	134 (129–140)	134 (126–142)	0.803
Serum potassium (mmol/L)	4.8 (4.2–5.5)	4.9 (4.5–5.5)	0.282
Serum chloride	96 (89–103)	94 (87–104)	0.731
Plasma osmolarity (mOsm/kg)	328 (304–348)	326 (312–371)	0.239
Total ketone body (μmol/L)	7,265 (1,887–13,441)	7,733 (2,322–13,775)	0.961
Acetoacetic acid (μmol/L)	2,125 (603–3,823)	1,873 (618–3,500)	0.477
3-hydroxybutyric acid (μmol/L)	5,214 (1,158–9,810)	4,450 (1,521–10,515)	0.783
pH	7.32 (7.20–7.37)	7.31 (7.13–7.37)	0.709
Bicarbonate ion concentration (mEq/L)	13.7 (7.1–21.1)	18 (8.4–23.1)	0.175
Base excess	–19.7 (–18.6– –3.0)	–7.3 (–19.1– –2.2)	0.426
Insulin flow rate at the initiation of CII (units/kg/hr)	0.09 (0.03–0.18)	0.09 (0.04–0.13)	0.891
Drip flow rate at the initiation of CII (units/kg/hr)	6.3 (2.6–9.7)	7.3 (3.4–9.9)	0.605

Data presented as median (interquartile range). DKA, diabetic ketoacidosis; HHS, hyperosmolar hyperglycemic state; CII, continuous intravenous insulin therapy; AST, aspartate aminotransferase; ALT, alanine aminotransferase; eGFR, estimated glomerular filtration rate. The Wilcoxon rank-sum test and chi-square test were used for analysis.

Model-building methods and feature selection

To select features for predicting CII-related hypokalemia, seven variables were extracted based on SHAP values (Supplementary Fig. 1). Four clinical models were constructed using these features (Fig. 1). The prediction accuracy of the decision tree (Fig. 1A), random forest, and GBDT models for CII-related hypokalemia was 0.60. Conversely, the logistic regression model showed a high accuracy of 0.75 (Fig. 1B). To build a lightweight model, insulin flow rate per body weight and serum potassium had high contribution rates among the features used in each predictive model. This result was consistent with the findings of our previous study [9]. Based on these findings, a logistic regression model to predict CII-related hypokalemia was built using these two features.

Fig. 1. Construction of machine-learning models to predict CII-related hypokalemia using seven features. The construction methods include (A) decision tree, (B) logistic regression, (C) random forest, and (D) gradient boosting decision tree.

Development of a clinical prediction model for CII-related hypokalemia

The calculation formula and receiver operating characteristic (ROC) curve for the logistic regression model constructed using the two features are shown in Fig. 2A, B. The ROC-area under the curve (AUC) for the logistic regression model predicting CII-related hypokalemia was 0.821. When the threshold for y was set to 0.35, recall for CII-related hypokalemia was 1.00. Still, precision was 0.35, the F1 score was 0.52, and the false-positive rate was high (Fig. 2C). In contrast, setting the y threshold to 0.5 improved the prediction accuracy, yielding an accuracy of 0.70 and F1 score of 0.62 (Fig. 2D). However, one case of hypokalemia was misclassified. These results suggest the potential of logistic regression model for predicting CII-related hypokalemia as a screening test with a threshold of 0.35 and high specificity at 0.5.

Fig. 2. Construction of logistic regression classification model to predict CII-related hypokalemia using machine learning with two features. (A) Calculation formula for the constructed logistic regression classification model. (B) ROC curve for the machine-learning model. (C) Confusion matrix when the threshold for y is set to 0.35. (D) Confusion matrix when the threshold for y is set to 0.5. The vertical axis of the confusion matrix represents the actual presence or absence of hypokalemia, whereas the horizontal axis represents the predicted CII-related hypokalemia label. In the table, “Hypo K” denotes the presence of CII-related hypokalemia, and “No hypo K” denotes its absence.

Fig. 3 shows the results when the model is applied to an external validation cohort. The ROC-AUC is 0.655 (Fig. 3A). In external validation, setting the threshold for y to 0.35 yields an accuracy of 40.0% and a sensitivity of 81.3%, comparable with the accuracy in internal validation (Fig. 3B). Furthermore, when the threshold for y is set to 0.5, the accuracy is 67.3% and the specificity is 74.4%, showing an increase in specificity compared with the threshold of 0.35 (Fig. 3C).

Fig. 3. Validation of predictive accuracy in an external validation cohort. (A) ROC curve. (B) Confusion matrix with a threshold of 0.35 for y. (C) Confusion matrix with a threshold of 0.5 for y. This clinical model predicts CII-related hypokalemia. The vertical axis of the confusion matrix represents the actual presence or absence of hypokalemia, whereas the horizontal axis represents the predicted CII-related hypokalemia label. In the table, “Hypo K” denotes the presence of CII-related hypokalemia, and “No hypo K” denotes its absence.

The decision curve analysis (DCA) demonstrated that the two-variable model provided a net clinical benefit across a wide range of threshold probabilities compared with treat-all or treat-none strategies (Supplementary Fig. 2A). Importantly, increasing model complexity did not lead to a meaningful improvement in net benefit, which supports the use of a parsimonious model in clinical practice. In the external validation cohort, DCA showed that the two-variable model provided a net clinical benefit within a clinically relevant range of threshold probabilities, particularly around the decision thresholds used in this study (Supplementary Fig. 2B). Despite the modest magnitude of the net benefit and some variability, increasing model complexity did not result in a consistent improvement, supporting the robustness of the parsimonious model.

Discussion

In this study, a clinically interpretable prediction model in which machine-learning techniques were used to assist feature selection was developed (Graphical Abstract). The clinically interpretable model readily identified cases warranting electrolyte monitoring. Known risk factors for CII-related hypokalemia include low serum potassium levels at CII initiation, high HbA1c levels, severe acidosis, and renal impairment [12, 13]. Furthermore, insulin administration promotes potassium uptake into cells and increases potassium excretion by the kidneys [14]. Consequently, high-dose insulin therapy and long-term insulin administration carry a risk of hypokalemia in patients with DKA treated in the pediatric care unit [6]. Serum potassium levels must be measured before initiating insulin therapy, and insulin administration be delayed and potassium correction prioritized if potassium levels are <3.3 mmol/L [13, 15]. Conversely, no internationally established quantitative indicators or scoring systems have been established for predicting CII-related hypokalemia. This study proposes a simple indicator for predicting CII-related hypokalemia using two variables: serum potassium levels, which are always measured at CII initiation, and insulin dosage per body weight. The model is expected to help rapidly identify patients who require electrolyte monitoring during the acute phase of a hyperglycemic crisis.

Graphical Abstract.

The constructed machine-learning models employed lightweight designs using minimal features to enable prompt treatment initiation for hyperglycemic crisis and rapid identification of patients requiring electrolyte monitoring. Serum potassium levels, the first feature, have been reported as a primary predictor of CII-related hypokalemia in American Diabetes Association guidelines and other clinical studies [6, 13, 15], making them a reasonable factor. Insulin flow rate per unit of body weight, the second feature, is a known risk factor for hypokalemia in patients with diabetes mellitus and is clinically valid [6, 14]. The clinical significance of this model lies in its ability to predict outcomes in patients experiencing hyperglycemic crises, specifically those with either DKA or HHS. In DKA and HHS, pre-CII serum potassium levels and insulin dosage are known risk factors for CII-related hypokalemia. However, the mechanisms of CII-related hypokalemia are different between the two conditions. In DKA, the severity of metabolic acidosis and acid–base imbalance due to renal dysfunction increases the risk, whereas HHS is characterized by extreme dehydration and potassium deficiency [12, 15]. When HHS and DKA overlap, distinguishing between them at treatment initiation is challenging [9, 16]. In such cases, this clinical model, which can preemptively identify cases requiring more careful electrolyte monitoring based on fewer characteristics, is expected to be useful.

Although experienced physicians may intuitively recognize the risk of hypokalemia based on serum potassium levels and insulin infusion rate, such assessments are often subjective and may vary depending on clinical experience. The proposed model provides a simple and objective framework based on these key variables, allowing for standardized risk assessment across different clinical settings, including less-experienced clinicians or resource-limited environments. The addition of more variables resulted in only minimal improvements in predictive performance, supporting the use of a parsimonious model for clinical application (Supplementary Tables 1, 2). Importantly, our findings present that increasing model complexity does not substantially improve predictive performance, supporting the use of a parsimonious model that closely aligns with clinically intuitive factors while enhancing reproducibility. To evaluate the benefits of performing electrolyte monitoring using the constructed clinical model, a DCA was conducted. The results showed that internal and external validation demonstrated significant benefits when monitoring was decided upon for high-risk cases (Supplementary Fig. 2A, B). These results help objectively visualize factors that experienced physicians have traditionally assessed intuitively for monitoring. However, as a retrospective observational study, it cannot determine whether using the model improves clinical decision-making. Furthermore, given that hypokalemia is a common CII-related complication, minimizing missed cases remains essential. However, if the proposed clinical model were directly integrated into electronic medical records or medical devices in the ICU or similar setting to generate automatic alerts regardless of the clinician’s intentions, a threshold of 0.35 could contribute to alarm fatigue. With further clinical validation of this model and improved predictive accuracy, the use of this threshold as a reference for clinical decision-making may help clinician actively assess the risk of CII-related hypokalemia while minimizing alert fatigue.

This study has several limitations. First, it is a retrospective observational study conducted at two facilities in Japan. Thus, further validation is needed to determine whether similar predictive accuracy can be maintained in different ethnic groups, regions, and healthcare systems. Because the external validation data included a population in which the incidence of CII-related hypokalemia was low, accuracy validation in populations with different characteristics is necessary. Second, hyperglycemic emergencies are rare conditions, resulting in a limited sample size for analysis. The decrease in accuracy in external validation compared with that in internal validation may be related to overfitting or differences in sample size. Sub-analyses of cases misclassified at two different thresholds could not be performed owing insufficient statistical power. Further improvement in predictive accuracy is necessary for clinical application. Finally, no clear protocol has been established for the timing of serum potassium retesting or insulin dosage, which were determined at the attending physician’s discretion. In cases with infrequent electrolyte monitoring, hypokalemia may have gone undetected.

Despite these limitations, the CII-related hypokalemia prediction model developed in this study is a highly readable machine-learning model. In this study, we intentionally prioritized interpretability and clinical usability over model complexity to support real-time decision-making rather than algorithmic optimization. The model holds potential as a suitable system for early detection of cases requiring more focused electrolyte monitoring. Although the model demonstrated good discrimination in internal validation, its moderate performance in external validation suggests the need for recalibration or refinement when applying it in different clinical settings. Systems such as the constructed clinical model are essential for enabling decision-making based on objective indicators rather than on conventional electrolyte monitoring, which relies on physician experience. Further investigation is warranted to improve predictive accuracy regarding CII-related hypoglycemia.

Authors’ contribution

Y.I. researched data and performed the analysis using the Python language and wrote the manuscript. Ha.I, F.T., Y.K., and F.K. collected data from Kawasaki Medical School General Medical Center. T.K., M.K., Y.O., R.I., T.I., T.S., K.D., Hi.I., Y.F., J.S., Ha.I., F.T., Y.K., F.K., M.S., S.N., K.K. participated in the discussion. H.K. participated in the discussion and reviewed the manuscript.

Funding

There is no funding associated with this manuscript.

Conflict of interests

H.K. has received honoraria for lectures, received scholarship grants from Novo Nordisk Pharma, Sanofi, Eli Lilly, Boehringer Ingelheim, Sumitomo Pharma, Daiichi Sankyo, Mitsubishi Tanabe Pharma, Manpei Suzuki Diabetes Foundation, Japan Arteriosclerosis prevention fund, Japan Association for Diabetes Education and Care. K.K. has been an advisor to, received honoraria for lectures from, and received scholarship grants from Novo Nordisk Pharma, Sanwa Kagaku, Taisho Pharma, Sumitomo Pharma, Astellas, Boehringer Ingelheim. S.N. has received honoraria for lectures from Novo Nordisk Pharma, Kowa and Daiichi Sankyo. T.K. has received honoraria for lectures from Sumitomo Pharma and Novo Nordisk Pharma. All other authors have no conflict of interests.

Data availability

All data generated or analyzed during this study are included in this published article. The code generated by the artificial intelligence in this study is pre-clinically applied and not publicly available.

Supplementary Material

Supplementary Tables

Supplementary Figs.