Development and validation of a hypoglycemia risk prediction tool for hospitalized patients with type 2 diabetes mellitus treated with insulin

Abstract

BACKGROUND

Insulin is the preferred clinical treatment for hospitalized patients with type 2 diabetes mellitus (T2DM) to control blood glucose effectively. Hypoglycemia is one of the most common adverse events. Accurate prediction of the risk of hypoglycemia is critical in reducing hypoglycemic events and related adverse events in hospitalized diabetic patients treated with insulin.

AIM

To develop and validate a hypoglycemia risk prediction tool for hospitalized patients with T2DM treated with insulin.

METHODS

This retrospective study included 802 hospitalized patients with T2DM in the Department of Endocrinology, the Third Affiliated Hospital of Sun Yat-sen University, between January 2021 and December 2021. The hypoglycemia risk prediction model was developed using logistic regression and nomogram models. The model was validated and calibrated using receiver operating characteristic curves and the Hosmer-Lemeshow goodness of fit test.

RESULTS

The incidence of hypoglycemia among the enrolled patients was 44.9%. The hypoglycemic risk prediction model included six predictors: Body mass index, duration of diabetes, history of hypoglycemia within 1 year, glomerular filtration rate, blood triglyceride levels, and duration of treatment. The hypoglycemia risk prediction model displayed high discrimination ability (area under the curve = 0.67) and good calibration power (goodness of fit, χ² =12.25, P = 0.14).

CONCLUSION

The hypoglycemia risk prediction model for hospitalized patients with T2DM on insulin therapy displayed high reliability and discrimination ability. The model is a promising tool for clinicians to screen hospitalized patients with T2DM and an elevated risk of hypoglycemia and guide personalized interventions to prevent and treat hypoglycemia.

Key Words: Hypoglycemia; Risk prediction tool; Type 2 diabetes mellitus; Insulin

Core Tip: The hypoglycemia risk-prediction model was developed using the logistic regression and nomogram models. The model was validated and calibrated using the receiver operating characteristic curves and the Hosmer-Lemeshow goodness of fit test. The incidence of hypoglycemia was 44.9%. The model included eight independent hypoglycemia risk factors. The hypoglycemia risk prediction model for hospitalized T2 diabetes mellitus patients treated with insulin showed high reliability and discrimination ability.

INTRODUCTION

Diabetes mellitus (DM) is a metabolic disorder associated with alterations in pathways involved in carbohydrate, fat, and protein metabolism[1]. Type 2 DM (T2DM) accounts for more than 90% of all DM cases, and it is the primary contributor to the global disease burden[2,3]. Insulin is the recommended clinical treatment of hyperglycaemia for hospitalized patients with T2DM[1,4]. Hypoglycemia is a key barrier to achieving euglycemic control in hospitalized patients with T2DM[5]. Hypoglycemia refers to a blood glucose level ≤ 3.9 mmol/L for hospitalized patients with T2DM[6]. Inpatient hypoglycaemia has been linked to adverse clinical outcomes, including mortality, longer stay in hospital, and cost to the healthcare system[5]. Furthermore, long-term severe hypoglycemia can cause multiple-organ function damage, including cardiovascular and cerebrovascular injury, loss of consciousness, and even death[7-10]. Hypoglycemia is very common in hospitalized patients with T2DM[11-13]. The global hypoglycemia assessment tool (HAT) study reported that the incidence of hypoglycemia was 46.5% among hospitalized patients with T2DM who received insulin[11]. The NaDIA reports that in England and Wales, 18.4% of inpatients with diabetes experience at least one episode of hypoglycaemia[12]. A recent Chinese study reported that the prevalence of hypoglycemia in hospitalized patients with DM was 14.68%[13].

Accurate prediction of the risk of hypoglycemia is critical for reducing hypoglycemic events and related adverse events in hospitalized patients with DM. Moreover, an objective and straightforward tool for predicting the risk of hypoglycemia during insulin therapy will help clinicians identify patients at high risk for hypoglycemia during insulin treatment, thereby providing a reference for appropriate insulin dosages. Although several predictive models of hypoglycemia in hospitalized patients with T2DM have been developed, the results are inconsistent[5,14,15]. Thus, this study aimed to develop and validate a hypoglycemia risk prediction tool for hospitalized patients with T2DM treated with insulin.

MATERIALS AND METHODS

Ethical approval

The Ethics Committee of the Third Affiliated Hospital of Sun Yat-sen University (Guangzhou, China) approved this study. All participants provided written informed consent before participating in this study.

Inclusion and exclusion criteria

This retrospective study enrolled 802 hospitalized patients with T2DM who received insulin treatment between January 2021 and December 2021 at the Department of Endocrinology, Third Affiliated Hospital of Sun Yat-sen University.

Inclusion criteria: The inclusion criteria were as follows: Diagnosis of T2DM as defined by the World Health Organization in 1999; age ≥ 18 years; eligible for insulin intensive therapy (including newly diagnosed patients with glycated hemoglobin A1c (HbA1c) > 9.0%, fasting plasma glucose (FPG) at admission > 11.1 mmol/L, or obvious symptoms of hyperglycemia; patients with significant blood glucose elevation (HbA1c > 9.0%) after combined treatment with two or more oral hypoglycemic agents; or patients with HbA1c > 7.0% after adequate dose-adjusted initial insulin therapy); and ongoing intensive insulin therapy, including basal pre-prandial insulin injection regimen, continuous subcutaneous insulin infusion, and premixed insulin analog injection regimen three times daily.

Exclusion criteria: The exclusion criteria were as follows: Severe acute complications of DM; serious disease of the heart, brain, lungs, and/or kidneys; pregnancy; and surgical treatment.

Baseline predictor variables and outcome measures

Outcome measures: The diagnostic criteria for hypoglycemia in this study were venous plasma glucose or capillary blood glucose (peripheral blood glucose) levels of ≤ 3.9 mmol/L combined with other clinical symptoms including sense of hunger, palpitation, and trembling hands.

Baseline predictor variables: Based on the review of previous studies, the following clinical data were collected from the medical records of all hospitalized patients: General factors including sex, age, education, body mass index (BMI), blood pressure, smoking history, alcohol consumption, occupation, education level, and marital status; disease-related information, including duration of diabetes, diabetic nephropathy, diabetic peripheral nerve disease, diabetic peripheral vascular disease, hypertension, hyperlipidemia, fatty liver, HbA1c, fasting blood glucose (FBG), random blood glucose at admission, standard deviation of blood glucose (SDBG), fasting insulin, fasting C peptide, homeostasis model assessment for insulin resistance (HOMA-IR), homeostasis model assessment of β-cell function, aspartate aminotransferase, alanine aminotransferase (ALT), creatinine, glomerular filtration rate (GFR), total cholesterol, triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol, and C-reactive protein; treatment-related information, including treatment before admission, insulin injection regimen, insulin dose, type of combined oral drugs (antidiabetic drugs, antihypertensive drugs), and total treatment time; and hypoglycemic event-related information, including frequency of hypoglycemia, specific blood glucose levels, hypoglycemia classification (clinically significant hypoglycemia, blood glucose alert value, pseudo-hypoglycemia), time of hypoglycemia occurrence (before and after meals, before sleep, or night/early morning), hypoglycemia symptoms (asymptomatic, hunger, sweating, vertigo, palpitation, fatigue, hand shaking, and others), hypoglycemia inducement (unreasonable diet, unreasonable exercise, and other reasons), treatment methods [eating food, oral intake of 50% glucose solution (GS), intravenous injection of 50% GS, or oral 50% GS + food], duration of hypoglycemia, and FBG before the onset of hypoglycemia.

Statistical analysis

SPSS 25.0 software (IBM, Armonk, NY, United States) was used for statistical analysis. P < 0.05 (two-sided) was considered statistically significant. Continuous variables were expressed as the mean ± SD. Categorical variables were expressed as numbers, percentages, or frequencies. Intergroup differences were analyzed using the independent-samples t-test, χ² test, or Mann-Whitney U test. A backward stepwise logistic regression model was used to screen the independent hypoglycemia risk factors and construct a hypoglycemia risk prediction model. Receiver operating characteristic (ROC) curves and the area under the curve (AUC) were used to verify the prediction performance of the risk model. The Hosmer-Lemeshow test was used to verify and calibrate the model. Ten-fold cross-validation and AUCs were used to validate the hypoglycemia risk prediction model.

RESULTS

Incidence and characteristics of hypoglycemia in patients with T2DM treated with insulin

The incidence of hypoglycemia in this study cohort was 44.9%, and 716 hypoglycemic events were recorded in 360 of 802 patients included in this study (Table 1). Among the patients with hypoglycemia, 53.9% experienced two or more episodes. The median (interquartile range) time for the occurrence of hypoglycemia during insulin treatment was 6.5 (4-9) days. Therefore, hypoglycemia generally occurred on days 4-9 of insulin treatment. Hypoglycemia occurred most frequently before lunch in 187 patients (26.1%), after breakfast in 117 patients (16.3%), and before breakfast in 106 patients (14.8%). Asymptomatic hypoglycemia was observed in 455 patients (63.5%), whereas symptomatic hypoglycemia occurred in 261 patients (36.5%). The cause of 88% of hypoglycemic events was unknown. The leading causes of hypoglycemia included reduced food intake (58 cases), excessive exercise (11 cases), examination-related hypoglycemia (15 cases), and excessive insulin dosage (one case).

Table 1 Incidence and characteristics of hypoglycemia events in patients with type 2 diabetes mellitus treated with insulin (n = 716).

	n (%) or mean ± SD		n (%)/M (P25, P75)
Frequency of hypoglycemic events		FPG before hypoglycemia	5.80 (4.80, 7.30)
0	442 (55.1)	Days of insulin treatment when hypoglycemia occurred	6.50 (4.00, 9.00)
1	166 (20.7)	Occurrence time
2	108 (13.5)	Before breakfast	106 (14.8)
3 and more	86 (10.7)	After breakfast	117 (16.3)
Hypoglycemia (mmol/L)	3.53 ± 0.40	Before lunch	187 (26.1)
Symptoms of hypoglycemia		After lunch	86 (12.0)
No symptom	455 (63.5)	Before dinner	69 (9.6)
Sweating	67 (9.4)	After dinner	48 (6.7)
Fatigue	37 (5.2)	Before sleep	76 (10.6)
Palpitation	56 (7.8)	Midnight/before dawn	27 (3.8)
Sense of hunger	54 (7.5)	Hypoglycemic trigger
Hands tremble	19 (2.7)	None	630 (88.0)
Dizzy	69 (9.6)	Reduced food intake	58 (8.1)
Blurred vision	6 (0.8)	Excessive exercise	11 (1.5)
Types of hypoglycemia		Examination related	15 (2.1)
Clinically significant	81 (11.3)	Excessive insulin dosage	2 (0.3)
Blood glucose alert value	617 (86.2)	Methods for alleviating hypoglycemia
Relative hypoglycemia	18 (2.5)	Eating food	475 (66.3)
Duration of hypoglycemia		Oral GS	172 (24.0)
15 minutes	617 (86.2)	Intravenous injection of GS	10 (1.4)
30 minutes	87 (12.1)	Eating food + Oral GS	59 (8.2)
> 30 minutes	12 (1.7)

FPG: Fasting blood glucose; GS: Glucose solution.

Univariate analysis of risk factors for hypoglycemia

Based on the occurrence of hypoglycemia, the patients in this study were categorized into the non-hypoglycemic group (n = 442) and hypoglycemic group (n = 360). Our data illustrated that age, BMI, and occupation significantly differed between these groups (all P < 0.05). The hypoglycemic group had an older age and lower BMI than the non-hypoglycemic group. Furthermore, fewer subjects in the hypoglycemic group were employed. The two groups exhibited significant differences in the total treatment time and the combined oral hypoglycemic drugs used (both P < 0.05). However, systolic blood pressure, diastolic blood pressure, smoking, alcohol consumption, and marital status did not differ between the groups. Compared with the findings in the non-hypoglycemic group, the hypoglycemic group had a longer duration of intensive insulin treatment and a lower rate of combined oral hypoglycemic agent use (Table 2).

Table 2 Differences in sociodemographic and treatment-related characteristics of patients with type 2 diabetes mellitus between the non-hypoglycemic and hypoglycemic groups, n (%), mean ± SD.

	Non-hypoglycemic group (n = 442)	Hypoglycemic group (n = 360)	t/χ2	P value
Age (years)	54.04 ± 13.84	56.06 ± 12.94	-2.11	0.04a
Age (years)			6.36	0.10
18-44	116 (63.0)	68 (37.0)
45-59	173 (52.6)	156 (47.4)
60-74	127 (52.3)	116 (47.7)
≥ 75	26 (56.5)	20 (43.5)
Gender			3.59	0.06
Male	147 (50.7)	143 (49.3)
Female	295 (57.6)	217 (42.4)
BMI (kg/m2)	24.79 ± 3.38	23.36 ± 3.34	5.98	< 0.001a
BMI (kg/m2)			26.56	< 0.001a
< 18.5	12 (37.5)	20 (62.5)
18.5-23.9	174 (47.0)	196 (53.0)
≥ 24.0	256 (64.0)	144 (36.0)
Systolic blood pressure (mmHg)	130.24 ± 20.27	130.88 ± 20.67	-0.44	0.66
Diastolic blood pressure (mmHg)	82.23 ± 11.53	82.40 ± 11.37	1.02	0.31
Smoking			0.23	0.63
No	345 (54.7)	286 (45.3)
Yes	97 (56.7)	74 (43.3)
Alcohol consumption			1.87	0.17
No	384 (54.2)	324 (45.8)
Yes	58 (61.7)	36 (38.3)
Education level			4.77	0.10
Primary school or lower	99 (52.1)	91 (47.9)
Middle school	191 (52.8)	171 (47.2)
College or higher	152 (60.8)	98 (39.2)
Occupational status			12.29	0.002a
Employed	278 (60.3)	183 (39.7)
Unemployed	49 (45.4)	59 (54.6)
Retired	115 (49.4)	118 (50.6)
Marital status			1.82	0.40
Married	412 (55.2)	335 (44.8)
Unmarried	22 (61.1)	14 (38.9)
Divorced/widowed	8 (42.1)	11 (57.9)
Treatment before admission			0.55	0.97
None	133 (55.4)	107 (44.6)
Oral antidiabetic drugs1	183 (54.5)	153 (45.5)
Insulin	45 (52.9)	40 (47.1)
Oral antidiabetic drugs2 + insulin	76 (57.6)	56 (42.4)
Others	5 (55.6)	4 (44.4)
Treatment time (days)	9.20 ± 3.13	10.34 ± 3.56	−4.81	< 0.001a
Insulin injection therapy			0.42	0.52
Basal-before meal	4 (44.4)	5 (55.6)
CSII	438 (55.2)	355 (44.8)
Total insulin dose (IU/d)	34.63 ± 9.14	34.07 ± 9.79	0.84	0.40
Combined oral hypoglycemic drugs3			6.18	0.01a
No	341 (53.0)	303 (47.0)
Yes	101 (63.9)	57 (36.1)
Combined antihypertensive drugs			1.25	0.26
No	307 (53.9)	263 (46.1)
Yes	135 (58.2)	97 (41.8)

^aP < 0.05.

¹Including metformin, SGLT2 inhibitor, alpha-glucosidase inhibitor, DPP4 inhibitor, sulfonylureas, triple hypoglycemic therapy.

²Including metformin, SGLT2 inhibitor, alpha-glucosidase inhibitor, DPP4 inhibitor, one or two antidiabetic drugs combined with insulin.

³Including metformin, SGLT2 inhibitor, alpha-glucosidase inhibitor, intensive insulin therapy combined with one or two oral drugs.

BMI: Body mass index.

Univariate regression analysis revealed that the duration of diabetes, hyperlipidemia, 1-year history of hypoglycemia, HOMA-IR, and FPG, fasting insulin, fasting C-peptide, ALT, GFR, TG, and HDL-C levels significantly differed between the two groups (all P < 0.05; Table 3). Compared with the results in the non-hypoglycemic group, the hypoglycemic group featured a longer course of diabetes; lower FPG, fasting insulin, fasting C-peptide, and ALT levels; lower HOMA-IR; a lower incidence of hyperlipidemia; a higher rate of low GFR; a higher probability of TG and HDL-C within the target range of T2DM control; and a larger proportion of subjects with a history of hypoglycemia within 1 year before hospital admission (Table 3).

Table 3 Differences in disease-related and biochemical characteristics of patients with type 2 diabetes mellitus between the non-hypoglycemic and hypoglycemic groups.

	Non-hypoglycemic group (n = 442)	Hypoglycemic group (n = 360)	Z/χ2	P value
	n (%)/M (P25, P75)	n (%) /M (P25, P75)
Duration of diabetes	5.00 (0.96, 10.00)	8.00 (1.00, 12.00)	-2.49	0.01
Duration of diabetes			12.26	< 0.001a
< 10 years	308 (59.7)	208 (40.3)
≥ 10 years	134 (46.9)	152 (53.1)
Diabetic nephropathy			0.10	0.76
No	326 (54.8)	269 (45.2)
Yes	116 (56.0)	91 (44.0)
Diabetic neuropathy			0.02	0.89
No	284 (54.9)	233 (45.1)
Yes	158 (55.4)	127 (44.6)
Diabetic vascular disease			0.05	0.83
No	182 (54.7)	151 (45.3)
Yes	260 (55.4)	209 (44.6)
High blood pressure			0.01	0.94
No	270 (55.2)	219 (44.8)
Yes	172 (55.0)	141 (45.0)
Cardiovascular diseases			1.04	0.31
No	381 (55.9)	301 (44.1)
Yes	61 (50.8)	59 (49.2)
Hyperlipidemia			7.59	0.01a
No	276 (51.7)	258 (48.3)
Yes	166 (61.9)	102 (38.1)
Fatty liver			3.46	0.06
No	235 (52.2)	215 (47.8)
Yes	207 (58.8)	145 (41.2)
History of hypoglycemia within a year			6.73	0.01
No	435 (55.9)	343 (44.1)
Yes	7 (29.2)	17 (70.8)
Blood glucose status
Random blood glucose (mmol/L)	13.1 (10.3, 17.3)	13.9 (10.53, 18.18)	-1.42	0.16
FPG (mmol/L)	7.6 (6.4, 9.2)	6.7 (5.2, 9.1)	-4.31	< 0.001a
SDBG (mmol/L)	3.75 (2.72, 4.97)	3.92 (2.86, 5.47)	-1.64	0.10
HbA1c (%)	9.5 (8.1, 11.2)	9.85 (8.1, 11.6)	-0.96	0.34
Pancreas islet function
HOMA-β (%)	40.27 (26.45, 75.93)	39.08 (22.91, 95.24)	-0.48	0.63
HOMA-IR	3.21 (1.91, 5.26)	2.03 (1.29, 3.86)	-6.50	< 0.001a
Fasting insulin (mu/L)	8.96 (5.77, 14.22)	6.82 (4.31, 11.17)	-5.01	< 0.001a
Fasting C-peptide (nmol/L)	0.36 (0.22, 0.55)	0.24 (0.15, 0.42)	-6.75	< 0.001a
Liver and renal functions
AST	19 (15, 26)	19 (15, 24)	-1.49	0.14
AST			1.21	0.27
≤ 40 U/L	395 (54.5)	330 (45.5)
> 40 U/L	47 (61.0)	30 (39.0)
ALT	23 (15, 34)	19 (14, 29)	-3.39	0.001a
ALT			2.87	0.09
≤ 35 U/L	342 (53.6)	296 (46.4)
> 35 U/L	100 (61.0)	64 (39.0)
Cr	64 (52, 77)	66 (54, 80)	-1.37	0.17
Cr			0.07	0.80
≤ 133 μmol/L	431 (55.2)	350 (44.8)
> 133 μmol/L	11 (52.4)	10 (47.6)
GFR	102.77 (89.30, 113.93)	98.675 (81.31, 109.32)	-3.31	0.001a
GFR			7.56	0.006a
< 90	114 (47.7)	125 (52.3)
≥ 90	328 (58.3)	235 (41.7)
Blood lipid level
TC	4.87 (3.89, 5.61)	4.73 (3.78, 5.41)	-1.36	0.18
TC			0.81	0.37
< 4.5 mmol/L	174 (53.2)	153 (46.8)
≥ 4.5 mmol/L	268 (56.4)	207 (43.6)
TG	1.52 (1.05, 2.42)	1.25 (0.89, 1.81)	-5.55	< 0.001a
TG			19.31	< 0.001a
< 1.7 mmol/L	249 (49.2)	257 (50.8)
≥ 1.7 mmol/L	193 (65.2)	103 (34.8)
HDL-C	0.92 (0.79, 1.08)	0.99 (0.82, 1.18)	-3.04	0.002a
HDL-C			6.86	0.01a
≤ 1 mmol/L	281 (58.9)	196 (41.1)
> 1 mmol/L	161 (49.5)	164 (50.5)
LDL-C	3.09 (2.16, 3.74)	3.06 (2.20, 3.80)	-0.40	0.69
LDL-C			0.30	0.58
< 2.6 mmol/L	163 (56.4)	126 (43.6)
≥ 2.6 mmol/L	279 (54.4)	234 (45.6)
CRP	1.8 (1.00, 3.73)	1.8 (0.80, 3.50)	-0.94	0.35
CRP			0.17	0.68
≤ 6 mg/L	327 (55.9)	258 (44.1)
> 6mg/L	115 (53.0)	102 (47.0)

^aP < 0.05.

FPG: Fasting blood glucose; SDBG: Standard deviation of blood glucose; HbA1c: Glycated hemoglobin A1c; HOMA-β: Homeostasis model assessment for β-cell function; HOMA-IR: Homeostasis assessment of insulin resistance; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; Cr: Creatinine; GFR: Glomerular filtration rate; TC: Total cholesterol; TG: Triglyceride; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; CRP: C-reactive protein.

Logistic regression analysis of predictors for hypoglycemia

As HOMA-IR is not routinely measured in clinical practice and is often unavailable in medical records, we selected seven variables to construct the hypoglycemia prediction model based on the results of single-factor logistic regression analysis (P < 0.2) after excluding factors with multicollinearity and including factors with clinical practical significance. The occurrence of hypoglycemia was the dependent variable in the model. In contrast, six variables, including BMI, duration of diabetes, history of hypoglycemia within 1 year, days of treatment, GFR, and TG, were the independent variables (Table 4).

Table 4 Logistic regression analysis of predictors for hypoglycemia.

Variables	β	S.E	Z	OR (95%CI)	P value
BMI	-0.10	0.02	-4.30	0.90 (0.86-0.95)	< 0.001
Duration of diabetes
< 10 years	Reference
≥ 10 years	0.26	0.16	1.63	1.30 (0.95-1.78)	0.10
History of hypoglycemia within 1 year
No	Reference
Yes	0.89	0.48	1.86	2.42 (0.99-6.56)	0.06
DII	0.09	0.02	3.68	1.09 (1.04-1.14)	< 0.001
GFR
< 90	Reference
≥ 90	-0.34	0.17	-2.02	0.71 (0.51-0.99)	0.04
TG
< 1.7 mmol/L	Reference
≥ 1.7 mmol/L	-0.48	0.16	-3.01	0.62 (0.45-0.85)	0.003

BMI: Body mass index; DII: Duration of intensive insulin therapy; GFR: Glomerular filtration rate; TG: Triglyceride.

Construction of the hypoglycemia risk prediction model and the nomogram

In patients with T2DM undergoing intensive insulin treatment, the risk of hypoglycemia significantly differed for every 1 kg/m² increase in BMI [odds ratio (OR) = 0.90, 95% confidence interval (CI): 0.86-0.95; P < 0.001]. Patients with a diabetes duration exceeding 10 years had a higher risk of hypoglycemia than those with a diabetes duration shorter than 10 years (OR = 1.30, 95%CI: 0.95-1.78; P = 0.10). The risk of hypoglycemia during intensive insulin therapy was significantly higher in patients with a history of hypoglycemia within 1 year of admission than in those without hypoglycemia (OR = 2.42, 95%CI: 0.99-6.56; P = 0.06). The risk of hypoglycemia increased for every additional day of intensive insulin therapy (OR = 1.09, 95%CI: 1.04-1.14; P < 0.001). The risk of hypoglycemia during insulin-intensive treatment significantly differed between patients with GFR ≥ 90 and those with GFR < 90 (OR = 0.71, 95%CI: 0.51-0.99; P = 0.04). The risk of hypoglycemia during insulin-intensive treatment significantly differed between patients with TG levels ≥ 1.7 mmol/L and those with TG levels < 1.70 mmol/L (OR = 0.62, 95%CI: 0.45-0.85; P = 0.003).

The multivariate logistic regression equation was as follows: Logit (P)= 1.89 - 0.10 × BMI + 0.09×DII - 0.42 × GFR -0.53 × TG. The nomogram model is presented in Figure 1.

Figure 1 A nomogram model for predicting the risk of hypoglycemia in hospitalized type 2 diabetes mellitus patients treated with insulin. For estimating the risk of hypoglycemia, the corresponding value is found for each variable on its corresponding axis. Then, a vertical line is drawn from the variable axis to the scoring scale to obtain the score for each variable. The total score is calculated from the sum of scores for all the variables. Based on the total points, a vertical line is drawn downwards to the risk axis to obtain the hypoglycemia risk for each individual patient. BMI: Body mass index; DII: Duration of intensive insulin therapy; GFR: Glomerular filtration rate; TG: Triglyceride.

Validation of the hypoglycemia risk prediction tool

The Hosmer-Lemeshow test was used to calibrate the logistic regression model. The hypoglycemia risk prediction model displayed a reasonable degree of fit (χ² = 12.54, P = 0.14). Based on ROC curve analysis, the model exhibited good discriminative ability with an AUC of 0.67 (95%CI: 0.63-0.71; P < 0.001; Figure 2). The risk prediction model based on six independent variables was evaluated by 10-fold cross-validation. The AUC for the cross-verified model was 0.66, thereby confirming its good discriminative ability (Figure 2).

Figure 2 Receiver operating characteristic curve analysis of the hypoglycemia risk prediction model on the training set and after 10-fold cross-validation. A: The receiver operating characteristic (ROC) curve generated using the training set; B: The ROC curve derived from predictions obtained through 10-fold cross-validation.

DISCUSSION

The present study found that the incidence of hypoglycemia in hospitalized patients with T2DM treated with insulin was very high. The present study also developed and validated a hypoglycemic risk prediction model, which has six predictors including BMI, duration of diabetes, history of hypoglycemia within 1 year, glomerular filtration rate, blood triglyceride levels, and duration of treatment. The model displayed high reliability and discrimination ability and is a promising tool for clinicians to screen hospitalized patients with T2DM with an elevated risk of hypoglycemia and guide personalized interventions to prevent and treat hypoglycemia. We have compared our results with other studies. In our study, the incidence of hypoglycemia in hospitalized patients with T2DM treated with insulin was 44.9%, which significantly exceeded the rates of 10.3%-25% reported by Shah et al[16]. However, our data were comparable with the hypoglycemia incidence rate of 46.5% reported in the global HAT study[11]. The differences in incidence of hypoglycemia may be related to the different definitions of hypoglycemia used in each study and the level of glycemic control[15].

In this study, the median (interquartile range) duration of treatment for patients with hypoglycemia was 6.5 (4-9) days. This likely corresponded to a gradual recovery of islet function after approximately 1 week of intensive insulin treatment[17]. The frequency of hypoglycemia was highest before lunch (26.1%), followed by after breakfast (16.3%) and before breakfast (14.8%). Pazos-Couselo et al[18] also reported that hypoglycemia occurred more commonly among patients with diabetes before lunch (32.0%). This might be related to the high carbohydrate content of the Chinese breakfast. The present study also found that hypoglycemia was recorded once in 46.1% of the patients and two or more times in 53.9% of the patients in the hypoglycemic group. This demonstrated that the probability of recurrent hypoglycemia was high, which could further aggravate patients’ medical conditions and reduce the likelihood of recovery. The findings of the present study suggest that it is essential to assess the dynamic risk of hypoglycemia.

The present study found that the risk of hypoglycemia increased as the treatment duration of insulin therapy increased. Moreover, the treatment duration was an independent risk factor for hypoglycemia in hospitalized patients with insulin-treated T2DM. Jones et al[19] reported that hypoglycemia increased with the length of hospitalization in patients with T2DM. We postulated that patients with T2DM achieved partial remission of islet function in the middle and later stages of intensive insulin therapy. Still, the risk of hypoglycemia increased if the treatment plan was not adjusted over time. This suggested that the hypoglycemia evaluation results at admission should not be used to distinguish patients into groups with high or low hypoglycemia risks during the overall hospitalization period. The present study also found that BMI, biomarkers including GFR and TG, were predictors of hypoglycemia in hospitalized patients with T2DM treated with insulin. The risk of hypoglycemia significantly increased when the patients had a smaller BMI, with GFR < 90, and TG < 1.7 mmol/L. The findings of the present study suggest that healthcare providers should pay more attention to the blood glucose in these T2DM patients with an increased risk for hypoglycemia and modify the dosage of insulin timely.

In the present study, we used logistic regression analyses to build a hypoglycemia risk prediction model for hospitalized patients with T2DM treated with insulin, which has six independent hypoglycemia risk factors, including BMI, disease duration, history of hypoglycemia within 1 year, duration of treatment, GFR, and TG. The prediction accuracy of the hypoglycemia risk prediction model was good, as indicated by the Hosmer-Lemeshow goodness-of-fit test results. Moreover, the AUC for the model and the AUC for the 10-fold cross-validation model suggested that the prediction performance of the model was good and stable. Using 0.43 as the optimal cutoff for ROC curve analyses, the sensitivity, specificity, and Youden index were 70%, 58%, and 0.28, respectively.

We have compared our model with others. Shah et al[16] also developed a hypoglycemia prediction tool for hospitalized patients with diabetes by including age, emergency department visit because of hypoglycemia in the previous 6 months, use of insulin, use of oral drugs that do not cause hypoglycemia, and severe chronic kidney disease. They reported an AUC of 0.64, sensitivity of 65%, and specificity of 69%. Ena et al[20] validated a predictive model that included estimated GFR, daily insulin dose, length of hospitalization, and hypoglycemia episodes in the previous 3 months and reported an AUC of 0.72, sensitivity of 40.2%, and specificity of 87.2%. Stuart et al[21] developed a hypoglycemia prediction tool for hospitalized patients with diabetes by including factors such as age > 75 years, insulin and sulfonylurea treatment, Black and Asian ethnicity, emergency admission, low estimated GFR, high C-reactive protein levels, hyponatremia, and hypoalbuminemia and reported an AUC of 0.73, sensitivity of 59.3%, and specificity of 73.7%. The differences in the models may be related to the different subjects and variables included, as well as the different statistical models used.

There are some limitations in this study. First, the study cohort came from a single hospital, and all data were extracted retrospectively from the electronic medical record system. Second, because the present model has not yet undergone external validation, its clinical utility remains uncertain. Future studies could use a multi-center prospective design to modify this predictive model, confirm its generalizability, and inform safe clinical implementation.

Although the present study has some limitations, we used a logistic model to transform the regression equation into a simple visual graph to simplify the results of the prediction model[22], making it convenient for clinicians to obtain quantitative scores and predict the risk of hypoglycemia in patients. The nomogram model can be used for individualized risk prediction. The hypoglycemia risk can be estimated in each patient with T2DM using the nomogram model score to identify those at higher risk of hypoglycemia.

CONCLUSION

We developed a hypoglycemia risk prediction model for hospitalized patients with T2DM treated with insulin using independent risk factors, namely BMI, duration of diabetes, history of hypoglycemia within 1 year, GFR, TG, and duration of treatment. The model exhibited good reliability and clinical performance. Our data illustrated that this model is a promising tool for clinicians to screen hospitalized patients with T2DM with an elevated risk of hypoglycemia and guide prevention and treatment.