This was a secondary analysis of a prospective observational cohort study (ChiCTR2000030846) (18). Patients were consecutively monitored in the five Pediatric Intensive Care Units (PICUs) in southwestern China from January to December 2020. To be eligible for inclusion, participants must: 1) receive vasoactive drug support for hypotension or ventilatory support for respiratory failure; 2) be between a corrected gestational age of 36 weeks and 16 years; 3) remain in the PICU for more than 24 hours. Exclusion criteria are: 1) a prior diagnosis of diabetes mellitus or inborn errors of metabolism; 2) readmission to the PICU; 3) use of glucocorticoids or total parenteral nutrition; 4) fewer than two arterial blood gas analyses daily during the first 72 hours of ICU admission. This secondary analysis was approved by the ethics committee of Children’s Hospital of Chongqing Medical University and was granted a waiver for informed consent (File No. 2022,125). Elaboration on the methods for original data collection, inclusion and exclusion criteria, insulin usage protocols, and quality control procedures can be referenced in our previous study (18).

We collected baseline demographic data (age, gender, and nutritional status), primary reasons for ICU admission (urgent organ dysfunction or critical disease conditions necessitating ICU intervention), clinical treatments (such as insulin, glucose infusion rate (GIR), and other supportive therapies), and illness severity assessment (The Pediatric Logistic Organ Dysfunction score-2, PELOD-2). The nutritional status index was assessed using the Z score for weight/height (for individuals aged ≤5 years) or Body Mass Index (for those older than five years) for age (BMZ), calculated through the Epi-info software. The PELOD-2 score was determined based on the most extreme physiological parameters and laboratory findings observed within the first 24 hours of admission, reflecting the severity of the patient’s condition, with a higher score indicating a worse condition and poorer prognosis (19).

Blood Glucose and lactate measurements were exclusively obtained from arterial blood gas analyses during the first 72 hours after PICU admission. We obtained the concentration of mean glucose (Glu_mean), maximum glucose (Glu_max) and calculated time-weighted mean glucose (Glu_twmean) (14). For lactate, corresponding measurements were collected, including mean lactate (Lac_mean), maximum lactate concentration (Lac_max), and time-weighted mean lactate (Lac_twmean). All blood gas samples were obtained through routine monitoring or measurements based on the clinical judgment of the attending physicians, with no additional blood gas analyses performed specifically for the purpose of the study. The blood gas analyzers in different research centers shared similar reference ranges and underwent regular testing as well as stringent laboratory quality control measures, including daily internal quality control and biannual external quality assessments.

A criterion of at least one episode of blood glucose concentration above 150 mg/dL (8.3 mmol/L) was established for defining SHG, as this threshold marked the onset of increased mortality risk and was recommended for initiating insulin therapy (5, 20). Meanwhile, cut-off blood glucose concentrations of 65 mg/dL (3.6 mmol/L) and 40 mg/dL (2.2 mmol/L) were adopted to define mild hypoglycemia (21) and severe hypoglycemia (7), respectively. HL was defined as a blood lactate concentration greater than 2 mmol/L, as this cut-off value was the threshold for septic shock definition and was highly correlated with an unfavorable prognosis (22).

The primary outcome of interest was 28-day mortality. The secondary outcomes comprised 28-day ventilator-free days and 28-day ICU-free days. A value of 0 was recorded for patients who did not survive within the 28 days, while the recorded value for survivors was 28 days minus the corresponding number of days on the ventilator or in the ICU, with a smaller value indicating a poorer prognosis (23).

All statistical analyses were performed using SPSS (Version 23, IBM, Armonk, NY, USA). The Kolmogorov-Smirnov test was used to assess the normality of continuous variables. The results showed that these variables had non-normal distributions, therefore, presenting as medians with interquartile ranges (IQR) and compared by the Wilcoxon rank-sum test. Categorical variables were expressed as numbers and percentages, and comparisons were made using either the chi-square test or the Yates corrected chi-square test, as appropriate. The Cochran-Mantel-Haenszel test was used to examine the association between SHG and 28-day mortality across different HL states, and the Breslow-Day test was used to assess the homogeneity of odds ratios (OR) across different strata. Interaction effects were evaluated by testing the product terms of lactate and glucose metrics in multivariable regression models. In cases of significant interaction effects, further subgroup regression analyses stratified by HL would be conducted. Depending on the type of outcome variable, either a multivariable logistic regression or linear regression model was employed. ORs with 95% confidence intervals (CI) derived from logistic regression were reported. For linear regression, coefficients, standard errors, and t-values were reported. Predefined covariates in the regression models were chosen based on demographics (gender, age groups cut-off at 36 months), nutritional status (BMZ), factors potentially influencing glycometabolism (GIR and insulin), and severity of illness (PELOD-2 score). To ensure the robustness of the results while adhering to predictor variable constraints and mitigating overfitting, additional covariates were sequentially introduced and replaced one at a time in subsequent steps. Sensitivity analyses were also performed by exploring the interaction effects of SHG and HL on outcomes in patients without hypoglycemia. A p-value < 0.05 for a two-sided test was considered statistically significant, except for interaction analyses where a p-value < 0.10 was used (24).

During the original study period, a total of 2,725 consecutive patients across the five participating units underwent stringent screening to determine their eligibility. Ultimately, a cohort of 433 pediatric patients was included ( Figure 1 ). Among them, the median age was 16.0 (IQR: 4.3 ~ 71.5) months, 255 (58.9%) cases were males, and 90 (20.8%) died within 28 days after PICU admissions. The leading primary reason for ICU admission was respiratory diseases, accounting for 152 cases (35.1%). At the primary endpoint of the 28-day mortality assessment, participants were divided into survivors and non-survivors groups. The deceased cases were associated with elevated GIR, increased requirements of insulin, continuous renal replacement therapy, mechanical ventilation, vasoactive drugs, and higher PELOD-2 scores (p < 0.05).

Within the first 72 hours of PICU admission, 4885 simultaneous glucose and lactate measurements from arterial blood gas analyses were obtained, with a median frequency of 12 (IQR: 9~13) per patient. Compared with survivors, more cases of SHG (64.4% vs. 32.7%), mild hypoglycemia (34.3% vs. 21.6%), severe hypoglycemia (7.8% vs. 1.7%), and HL (71.1% vs. 36.2%) were detected in non-survivors (p<0.5) ( Table 1 ).

The three-dimensional column diagram illustrated 28-day mortality across combined quartiles of lactate and glucose metrics( Figure 2 ). Figure 2A revealed that the combined highest Lac_mean and Glu_mean quartiles (upper-right column) were associated with the highest 28-day mortality, whereas relatively low Lac_mean (Q1~Q2) even with the highest Glu_mean didn’t correlate with high 28-day mortality (upper-left columns); This pattern was consistently observed across quartiles of Lac_twmean and Glu_twmean ( Figure 2B ), as well as Lac_max and Glu_max ( Figure 2C ), concerning 28-day mortality.

The Cochran’s and Mantel-Haenszel test was used to explore the relationship between SHG, HL, and 28-day mortality ( Figure 3 ). In the non-HL group, SHG did not significantly affect PICU mortality (OR = 1.05, 95% CI 0.42~2.62, p = 0.922). Conversely, in the HL group, SHG was significantly associated with 28-day mortality (OR = 4.92, 95% CI 2.43~9.96, p < 0.001). The Test of Homogeneity of ORs revealed a significant difference in the ORs between the HL and non-HL groups (p of Breslow-Day = 0.007), indicating that the effect of SHG on mortality varied significantly with HL states.

In the multivariate logistical regression analyses, noteworthy interaction effects were observed between Glu_mean and Lac_mean, Glu_twmean and Lac_twmean, Glu_max and Lac_max, as well as SHG and HL concerning mortality, yielding p-values of interaction as 0.031, 0.020, 0.093 and 0.025, respectively. Stratified multivariate regression analyses revealed that in the subgroup without HL, quartiles of Glu_mean, Glu_twmean, and Glu_max, as well as SHG, were not independent predictors of mortality (p = 0.377, 0.439, 0.572, and 0.656, respectively). Conversely, within the HL subgroup, all these glucose indices were independently associated with mortality: Glu_mean (OR 1.43, 95% CI 1.02~1.99, p = 0.036), Glu_twmean (OR 1.42, 95% CI 1.01~2.00, p = 0.042), Glu_max (OR 1.91, 95% CI 1.28~2.86, p = 0.002), and SHG (OR 3.55, 95% CI 1.62~7.80, p = 0.002) ( Table 2 ). In addition, the interaction effects between other categorical covariates and HL/SHG were also checked, no significant interactions were found (p>0.1) ( Supplementary Table S1 ).

Similarly, significant interaction effects between SHG and HL on 28-day ventilator-free days and 28-day ICU-free days were detected using multivariate linear regression, with p-values of 0.006 and 0.014, respectively. In the non-HL subgroup, no significant association between SHG and either 28-day ventilator-free days (p = 0.916) or 28-day ICU-free days (p = 0.914) was found. In contrast, within the HL subgroup, SHG was associated with a reduction of 5.04 days in 28-day ventilator-free days (p = 0.003) and 4.10 days in 28-day ICU-free days (p = 0.004), suggesting that HL may modulate the effect of SHG on these outcomes ( Table 3 ).

Additionally, to strengthen the observation that HL modulated the association between SHG and 28-day mortality, additional covariates were included. To adhere to predictor variable constraints and avoid overfitting, Mild hypoglycemia, severe hypoglycemia, continuous renal replacement therapy, mechanical ventilation, and vasoactive drugs were introduced and replaced sequentially. The interaction effect of SHG and HL remained robustly significant, with p-values of 0.019, 0.015, 0.022, 0.027, and 0.032, respectively. In the non-HL subgroup, SHG did not correlate with 28-day mortality (p > 0.05). In contrast, in the HL subgroup, SHG was consistently associated with mortality (p < 0.05) ( Table 4 ).

Recognizing the nature of the observational data, sensitivity analyses were further conducted. Patients with hypoglycemia at 3 thresholds: glucose <2.2, 3.3, and 3.6 mmol/L were sequentially excluded and formed 3 subgroups (excluded n=13, 59, 105; remaining n=420, 374, 328) (7, 21, 25). SHG×HL interactions on 28-day mortality and secondary outcomes of these 3 subgroups were retested. All subgroups showed consistent statical significance of interaction effects between HL, SHG and poor outcomes (p>0.1) ( Supplementary Tables S2 – S7 ).