Early prediction cardiac arrest in intensive care units: the value of laboratory indicator trends

The incidence of in-hospital cardiac arrest (IHCA) has increased over the past decade, with more than half occurring in intensive care units (ICUs).[1] ICU cardiac arrest (ICU-CA) presents unique challenges, with worse outcomes than those in monitored wards, highlighting the need for early detection and intervention.[2] Up to 80% of patients exhibit signs of deterioration hours before IHCA.[3] Although early warning scores based on vital signs are useful, their effectiveness in ICUs is limited due to abnormal physiological parameters.[4] Laboratory markers, such as sodium, potassium, and lactate, are predictive of poor outcomes,[5] but static measurements may not capture the patient’s trajectory. Trends in laboratory indicators, such as variability and extremes, may offer better predictive value.[6] This study aimed to evaluate ICU-CA predictive factors, with a focus on vital signs and trends of laboratory indicators.

We conducted a retrospective matched case-control study from December 1st, 2016, to November 31st, 2019, in six ICUs of Qilu Hospital of Shandong University. Patients aged >14 years who experienced ICU-CA, defined as circulatory arrest receiving cardiopulmonary resuscitation (CPR), including chest compressions and/or cardiac defibrillation, were included. The exclusion criteria were do-not-attempt resuscitation orders, cardiac arrest (CA) before ICU admission, ICU stays <48 h, or intraoperative CA. Controls were patients without CA events during their ICU stays and were selected at a 1:2 ratio after being matched to CA cases by ICU, sex, admitting diagnosis, and admission date. All controls must have an ICU length of stay longer than the time to CA event of its corresponding case, allowing valid comparisons between these two groups within the time interval.

All clinical variables were collected retrospectively from the medical records. Vital signs (temperature, heart rate, respiratory rate, blood pressure, and oxygen saturation) and laboratory indicators (sodium, potassium, calcium, hemoglobin, lactate, and pH) were collected at 48, 36, 24, 12, and 8 h before ICU-CA or index time for controls. Missing data were filled using the last observation carried forward (LOCF) method.[7] Variability in laboratory indicators was assessed via mean, maximum, minimum, and range values.

Descriptive statistics are presented as median (interquartile range) or frequencies (%). Group comparisons were made via t-tests, Wilcoxon rank-sum tests, and χ² or Fisher’s exact tests. Univariable and multivariable logistic regression models were applied to assess predictors of ICU-CA, with odds ratios (ORs) and 95% confidence intervals (95% CIs), and the area under the receiver operating characteristic (AUROC) curve was used to indicate the predictive power of the laboratory indicator trends and variabilities for ICU-CA. P-values <0.05 were considered statistically significant. Analyses were conducted via IBM SPSS statistics 25.0. Detailed statistical methods are available in the supplementary material.

During the study period, 6,164 patients were treated in the six ICUs, with 1,042 identified as ICU-CA patients. After exclusions, 391 ICU-CA patients were included, and 782 patients were selected as controls (supplementary Figure 1). The baseline characteristics revealed a median age of 63 years, and the patients were predominantly male (70.9%), with frequent comorbidities such as hypertension, coronary artery disease (CAD), and diabetes mellitus (DM). Compared with the controls, the ICU-CA patients were slightly older and had higher rates of CAD (Table 1).

Table 1.

Baseline characteristics of the ICU-CA patients and controls

Logistic regression identified several vital signs and laboratory markers significantly associated with ICU-CA. Heart rate (HR) and systolic blood pressure (SBP) at specific time points, as well as high sodium (Na⁺), potassium (K⁺), hemoglobin (HGB), lactate (Lac), and abnormal pH levels, were significant predictors of ICU-CA (supplementary Table 1). In the final model, Na⁺ >145 mmol/L, elevated K⁺, lower HGB, and elevated Lac were independently associated with ICU-CA (Table 2).

Table 2.

Multivariable multilevel logistic regression analysis of the association of vital signs and laboratory indicators with ICU-CA

Trends in laboratory markers, particularly Lac and pH, improved the predictive accuracy closer to the CA event, with Lac showing the highest predictive value at the -8 h time point (AUROC 0.773). The variability in markers including Na⁺, K⁺, HGB, Ca²⁺, pH, and Lac further enhanced the prediction, with the maximum value of Lac being the most accurate indicator (Figure 1). These findings suggest that both the trend and variability of laboratory indicators, especially Lac, offer valuable insights for the early prediction of ICU-CA, potentially improving clinical outcomes through timely interventions.

Figure 1.

AUROCs for the trend and variability of laboratory indicators. AUROC: area under the receiver operating characteristic curve; Na⁺: sodium; K⁺: potassium; Ca²⁺: calcium; HGB: hemoglobin; Lac: lactic acid; Max: maximum; Min: minimum.

In this study, we identified laboratory indicators, including Na⁺, K⁺, HGB and Lac, as independent predictors of ICU-CA. Our analysis of trends and variability in these indicators revealed that their predictive accuracy improved closer to the event, with Lac showing the greatest improvement. Specifically, the minimum value of HGB, the range value of Na⁺, K⁺, Ca²⁺, pH, and the maximum value of Lac were the most accurate predictors. This highlights the importance of laboratory trends and variability in predicting ICU-CA, and this approach may offer a feasible method for the early detection of ICU-CA.

Previous studies[7,8] have explored the early prediction of IHCA, combining vital signs and laboratory findings. Recent advancements in the prediction of CA have led to a shift toward more efficient and precise methodologies, such as machine learning and artificial intelligence, which have become powerful tools and offer enhanced capabilities for predictive analytics of ICU-CA.[9-12] Ueno et al[13] proposed that a model that considers both vital signs and laboratory results outperforms a model that is based on vital signs alone in the prediction of IHCA in the ICU. In our study, vital signs, including HR and SBP, were associated with ICU-CA at certain time points, whereas laboratory values, such as Na⁺, K⁺, HGB and Lac, consistently predicted ICU-CA, suggesting that laboratory indicators might be more valuable in the ICU setting. Vital signs in ICUs are often normalized quickly through interventions, potentially masking underlying physiological disturbances, whereas laboratory tests provide a clearer picture of internal homeostasis. Furthermore, a study supports the use of variability in vital signs for IHCA prediction.[5] In our analysis, laboratory indicators, not vital signs, were more closely linked to ICU-CA, with trends and variability enhancing predictive accuracy. Since laboratory data are frequently available in ICUs, particularly through arterial blood gas analysis, this approach offers a feasible method for the early detection of ICU-CA.[14] Future models using interpretable machine learning algorithms for ICU-CA prediction could incorporate a broader range of vital signs and laboratory indicators, such as trends and variability, thereby allowing for more precise and patient-specific predictions.

This study has several limitations. First, this study was conducted in a single center, limiting the generalizability of the results. Second, the retrospective data may have had missing points. Although the LOCF method was used to handle missing data, some information bias remains. Third, despite the suitability of the study design, some factors may not have reached statistical significance because of the difference between the actual incidence of ICU-CA and the ratio observed in our study population.

In conclusion, laboratory indicators, including Na⁺, K⁺, HGB, and Lac, were independently associated with ICU-CA. The minimum HGB, range value of Na⁺, K⁺, Ca²⁺, pH, and maximum Lac were the most accurate predictors. Adding trends or variabilities of laboratory indicators may increase the accuracy of models designed to detect critical illness in the ICU in the future.