Acute respiratory failure requiring the initiation of invasive mechanical ventilation remains commonplace in the pediatric intensive care unit (PICU) (1). Emergent endotracheal intubation in children is associated with an increased risk of complications as compared to intubations performed electively (2). Early recognition of patients at risk for respiratory failure may provide clinicians with the opportunity to intervene and potentially improve outcomes (3).

Predictive analytics leverages physiologic and biochemical data in the development of algorithms that can identify signatures of illness present early in the process of clinical decompensation (4–6). Display of algorithmic-derived scores or indices alert care providers to patients at risk for clinical decompensation. Our group previously found a signature of respiratory distress leading to unplanned intubation in the adult intensive care unit (ICU) population that we validated externally at another site (3, 7, 8). Display of a risk estimate for this event in a surgical ICU was associated with a fall in septic shock by 50% that was not matched in a medical ICU that did not have the display (9). Through the development of a random forest model to identify patients at risk for requiring urgent unplanned intubation, we tested the hypothesis that subtle signatures of illness are present in physiological and biochemical time series of PICU patients in the early stages of respiratory decompensation (10).

Table 1 shows the study cohort. There were 2767 PICU admissions included in the study. When we require a non-elective indication for intubation using the UVA contribution to the NEAR4KIDS database we found 92 admissions with 116 unplanned intubations. The most common indications noted in the NEAR4KIDS database for unplanned intubations were respiratory failure (64%), both respiratory failure and hemodynamic instability (14%), and hemodynamic instability alone (12%). Patients who experienced unplanned intubation during their stay were younger and were more likely to have had cardiac surgery during their stay. When we did not require a non-elective indication from the UVA contribution to the NEAR4KIDS database we found 944 admissions using the computable phenotype alone.

Figure 1 shows the characteristics of the study cohort. In Figure 1A, patients admitted for cardiac surgery were predominantly infants admitted with complex congenital heart disease, while medical and non-cardiac surgery patients were more distributed across the pediatric age range. In Figures 1B–D, the estimates are made at each quartile of age for the specific patient type. Unplanned intubation was more likely for younger patients regardless of patient type, see Figure 1B. Unplanned intubation increased length of stay all patient types and ages, and increased mortality for most in Figures 1C,D.

Figure 2 shows an example of the physiological signature of unplanned intubation for cardiac surgery (left), medical (center), and non-cardiac surgery (right) patients. In each panel, each colored column estimates the empirical relative risk for unplanned intubation as a function of respiratory rate for one decile of age based on the surrounding quintile. That is, for all measurements in each decile of age and respiratory rate, we calculated the probability of unplanned intubation in the next 12 h and divided by the average rate of unplanned intubation in the cohort (0.0063). The youngest 10% of medical patients, for example, have about 3-fold more risk than average (redder) for unplanned intubation when respiratory rate is in the lower 30% (less than about 27 breaths per minute) independent of all other features. Averaging across all columns and panels yields an age and patient type marginalized risk profile for respiratory rate as shown in Figure 3. The deciles of respiratory rate and age are maintained across the rows and panels. There was a distinct physiological signature of illness for unplanned intubation depending on patient type. High respiratory rate is a signature of unplanned intubation for non-cardiac surgery patients and older medical patients, while in younger medical patients it is low respiratory rate and for cardiac surgery patients there is not a strong signature in respiratory rate. See Supplementary Figure A2 for all features.

Figure 3 shows the marginal risk profiles of several example features for the random forest model based on all patients. Each panel shows the log odds of unplanned intubation (ordinate) as a function of the value of each feature (abscissa). We estimated log odds by varying each feature across its range and estimated the log odds from the predicted probability while keeping all other features at their median values, then calculated the logit of predicted probability based on the model. Increasing heart rate from about 70 to 150 with all other features at the median values for the sample, for example, increases the log odds of unplanned intubation in the next 12 h from about −6 to −5, or an increase in probability from 0.0025 to 0.0067. Note that this marginalized representation removes the dependence captured by the model of each feature on age. See Supplementary Figure A3 for all features. All features were found to be more important than the random feature. Small changes in heart rate, respiratory rate, blood pressure, and others, alone or in combination, can be signatures of changing risk for unplanned intubation.

The AUC for the random forest model based on cross-validated predictions is 0.696 for all patients (0.692 for medical, 0.673 for cardiac surgery, and 0.727 for non-cardiac surgery). The AUC depends on the definition of case data (24): the AUC for the random forest model using a definition of 4 h before the event (rather than 12 h before) is 0.766 (0.738 for medical, 0.755 for cardiac surgical, and 0.797 for non-cardiac surgical). For comparison, the AUC of HRC monitoring for neonatal sepsis [that resulted in 20% mortality reduction (28)] within 24 h was 0.70 in the development cohort (29). In adults, prospective display of models for unplanned intubation and hemorrhage with AUC of 0.68 and 0.71, respectively (7), were associated with a 50% reduction of septic shock in an adult surgical ICU (9).

Figure 4 shows a sensitivity analysis for using a computable phenotype rather than chart review, and for excluding continuous cardiorespiratory monitoring features from the model. Each AUC is based on the cross-validated probabilities. Models using continuous cardiorespiratory monitoring features perform better than those without, and models based on chart reviewed events perform better than those based on a computable phenotype: the AUC using a computable phenotype is 29% lower than using manual chart review (0.638 and 0.696, respectively) and the AUC using chart review but excluding continuous cardiorespiratory monitoring features is 12% lower (0.671).

Figure 5 shows the average model output leading up to the time of unplanned intubation. For each event of unplanned intubation, we estimated the relative risks are based on the cross-validated predicted probabilities from the model. We then averaged the relative risks at each time to event for all events with data. To test that the predicted risk is increasing leading up to the time of unplanned intubation, we tested the null hypothesis that risk estimates are less than or equal to the values for the same patients 8 h prior. Open circles indicate that we reject the null hypothesis at the 0.05 significance level based on a Wilcoxon rank sum test. The figure indicates that all patient types have indications of physiological derangement 4–6 h prior to unplanned intubation.

Like adults, children have a physiologic signature of illness preceding unplanned intubation in the PICU. Generally, it comprises younger age, and abnormalities in electrolyte, hematologic and vital sign parameters. Given the heterogeneity of the patient population, it is not surprising that there are differences in the presentation among the major patient groups – medical, non-cardiac surgical, and cardiac surgical. We have argued elsewhere that predictive analytics monitoring is not a one-size-fits-all enterprise (30).

There are a variety of indications for unplanned intubation including oxygen failure, ventilation failure, and unstable hemodynamics. For the individual patient, it is often the case that multiple indications for intubation exist, demonstrating the complexity of physiology at play (31). For example, a patient with infantile botulism or spinal muscular atrophy may first exhibit signs of decompensation secondary to neuromuscular weakness which can progress to ventilation failure and ultimately oxygenation failure. Identifying physiologic and biochemical signatures of illness that help discriminate patients at risk for decompensation may allow for earlier intervention and potentially even prevention of the need for intubation.

We observed overlap in the signatures of illness between the medical and non-cardiac surgical patients. The leading indication for intubation in both the medical and non-cardiac surgical patients was respiratory failure (e.g., oxygen and/or ventilation failure) typically driven by primary disorders of respiratory physiology (e.g., pneumonia, acute lung injury). This is best illustrated by the heat map in Figure 2 exhibiting the influence of respiratory rate on the risk of intubation. Cardiac surgical patients, on the other hand, may be more likely to exhibit hemodynamic instability necessitating intubation, distinguishing them from the other two patient types.

This work is presented in the context that the multivariable statistical models that we use to look for signatures of illness can be repurposed for bedside risk prediction to supplement clinical decision making, including potential improvement in patient outcomes (32–36). The precedent is the heart rate characteristics index, a logistic regression model that used time series mathematical measures to identify early signatures of infection in preterm infants (37). Preterm infants in intensive care represent a mostly homogeneous cohort with a consistent signature of deterioration well-suited for logistic regression. The PICU, by contrast, may consist of multiple patient cohorts over a wide range of ages with different physiology and therefore physiological signatures, thus requiring more adaptable modeling approaches.

We found components of the signature in the labs (white blood cell count), vital signs (BP), and from the bedside monitor (every-2-second respiratory rate, variability of the SpO₂). We affirmed the importance of continuous cardiorespiratory monitoring data in discovering dynamic signatures of illness (38–40) excluding it reduced the AUC by 12%. This finding underscores the importance of using all available sources of information in bedside predictive analytics monitoring (41).

Individual chart review proved to be of great importance in defining the signature of illness. When we used a computable phenotype, we found cases in 852 more patient admissions, but the statistical models returned poorer performance with a 29% reduction in AUC. Like sepsis, the gold standard for case identification for model training is individual chart review (42).

Our study has several limitations including the single-center observational nature. Our model was constructed using data from a mixed cardiac/medical-surgical PICU, limiting its applicability broadly to all PICUs and patient populations. We were limited to the time of the intubation, rather than the time the clinical decision to intubate was made, to define the event. We were also limited by the indication(s) for intubation (e.g., respiratory failure, hemodynamic instability) as recorded in the NEAR4KIDS database. To our knowledge, the inter-rater reliability for establishing intubation indication(s) for the NEAR4KIDS database has not been evaluated. The small number of admissions from teenagers, and the low rate of unplanned intubation in that age range, limits the ability to capture a complete signature of illness. Physiological monitoring data included epochs with artifact due to, for example, physical therapy or suctioning. We also did not control for the impact of routine medications, such as sedatives and vasoactives, or non-invasive respiratory support on physiology. We acknowledge that these factors contribute to the complex and nuanced clinical scenarios in the PICU, and that results of a predictive model are to be taken into consideration with many other factors.

Additionally, we were limited to the time of the intubation, rather than the time the clinical decision to intubate was made, to define the event. Several factors may influence the timing of the clinical decision to intubate. While one patient may demonstrate clinical decompensation over a period of hours and another may experience an acute catastrophic event, we were encouraged by the observation of improved AUC at four hours prior to the event. Next steps include external validation of our models and evaluation through prospective integration into clinical care.

We have studied physiological and biochemical dynamics leading up to the event of unplanned intubation in a PICU. We observed that patients who experienced unplanned intubation were younger, more likely to have a co-morbid condition, and more likely to have undergone cardiac surgery. In addition, we identified different signatures of illness for medical, cardiac surgical, and non-cardiac surgical patients. While all this information is already incorporated into clinical practice by PICU teams, using predictive analytics prospectively to accurately estimate risk based on high-dimensional physiological signatures can support clinical decision making, direct new approaches for continuous monitoring in the PICU, and improve patient outcomes.