Near-infrared spectroscopy (NIRS) has gained popularity in the PICU, as it allows for continuous and noninvasive assessment of cerebral or somatic tissue oxygenation at the bedside (1–4). NIRS-detected cerebral hypoxia has been validated in relation to venous oximetry and anaerobic metabolism (3,5,6). Therefore, cerebral NIRS monitoring is used as a hemodynamic monitor for early recognition of an inadequate global oxygen supply/demand relationship, and to adapt interventions to minimize the risk of secondary organ dysfunction.
Evidence on the usefulness of NIRS monitoring both for prognosis and to guide management decisions in pediatrics is limited and has focused on cardiac surgery settings. A single study (7) suggests that NIRS-guided early detection of decreased tissue perfusion and oxygenation could allow initiation of prompt therapies to avoid organ damage and hence improve patient outcome. However, there is currently no consensus on which critical NIRS thresholds could guide patient care, trigger interventions, or be used for prognostication (8–10). Given the relatively high cost of the monitoring sensors, it is crucial to determine specific settings in which NIRS monitoring could improve patient care.
One such setting, to which impaired global perfusion contributes significantly, is acute kidney injury (AKI). AKI affects approximatively 40% of children after pediatric cardiac surgery (11,12), and is associated with prolonged duration of ICU and hospital stay and increased mortality (11–15). A reduced systemic oxygen delivery or its variability may be correlated with reduction in renal substrate delivery and AKI. Indeed, a prior study has shown that decreased renal NIRS oxygen saturation was predictive of AKI after adult cardiac surgery (16). Whether the more common cerebral site of NIRS monitoring is equally predictive of AKI is unclear.
In this study, we hypothesized that cerebral NIRS oximetry could adequately discriminate the risk of severe AKI in children following cardiac surgery and could be combined with routinely collected patient information to improve predictive performance.
MATERIALS AND METHODS
This prospective blinded observational study (ClinicalTrials.gov NCT01706497) was performed between October 2012 and November 2015, in the PICU of the Leuven University Hospitals, Leuven, Belgium. The study aimed at evaluating associations between NIRS and ICU short- and long-term outcomes (17) and hemodynamic instabilities, including AKI. Patient enrollment and clinical data collection protocol, including a waiver of parental consent for study participation, were approved by the Institutional Review Board.
Children after cardiac surgery, younger than 12 years old, with an arterial line in place, mechanically ventilated upon PICU admission or intubated after admission, and expected to stay at least 24 hours in the PICU, were eligible for the study. Patients were excluded if they had actual or potential brain damage, or if they had a condition or a wound that prohibited the placement of the forehead NIRS sensors. In addition, we excluded patients who developed AKI within 6 hours after admission and patients in whom NIRS monitoring was initiated after AKI onset.
Near-Infrared Spectroscopy Monitoring
The cerebral oxygen saturation was continuously measured with NIRS in all eligible patients from PICU admission until they were weaned off mechanical ventilation, using the FORESIGHT cerebral oximeter (CAS Medical Systems, Branford, CT). The monitor screens were blinded to the bedside clinicians with a sealed screen-cover in order not to influence the predictive value of the signal.
Prediction of Acute Kidney Injury
The primary outcome was prediction of severe AKI 6 hours before its occurrence during the first week of ICU stay. Severe AKI was defined according to stage 2 and 3 of the Kidney Disease: Improving Global Outcome criteria (18), as described in the supplemental data (Supplemental Digital Content 1, http://links.lww.com/CCX/A126).
Prediction of severe AKI was performed, first using the signal from the cerebral NIRS oximeter (NIRS signal), second using routinely collected clinical data, and third by combining NIRS and clinical information. Data were retrieved during the observation window, which lasted from admission until prediction (supplemental data, Supplemental Digital Content 1, http://links.lww.com/CCX/A126).
A NIRS model was developed using the minute-by-minute NIRS signal after preprocessing, as described in (17), and transformation to relevant predictors (for complete list, see eTable 1, Supplemental Digital Content 1, http://links.lww.com/CCX/A126), including value-based metrics, variability metrics, frequency components, time and dose below and above various hypoxic and hyperoxic thresholds (19,20). Variability metrics included root mean square of successive differences (RMSSD), sd, and smoothed sd (sd-s) (17). It is important to note that up-to-date cerebral oximeters only display value-based metrics and the dose below a user-defined threshold. For the most predictive NIRS predictors, interaction analyses were conducted with mean blood pressure and arterial carbon dioxide (co2) independently, as they may influence the NIRS signal.
A clinical prediction model was developed based on patient demographics and prospectively collected clinical information recorded during surgery and ICU stay (Table 1). Monitoring information included minute-by-minute heart rate and systolic blood pressure, daily lactate levels, hemoglobin levels, arterial oxygen saturation, and central venous oxygen saturation measured using intermittent blood sampling and treatment with extracorporeal membrane oxygenation. The median values and variability metrics of minute-by-minute monitoring signals were used as clinical predictors.
Combination of NIRS and Clinical Models.
Predictors included in the NIRS model and in the clinical model were combined to assess the added predictive value of the NIRS signal as compared with routinely collected data.
Data are presented as means and sd, medians and interquartile ranges, and numbers and proportions, where appropriate. Differences between characteristics of patients with and without AKI were compared with one-way analysis of variance for continuous variables and Fisher exact test for categorical variables. Statistical significance was set at p value of less than 0.05. All analyses were performed using Python Version 3.7 (Python Software Foundation, http://www.python.org), Scipy Version 1.3 (SciPy.org).
Reporting of the study was performed using the Strengthening the Reporting of Observational Studies in Epidemiology guidelines (21).
Prediction Model Development.
Prediction models were developed using logistic regression. For each model, predictors with a univariable p value of less than 0.05 were included in the models. To reduce overfitting and optimize model generalizability, backward feature selection was used to identify the smallest and most accurate model. Model performance and stability were internally validated using 1,000 bootstrap replicas (22).
Diagnostic Accuracy Assessment.
Model performance was reported using discrimination, calibration, and net benefit (23). Discrimination between patients with and without AKI was evaluated with the receiver operating characteristics (ROCs) curve and the area under the ROC curve (AUROC). Calibration refers to the agreement between the observed frequency of AKI in the population and the model predictions. Calibration was assessed using calibration belts together with the distribution of patient numbers (24). A statistically significant difference from perfect calibration is reported by a calibration test p value of less than 0.05 (24). Finally, the net benefit of the model was assessed by the difference between the expected benefit and the expected harm associated with model classification of AKI. Net benefit was visualized using decision curves and reported using ranges above “treat-all” and “treat-none” curves (25,26).
A total of 177 patients met inclusion criteria. Twenty-one patients were excluded (Fig. 1): four had AKI within 6 hours after admission, 16 were monitored with NIRS after AKI onset, and one had NIRS monitoring initiated later than 72 hours, which prevented the fit of prediction time for this patient. The remaining 156 patients were included in the analysis.
Demographics and outcomes of study participants are reported in Table 1. Fifty-five patients (35.3 %) developed severe AKI. Compared with patients without AKI, patients with AKI had a higher PICU and 90-day mortality risk (p = 0.02 and p = 0.05, respectively), longer PICU and hospital stay (p < 0.001) and longer duration of mechanical ventilation (p < 0.001).
Performance of NIRS Model
Compared to patients without AKI, patients with AKI had a lower maximum value of NIRS signal (eTable 1, Supplemental Digital Content 1, http://links.lww.com/CCX/A126) (p = 0.01), lower mean (p = 0.007), lower RMSSD (p < 0.001), and increased time and dose below 50% (p = 0.01, p = 0.03, respectively) and 60% (p = 0.001). These predictors were calculated 6 hours prior to AKI diagnosis, suggesting that the NIRS signal carries meaningful information for severe AKI prediction. However, given the low occurrence rate of saturations below 50% and 60%, the robustness of these predictors could not be assessed with the bootstrapping approach.
Performance of each significant NIRS predictor for severe AKI is reported in eFigure 1 (Supplemental Digital Content 1, http://links.lww.com/CCX/A126) and Table 2. Overall, the RMSSD (Fig. 2), a measure of NIRS variability, was more discriminant than the mean and the maximum values of NIRS (AUROC, 0.68; 95% CI, 0.67–0.68; p < 0.001), had wider (25–92%) and higher ranges of net benefit, and was well calibrated (p = 0.6). A patient with low NIRS RMSSD (e.g., 0.56, 25th percentile) (eTable 1, Supplemental Digital Content 1, http://links.lww.com/CCX/A126) had a 28% larger probability of developing AKI than a patient with high NIRS RMSSD (e.g., 0.98, 75th percentile) (eTable1, Supplemental Digital Content 1, http://links.lww.com/CCX/A126). Difference was 17% between low (68) and high (83) maximum NIRS value, and 20% between low (60.2) and high (75.1) mean value.
Combining the NIRS predictors slightly improved discrimination (eFig. 2 and eTable 2, Supplemental Digital Content 1, http://links.lww.com/CCX/A126) (AUROC, 0.69; 95% CI, 0.69–0.70; p < 0.001). However, only the RMSSD remained in the NIRS model after backward feature selection.
Finally, no interaction effect was found between NIRS predictors and mean blood pressure and arterial co2 (eTables 3 and 4, Supplemental Digital Content 1, http://links.lww.com/CCX/A126).
Performance of Clinical Model
Table 1 reports univariable differences in clinical variables between patients with and without AKI. After backward feature selection, the remaining significant predictors were baseline serum creatinine, cyanotic heart defect prior to surgery, and median blood pressure and heart frequency (p < 0.001 for all). The multivariable clinical model achieved good discrimination (Fig. 3A–C and Table 2; AUROC, 0.75; 95% CI, 0.75–0.75), a wide range of net benefit (22–93%) and was well calibrated (p = 0.56).
Comparison of NIRS and Clinical Models
The NIRS RMSSD was less discriminant for severe AKI than the clinical model (p < 0.001). However, combining them improved model sensitivity which translated in significantly higher discrimination (Fig. 3D–F and Table 2; AUROC, 0.79; 95% CI, 0.79–0.80; p < 0.001), wider ranges of net benefit (14–100%), and larger benefit in the 14–68% range compared with the clinical model (eFig. 3, Supplemental Digital Content 1, http://links.lww.com/CCX/A126). For risk thresholds lower than 14%, the highest net benefit is achieved by considering that all patients have AKI. For risk thresholds comprised between 14% and 68%, the highest net benefit is achieved by combining NIRS RMSSD with the clinical model. Finally, above 68%, the highest net benefit is achieved by using the clinical model only. Crucially, at the prevalence of AKI in this cohort and for prevalence reported in the literature (11,12), NIRS would provide benefit. These are the risk thresholds for which the tool is likely to have clinical impact. Depending on the intended use of the predictions, only if one would need a more sensitive model, would NIRS monitoring provide additional value.
Currently, there is little evidence available to support the use of cerebral NIRS oximetry after pediatric cardiac surgery. In this study, we aimed to determine the potential of this cerebral oximeter to predict AKI by prospectively monitoring children after surgery for correction of a congenital heart disease. Overall, we found that the data currently displayed in cerebral oximeters have only fair discrimination for 6-hour-ahead prediction of severe AKI. Retrospective calculation of NIRS variability with RMSSD achieved better performance, although not sufficient to outperform a clinical model based on routinely collected patient information. However, combining NIRS variability with the clinical model significantly improved predictive capabilities, suggesting that the NIRS signal carries independent relevant information in addition to routinely collected clinical data.
Cerebral NIRS oximetry is appealing as it allows measuring brain oxygen levels noninvasively and continuously. In a pediatric perioperative setting where the use of invasive catheters is not feasible or available, such a monitor has high potential provided its clinical value is proven. Currently, there is little evidence to support the use of cerebral oximetry at the pediatric patient bedside. Two randomized trials (7,27) in coronary artery bypass patients and preterm infants used an active NIRS display and treatment intervention protocol based on NIRS desaturation. They showed that NIRS monitoring avoided profound cerebral desaturation but were not powered to detect differences on mortality and morbidity. Other observational and retrospective studies have reported an association between (dose of) NIRS desaturation and worse (neurodevelopmental) outcomes (4,28,29). A single study has investigated the use of NIRS oximetry for AKI prediction in adults. The authors found that decreased renal NIRS oxygen saturation was predictive for AKI development after cardiac surgery (16). In accordance with these studies, we found that NIRS-detected low oxygen saturation was predictive for severe AKI. However, the mean and maximum NIRS values were only fair predictors of AKI, and although the time and dose of desaturation below 50% and 60% were predictive, they did not occur frequently enough for inclusion in the predictive model. When cerebral autoregulation is intact, blood flow and oxygen delivery to the brain may be preserved in low cardiac output states. Therefore, the cerebral NIRS values might not represent the perfusion to the other organs during low cardiac output and could partially explain the lack of predictive value for AKI.
Interestingly, we found that decreased NIRS variability had more discriminative power than the NIRS value currently displayed at the patient bedside. Spaeder et al (30) investigated NIRS variability previously and found that decreased NIRS RMSSD, a notion of high frequency variability, was associated with poor neurodevelopmental outcomes in neonatal survivors of congenital heart disease surgery. In a previous study (17), we found that the low frequency variability (sd-s) was associated with prolonged ICU stay and prolonged duration of mechanical ventilation. Here, as patients developed AKI early during ICU stay, the monitoring period before prediction was short which could have explain the lack of predictive capability from the sd-s. To our knowledge, this is the first study investigating the use of NIRS variability for acute clinical events, such as severe AKI. This study supports displaying such a metric at the patient bedside, as it may improve the predictive value of NIRS-based cerebral oximeters.
The current study not only reports an added predictive value of cerebral NIRS monitoring but also highlights the predictive performance of routinely collected patient information to predict severe AKI after pediatric cardiac surgery. AKI is an increasingly recognized concern in pediatric patients (31). Therefore, it is essential to identify children whose kidney function will deteriorate to provide appropriate intervention to mitigate AKI (31). For that purpose, several biomarkers have shown great diagnostic and predictive abilities (32,33). However, due to their high cost, serial measurements is not always applicable (34). The model developed in the current study could be used continuously at the patient bedside, after translation to an online predictor such as similar models for adults (35) or encoded in electronic health records.
Our study has several strengths: its prospective and blinded design that excludes treatment bias, its large sample size as compared with other studies investigating NIRS oximetry (3,5,6,27,30), and the thorough statistical analysis of the cerebral NIRS oxygen saturation which included not only value-based metrics and dose below hypoxic and above hyperoxic thresholds as previous studies (6,19), but also measures of variability and frequency.
Our study has several limitations. First, as no precise information on the cause of AKI was available, the clinical model did not include this information, nor did it include medication, while both could have influenced model performance. However, the clinical model included relevant surgical data that were associated with AKI. Second, it is unclear how the findings are affected by cerebral autoregulation. Indeed, in case of decreased systemic perfusion and intact cerebral autoregulation, the cerebral blood flow would initially remain unchanged. Therefore, the NIRS signal would not change. On the other side, in case of impaired cerebral autoregulation, disturbances in systemic perfusion would be reflected by cerebral perfusion and by the NIRS signal. Therefore, the lack of predictive power of cerebral NIRS oximetry might be explained by a late indication of reduced systemic oxygen delivery. However, we did not find any interaction effect between NIRS and mean blood pressure or arterial co2. Third, we could not compare or combine the performance of the models with renal NIRS oximetry, which might be a more sensitive marker of kidney perfusion than cerebral NIRS oximetry, in particular in case of deficient autoregulation (6,16,36). However, major drawbacks limit the use of renal NIRS oximetry (37,38). First, renal oxygen saturation is influenced by the distance between the body surface and the kidney, which largely varies between patients. Additionally, the presence of subcutaneous fat alters the measurement of light absorption by hemoglobin. These limitations explain why the most common clinical application of the NIRS technology has been in assessing cerebral oxygen saturation, unaffected by these limitations. Fourth, as a single-center study, our findings might not generalize to other centers or to different (noncardiac) populations. Although we performed internal validation via bootstrapping to improve generalizability, validation in external centers is required. Fifth, the lack of standardization between NIRS monitors additionally contributes to the difficulty of identifying NIRS critical thresholds in previous studies, and it is well known that the different devices have discordant values in certain ranges (39); therefore, the findings of our study might not generalize to different cerebral oximeters. Finally, prospective validation of the models developed in this study is warranted to identify potential clinical benefit for their use at the patient bedside.
The predictive value of cerebral NIRS oximetry as displayed in the current monitors, without monitoring cerebral autoregulation, is limited for prediction of severe AKI after pediatric cardiac surgery. However, NIRS variability, in particular combined with routinely collected patient information, showed improved discrimination. Future studies are required to validate these findings as compared with renal NIRS and to identify whether implementation of NIRS variability at the bedside could help identify children whose kidney function will deteriorate early enough to initiate mitigating interventions.
We are grateful to the PICU nurses and residents for the patient care and setting up the near-infrared spectroscopy monitoring, in particular to Stoffel Lamote, MD, Heidi Delrue, MD, and Marc Beckers, MD, for patient management and data collection; to Pieter Wouters, MSc, for database exports; to Koen Vanhonsebrouck, PICU head nurse, for the assistance in coordinating the monitoring setup; to Tom Fivez, MD, for patient management and signaling eventual technical problems; to Marc Denturck, Fredrik Hermans, and Jan Lauwers (biotechnology department UZ Leuven) for blinding the FORESIGHT monitors.
1. Lima A, Bakker J. Noninvasive monitoring of peripheral perfusion.Intensive Care Med2005311316–1326
2. Garvey AA, Dempsey EM. Applications of near infrared spectroscopy in the neonate.Curr Opin Pediatr201830209–215
3. McQuillen PS, Nishimoto MS, Bottrell CL, et al. Regional and central venous oxygen saturation monitoring following pediatric cardiac surgery: Concordance and association with clinical variables.Pediatr Crit Care Med20078154–160
4. Murkin JM, Arango M. Near-infrared spectroscopy as an index of brain and tissue oxygenation.Br J Anaesth2009103Suppl 1i3–13
5. Kaufman J, Almodovar MC, Zuk J, et al. Correlation of abdominal site near-infrared spectroscopy with gastric tonometry in infants following surgery for congenital heart disease.Pediatr Crit Care Med2008962–68
6. Chakravarti SB, Mittnacht AJ, Katz JC, et al. Multisite near-infrared spectroscopy predicts elevated blood lactate level in children after cardiac surgery.J Cardiothorac Vasc Anesth200923663–667
7. Murkin JM, Adams SJ, Novick RJ, et al. Monitoring brain oxygen saturation during coronary bypass surgery: A randomized, prospective study.Anesth Analg200710451–58
8. Tweddell JS, Ghanayem NS, Hoffman GM. Pro: NIRS is “standard of care” for postoperative management.Semin Thorac Cardiovasc Surg Pediatr Card Surg Annu20101344–50
9. Hirsch JC, Charpie JR, Ohye RG, et al. Near infrared spectroscopy (NIRS) should not be standard of care for postoperative management.Semin Thorac Cardiovasc Surg Pediatr Card Surg Annu20101351–54
10. Aneman A, Brechot N, Brodie D, et al. Advances in critical care management of patients undergoing cardiac surgery.Intensive Care Med201844799–810
11. Singh SP. Acute kidney injury after pediatric cardiac surgery.Ann Card Anaesth201619306–313
12. Blinder JJ, Asaro LA, Wypij D, et al. Acute kidney injury after pediatric cardiac surgery: A secondary analysis of the safe pediatric euglycemia after cardiac surgery trial.Pediatr Crit Care Med201718638–646
13. Li S, Krawczeski CD, Zappitelli M, et al.; TRIBE-AKI ConsortiumIncidence, risk factors, and outcomes of acute kidney injury after pediatric cardiac surgery: A prospective multicenter study.Crit Care Med2011391493–1499
14. Morgan CJ, Zappitelli M, Robertson CM, et al.; Western Canadian Complex Pediatric Therapies Follow-Up GroupRisk factors for and outcomes of acute kidney injury in neonates undergoing complex cardiac surgery.J Pediatr2013162120–127.e1
15. Greenberg JH, Zappitelli M, Jia Y, et al. Biomarkers of AKI progression after pediatric cardiac surgery.J Am Soc Nephrol2018291549–1556
16. Choi DK, Kim WJ, Chin JH, et al. Intraoperative renal regional oxygen desaturation can be a predictor for acute kidney injury after cardiac surgery.J Cardiothorac Vasc Anesth201428564–571
17. Flechet M, Güiza F, Vlasselaers D, et al. Near-infrared cerebral oximetry to predict outcome after pediatric cardiac surgery: A prospective observational study.Pediatr Crit Care Med201819433–441
18. Kidney Disease: Improving Global Outcomes (KDIGO) acute kidney injury work groupKDIGO clinical practice guideline for acute kidney injury.Kidney Int Suppl201221–138
19. Slater JP, Guarino T, Stack J, et al. Cerebral oxygen desaturation predicts cognitive decline and longer hospital stay after cardiac surgery.Ann Thorac Surg20098736–44discussion 44–45
20. Yao FS, Tseng CC, Ho CY, et al. Cerebral oxygen desaturation is associated with early postoperative neuropsychological dysfunction in patients undergoing cardiac surgery.J Cardiothorac Vasc Anesth200418552–558
21. von Elm E, Altman DG, Egger M, et al.; STROBE InitiativeThe strengthening the reporting of observational studies in epidemiology (STROBE) statement: Guidelines for reporting observational studies.Ann Intern Med2007147573–577
22. Efron B, Tibshirani R. Improvements on cross-validation: The 632+ bootstrap method.J Am Stat Assoc199792548–560
23. Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: A framework for traditional and novel measures.Epidemiology201021128–138
24. Nattino G, Finazzi S, Bertolini G. A new calibration test and a reappraisal of the calibration belt for the assessment of prediction models based on dichotomous outcomes.Stat Med2014332390–2407
25. Vickers AJ, Elkin EB. Decision curve analysis: A novel method for evaluating prediction models.Med Decis Making200626565–574
26. Fitzgerald M, Saville BR, Lewis RJ. Decision curve analysis.JAMA2015313409–410
27. Hyttel-Sorensen S, Pellicer A, Alderliesten T, et al. Cerebral near infrared spectroscopy oximetry in extremely preterm infants: Phase II randomised clinical trial.BMJ2015350g7635
28. Fischer GW, Lin HM, Krol M, et al. Noninvasive cerebral oxygenation may predict outcome in patients undergoing aortic arch surgery.J Thorac Cardiovasc Surg2011141815–821
29. Scott JP, Hoffman GM. Near-infrared spectroscopy: Exposing the dark (venous) side of the circulation.Paediatr Anaesth20142474–88
30. Spaeder MC, Klugman D, Skurow-Todd K, et al. Perioperative near-infrared spectroscopy monitoring in neonates with congenital heart disease: Relationship of cerebral tissue oxygenation index variability with neurodevelopmental outcome.Pediatr Crit Care Med201718213–218
31. Jefferies JL, Devarajan P. Early detection of acute kidney injury after pediatric cardiac surgery.Prog Pediatr Cardiol2016419–16
32. Dong L, Ma Q, Bennett M, Devarajan P. Urinary biomarkers of cell cycle arrest are delayed predictors of acute kidney injury after pediatric cardiopulmonary bypass.Pediatr Nephrol2017322351–2360
33. Parikh CR, Devarajan P, Zappitelli M, et al.; TRIBE-AKI ConsortiumPostoperative biomarkers predict acute kidney injury and poor outcomes after pediatric cardiac surgery.J Am Soc Nephrol2011221737–1747
34. de Geus HR, Betjes MG, Bakker J. Biomarkers for the prediction of acute kidney injury: A narrative review on current status and future challenges.Clin Kidney J20125102–108
35. Flechet M, Güiza F, Schetz M, et al. Akipredictor, an online prognostic calculator for acute kidney injury in adult critically ill patients: Development, validation and comparison to serum neutrophil gelatinase-associated lipocalin.Intensive Care Med201743764–773
36. Hoffman GM, Ghanayem NS, Scott JP, et al. Postoperative cerebral and somatic near-infrared spectroscopy saturations and outcome in hypoplastic left heart syndrome.Ann Thorac Surg20171031527–1535
37. Steppan J, Hogue CW. Cerebral and tissue oximetry.Best Pract Res Clin Anaesthesiol201428429–439
38. Gist KM, Kaufman J, da Cruz EM, et al. A decline in intraoperative renal near-infrared spectroscopy is associated with adverse outcomes in children following cardiac surgery.Pediatr Crit Care Med201617342–349
39. Schmidt C, Heringlake M, Kellner P, et al. The effects of systemic oxygenation on cerebral oxygen saturation and its relationship to mixed venous oxygen saturation: A prospective observational study comparison of the INVOS and foresight elite cerebral oximeters.Can J Anaesth201865766–775