Introduction
With advances in pediatric intensive care, the survival rate of children admitted to the pediatric intensive care unit (PICU) has increased dramatically in the past decades [
1,
2]. Yet, long-term morbidity after PICU admission is a growing concern [
3,
4]. Sequelae are described in physical, neurocognitive, and psychosocial health [
3‐
7]. Adverse neurocognitive outcomes are known to interfere with development in other major domains of functioning, such as physical and mental health [
8,
9], academic achievement [
10], and socioeconomic success (as measured by education, occupation, and income) [
11], highlighting neurocognition as an important outcome after PICU admission.
In the literature, multiple pathophysiological mechanisms have been proposed that may contribute to long-term neurocognitive outcome of critically ill patients, including hypoxia, metabolic derangements such as glucose dysregulation and ischemia [
12‐
14]. Such mechanisms may be triggered by the underlying disease [
15], the critical deterioration [
7,
16], and/or the associated treatments at the PICU [
17,
18]. In addition, also demographic characteristics such as age at PICU admission, age at follow-up, sex, and socioeconomic status have been found related to neurocognitive outcome after PICU admission [
7,
18‐
20]. As understanding of the origin of difficulties in neurocognitive functioning is a prerequisite for successful prevention and intervention, it is important to unravel the factors that affect neurocognitive functioning after PICU admission.
Digitalization of health care provides increasingly more data that can importantly contribute to better prediction and understanding of long-term outcome after PICU admission. Nevertheless, the increasing wealth of clinical data produced by medical devices involves very long time series representing a great number of characteristics with potential complex inter-relations that are relevant for outcome. Therefore, novel data sources challenge conventional statistical methods, such as linear regression, which are not suitable to handle larger numbers of predictors and have limited potential to capture complex relations between predictors and outcome. Compared to conventional statistics, machine learning has great potential to capture this complexity thanks to the capability to process vast amounts of data and model non-linear and highly complex interactions [
21]. Machine learning is a rapidly growing field of artificial intelligence that is increasingly applied in health care settings [
22‐
25]. Given the large number of factors and mechanisms that have been implicated on the long-term neurocognitive outcome of critically ill patients, machine learning may have added value compared to linear regression to improve neurocognitive outcome prediction. However, the value of machine learning in investigating the relation between PICU admission and long-term neurocognitive outcome has not been investigated thus far and is therefore currently unclear.
This study aims (1) to elucidate the potential relevance of patient and PICU-related characteristics for long-term adverse neurocognitive outcome after PICU admission for bronchiolitis, and (2) to perform a preliminary exploration of the potential of machine learning as compared to linear regression to improve neurocognitive outcome prediction in a relatively small sample of children after PICU admission.
Discussion
This study aimed (1) to elucidate the potential relevance of patient and PICU-related characteristics for long-term adverse neurocognitive outcome after PICU admission for bronchiolitis, and (2) to perform a preliminary exploration of the potential of machine learning as compared to linear regression to improve neurocognitive outcome prediction in a relatively small sample of children after PICU admission. The results provide no evidence for the added value of machine learning models as compared to conventional linear regression analysis in the prediction of long-term neurocognitive outcome after PICU admission for bronchiolitis. As may be expected, linear regression analysis revealed that neurocognitive outcome was associated with demographic and perinatal characteristics (socioeconomic status, age at follow-up, and birth weight). Moreover, children with greater exposure to acidotic events during PICU admission for bronchiolitis had poorer verbal memory outcome. As the involvement of the central nervous system in the pathology of bronchiolitis is unlikely [
26,
27], the relation between acidotic events and neurocognitive outcome may reflect either potentially harmful effects of acidosis itself, or reflect related processes such as hypercapnia, hypoxic, and/or ischemic events during PICU admission.
Given the large number of factors and mechanisms that have been proposed to contribute to long-term neurocognitive outcome of critically ill patients, characteristics of machine learning models (such as flexibility, ability to model non-linear relationships, more advanced inherent selection strategies) may provide potential to improve neurocognitive outcome prediction. We used machine learning in the current sample to perform a preliminary exploration of the potential value of machine learning to improve outcome prediction in a relatively smaller sample, although comparable in size to other post-PICU follow-up studies [
53]. Regarding comparison of prediction models, we found no evidence for added value of the Regression Trees and k-Nearest Neighbors machine learning models as compared to conventional linear regression analysis. The wide confidence intervals, potentially reflecting the small sample size of the blind test set, provided limited sensitivity for model comparisons. Nevertheless, the findings suggest that machine learning models may not have added value in smaller sample sizes. Although there are examples of successful machine learning applications in small datasets [
54], machine learning flourishes by large datasets not easily obtained in clinical settings [
55]. This further stresses the importance of multicenter (international) collaborations [
56] to pool clinical data and acquire larger datasets for clinical research into advanced outcome prediction using machine learning. In this study, model performance (assessed by
R2) was not sufficient to have utility for individual outcome prediction. Nevertheless, it should be stressed that additional measures of model performance (e.g., precision and calibration) are critical to evaluate when evaluating the value of prediction models for individual outcome prediction [
57]. In this study, we found no evidence for a typical pattern of overfitting (i.e., relatively high performance on training data combined with relatively low performance on blind test data). Conversely, a pattern of relatively high performance on the blind test data combined with relatively low performance on training data can be observed, indicating instable model performance. Considering the relatively larger number of predictors relative to the size of the study sample, more comprehensive data reduction and predictor selection methodology could decrease the amount of predictors for each model and potentially improve the performance of machine learning in future work, and is considered particularly important for smaller samples.
The results of our study further show that lower socioeconomic status was associated with lower intelligence after PICU admission for bronchiolitis. Abundant research has documented the relation between lower socioeconomic status and poorer neurocognitive functioning, of which the origin is matter of debate [
18,
19,
28]. For example, poverty in early childhood and adverse environmental influences have been found related to neurocognitive functioning later in life [
28]. In addition, literature shows that enriched environments throughout development influence brain plasticity and gene expression and resultant phenotypic cognitive traits [
28]. We also observed that younger age at follow-up was associated with poorer neurocognitive functioning (i.e., poorer speed and attention and verbal memory). Most likely, this finding reflects a developmental effect, i.e., reflecting the commonly observed age-related improvements in neurocognitive functioning [
42]. Children with younger age at follow-up also had shorter recovery time (
r = .98), which could theoretically also have contributed to relatively poorer neurocognitive performance in younger children. Indeed, literature shows an association between younger age at follow-up and poorer neurocognitive functioning in some PICU subgroups, such as children admitted after heart- or heart-lung-transplantation [
7], although contradicting findings have been reported in children and adolescents who survived meningococcal septic shock [
20]. Furthermore, lower birth weight was associated with lower intelligence and poorer verbal memory. This result is consistent with existing work reporting an association between lower birth weight and poorer neurocognitive functioning [
58‐
60].
The findings of our study further suggest that greater exposure to acidotic events during PICU admission is associated with poorer verbal memory outcome. In experimental studies, several mechanisms have been proposed that may explain a potential negative effect of acidosis on the central nervous system, such as acidosis causing denaturation of proteins and nucleic acids, triggering cell swelling potentially leading to cellular edema and osmolysis, and inhibition of excitatory neurotransmission in the hippocampus, and influencing neuronal vulnerability indirectly by damaging glial cells [
61,
62]. Although the translation of these findings from the literature to our study findings is unclear, our findings indicate that acidotic events may be implicated in negative effects on the central nervous system, whether or not through other neurotoxic processes such as hypercapnia, hypoxia, or ischemia. In our exploratory analyses, we found additional evidence indicating that higher pCO
2 measurements, compatible with a respiratory origin of acidosis, were also related to poorer verbal memory outcome. Regardless of the exact mechanisms at play, our findings suggest that children with greater exposure to acidotic events are at risk of adverse long-term neurocognitive outcome after PICU admission for bronchiolitis, a finding that awaits replication in future prospective studies.
Bronchiolitis is a relatively mild indication for PICU admission that seldom manifests neurologically [
26,
27] and is therefore not expected to affect neurocognitive functioning in itself. The observed adverse long-term neurocognitive outcomes may suggest that (a combination of) secondary consequences of bronchiolitis and/or PICU treatment may negatively affect outcomes after PICU admission. In previous work, we found no evidence for a relationship between exposure to sedatives, analgesics, anesthetics (per local protocol that was used at that time at our PICU) and a range of neurocognitive outcomes in the current sample [
35]. In addition, duration of invasive mechanical ventilation was also not associated with neurocognitive outcomes [
35]. In recent years, PICU therapy for bronchiolitis shifted to less invasive mechanical ventilation and more high-flow nasal cannula, with potential relevance for long-term outcome. Nevertheless, we found no association between invasive mechanical ventilation and neurocognitive outcomes, suggesting that the shift towards less invasive ventilation is unlikely to influence neurocognitive outcome. Furthermore, other factors such as hypoxic episodes, hypotension associated with mechanical ventilation, and metabolic derangements may have negatively affected children’s neurocognitive outcome after PICU admission [
12‐
14,
63]. As understanding of the exact nature and origin of difficulties in neurocognitive functioning is a prerequisite for successful prevention and intervention, the findings of our study highlight the importance of large prospective studies aimed at identifying the combination of factors that may account for adverse neurocognitive outcome in children admitted to the PICU for bronchiolitis, and for PICU admission in general.
Although prevention strategies, such as respiratory syncytial virus vaccine in pregnancy [
64], show promising results, children will continue to be admitted to the PICU for bronchiolitis and for other admission indications. The findings of this study suggest that these children may be at risk of adverse neurocognitive outcome, even in the absence of a clear neurological manifestation of the underlying disease. Neurocognitive impairments are known to interfere with development in crucial outcome domains [
8‐
11]. In addition, the results of our previous study [
65], in which we investigated the same children included in the current study, showed that these children are at risk of long-term adverse daily life outcomes in terms of academic performance and health-related quality of life regarding school functioning 6–12 years after PICU admission for bronchiolitis. Furthermore, the findings of that study [
65] suggest that lower intelligence may contribute to academic difficulties after PICU admission. Our findings underline the importance of long-term structured follow-up after PICU admission, even in the absence of underlying disease with neurological manifestation, enabling early identification and appropriate management of adverse outcomes. Furthermore, as it is unclear whether adverse neurocognitive outcomes can be catched up later in life, it may be warranted to continue follow-up monitoring into adulthood.
A limitation of our study is that a substantial number of eligible children (45.4%) did not participate in our study, mainly because they were not reached despite our efforts. However, we deem it unlikely that this has caused important selection bias because the study sample did not differ from the total cohort of eligible children in terms of demographic characteristics and illness severity. A second limitation relates to the operationalization of socioeconomic status as the average level of parental education. The use of parental education is only one attribute of the multifaceted construct of socioeconomic status, not accounting for the roles of, for example, income and level of professional functioning [
66]. This may limit the generalizability of the study to communities with wide disparities according to for example race, ethnicity, economic opportunities, and/or insurance status. Furthermore, we acknowledge that the reported associations between risk factors and outcome may not reflect causal relationships [
67]. Important to note, is that the number of acidotic events was determined on blood gas analyses measured based on clinical signs of respiratory distress. Therefore, the number of assessed blood gas analyses varied between patients based on the presentation of clinical state. In addition, we included both arterial and capillary blood gas and lactate measures, as only a minority of the children had an arterial line. Yet, capillary blood gases accurately reflect arterial pH and pCO
2 in most PICU patients (in particular in hemodynamic stable patients) [
68]. Another limitation of this study is that we did not perform external validation of the models, such as by an independently collected dataset sample of another hospital. Therefore, the hypothesis that acidotic events may increase the risk of adverse verbal memory outcome awaits replication in future work. At last, this study has modest sample size and hence had limited statistical power [
69]. A strength of our study is that we extensively investigated patient and PICU-related characteristics in the relation between PICU admission and neurocognitive outcome. In addition, we focused on children admitted to the PICU for bronchiolitis, in an attempt to control for the confounding effect of underlying disease on outcome.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.