Artificial Intelligence to Improve Health Outcomes in the NICU and PICU: A Systematic Review

This systematic review was reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analysis guidelines.³⁸  Our protocol was registered with the Open Science Framework on October 21, 2020.³⁹ 

We searched the PubMed, IEEE Xplore, Cochrane, and Web of Science databases to identify available, peer-reviewed articles within the scope and eligibility criteria of this systematic review. All databases were last accessed in January 2021. Search terms and phrases were selected to identify studies in which researchers described (1) AI, (2) pediatrics and (3) intensive care. Specifically, AI terms included deep learning, reinforcement learning, and supervised or unsupervised ML–supervised ML learning models are trained by using labeled structure data typically for classification and regression type problems, and unsupervised models are used to extract patterns from unlabeled data typically for clustering analysis. Keywords were initially chosen on the basis of a preliminary analysis of the literature and Medical Subject Headings (MeSH) terms. They were then modified on the basis of feedback from subject experts as well as our institution’s librarian. Our completed search strategy is shown in Supplemental Table 3.

We included articles that satisfied the following criteria: (1) use of AI algorithms in a NICU or PICU setting, (2) description of the impact of AI on at least 1 patient health outcome (eg, mortality), (3) pediatric-only populations (defined as patients <18 years of age) and (4) publications printed in English from June 1, 2010, to May 31, 2020. Studies involving adult patients (defined as >18 years of age) or that included secondary or gray literature (eg, editorials) were excluded.

Using a multidatabase query, we identified 287 publications, of which 32 met inclusion criteria after the title, abstract, and full-text review (Fig 1). After removing duplicates, 2 authors (C.A. and A.C.) independently evaluated all publications for eligibility. The initial title and abstract screening of each publication was followed by a full-text review. To minimize selection bias, all discrepancies were resolved in a stepwise fashion, through discussion, and required consensus from 2 additional authors (O.A. and M.K.).

Two authors (C.A. and A.C.) used a data abstraction form to independently record details from each publication, including study aim(s) and/or objective(s), design, setting, patient population, disease process and health outcome(s), AI type and area of impact. We also captured data types and sources, attributes defining the AI classification and learning types, and whether results were cross-validated and compared with clinicians. Studies were then categorized based on AI impact on outcomes (direct or indirect) as well as by impact focus area, all defined below.

The studies in our review had marked heterogeneity in design and outcomes. Coupled with the multiple classification and mixed learning types, this level of heterogeneity precluded our ability to combine results. Given that statistical and study heterogeneity reflects a variability in outcomes so great that it cannot be explained by measurement error alone, a meta-analysis of pooled data was not performed.

The impact of AI on health outcomes was assessed by using a definition and focus area framework developed by the European Institute of Innovation and Technology (EIT) Health joint report.⁴⁰  In this report, the authors defined direct impact as AI that augmented any patient health outcome after clinical diagnosis or that aided physicians by allowing them to “spend more time in direct patient care (while reducing provider burnout).”⁴⁰  We derived the term indirect impact as the inverse of this concept and defined it as AI that did not directly, or actively, change the patient health outcome after clinical diagnosis or aid physicians in spending more time in direct care. One such example involves early detection of critical events in infants with single-ventricle physiology between first- and second-stage cardiac repair.⁴¹  Although this AI modality had a “statistically significant higher performance” compared with expert-defined assessment tools, the authors did not report an improvement in patient health outcomes following its implementation.

In another European technical report, the authors detail a methodology for assessing AI real-world applications using technology readiness levels (TRLs) (Fig 2). TRL is used to classify any given technology into 9 categories. TRL 1 is assigned to technologies with that have principles observed and reported (lowest level to technology readiness); TRL 2 is assigned to technologies with a ready concept and/or application formulated; TRL 3 identifies technologies with a proof of concept; TRL 4 and 5 are for those technologies with tests carried out in the laboratory or components validated in the relevant environment, respectively; TRL 6 and 7 identifies technologies with a prototype functioning in a relevant environment or operational environment; TRL 8 is for technologies that are completed and qualified through test and demonstration; and finally TRL 9 is for technologies that have been proven through successful mission operations. These 9 TRLs are used to evaluate the stages of AI real-world application ranging from laboratory to operational environments, with research to implementation goals and evaluations from the prototype to deployment phases.⁴²  We classified all included studies using the TRL framework. For AI applications in some stage of bedside or clinical implementation, we further categorized them into the 6 impact focus areas in accordance with the EIT Health joint report: self-care/prevention/wellness, triage and early diagnosis, diagnostics, clinical decision support, care delivery, and chronic management.⁴⁰ 

In Table 1, we summarize study characteristics. The majority of study participants were premature infants or newborns hospitalized at tertiary academic centers. In several studies, researchers used AI to provide early diagnosis or improvement of outcomes among diseases like sepsis and severe respiratory infections. Researchers used AI models, particularly deep learning frameworks and neural networks, to enhance survival prediction, hone patient risk stratification, and modify diagnostic approaches. Our analysis included 22 NICUs and 10 PICUs. Approximately one-half (53%; n = 17) of studies took place in ICUs outside of the United States, of which 12 studies originated from high-income countries and 5 from low- to middle-income countries. Among the 32 studies, 14 had a retrospective cohort design, whereas 5 were described by authors as prospective longitudinal, 4 as atypical, 3 as exploratory, 2 as clinical trials, 2 as case-control, and 2 as observational.

A direct impact of AI on at least 1 health outcome was described in 7 studies (22%). In each study, researchers documented an improvement in health outcome after AI implementation. These studies aligned with the following 4 impact focus areas: triage and early diagnosis (43%; n = 3), clinical decision support (29%; n = 2), diagnostics (14%; n = 1), self-care/prevention/wellness (14%; n = 1) (Table 1). For example, AI was used in predictive models to identify and reduce the risk of morbidity and mortality among children with clinical illness.^30,41,43  In one study, trained, tested, and validated AI algorithms were used to successfully identify preterm infants with congenital heart disease who were at risk for respiratory syncytial virus and required vaccinations to prevent severe illnesses and increased mortality.⁴⁴  These preterm infants were noted to have a decreased oxygen need and an overall improved health outcome following vaccination. In another study, AI was used to accurately diagnose 3 out of 7 critically ill PICU patients with genetic diseases.⁴⁵  Finally, researchers from Spain used AI to develop relaxing music tunes for preterm infants in the NICU.⁴⁶  These neonates had reduced stress levels, identified by changes in vital signs.

In the remaining 25 studies, researchers described AI algorithms with a more limited, indirect impact on health outcomes. In these studies, AI models outperformed current non-AI modalities (eg, expert human opinion and statistical analysis models), However, a change in outcome, such as reduced disease burden, improved recovery times, and decreased length of stay or mortality, was not reported. Other outcome measures, such as surgical outcomes or effects on morbidity or mortality, were also not reported. None of the studies suggested AI use was associated with worsened health outcomes or patient care.

Applying the TRL system to the articles in our review, only 1 AI model and/or algorithm was in the proof-of-concept phase. Majority were in prototype testing in an operational environment with near-implementation readiness. A handful of studies were in the implementation phase, and had been either certified for use or deployed into an ICU setting (Table 1).

In Table 2, we detail AI model data sources and classification metrics. Most relied on electronic medical record (EMR) data, and 29 had defined statistical classification types. In the remaining 3 articles, there were no clear details on the statistical method used to compare AI models to more traditional modalities. Binary classification was the predominant classification type and represented 47% (n = 15) of included studies. In addition to this statistical categorization, researchers also reported a measurable quantitative outcome to compare AI models to current gold standards. Quantitative outcomes or metrics ranged from area under the receiver operating curve (AUROC) (most common at 56% [n = 18] to sensitivity and/or recall [47%; n = 15] to specificity [38%; n = 12]) (Supplemental Table 4). Researchers used anywhere from 1 to 6 different types of quantitative outcomes or metrics. For instance, in 1 study, researchers used accuracy, AUROC, recall, precision, F-measure, and specificity, all to improve reporting quality. Two articles discussed the analysis of AI’s impact on health status, but did not clearly define the quantitative metric or outcome.

AI has the potential to transform pediatric care by improving clinical decision-making and optimizing care delivery.⁴⁷  This is particularly true in fast-paced, critcal care environments where physicians regularly make life-saving decisions. This systematic review is the first in which researchers explore the use of AI to improve health outcomes for patients in NICUs and PICUs. Bedside application of AI was identified in 7 studies, and all were associated with an improved health outcome. Additionally, the articles in this review, were categorized into specific impact focus areas, and appear consistent with the makeup of health care AI today: applications to support clinical decision-making, patient-triage, and early disease detection.

Although some experts believe AI will change the future of health care, most emerging AI technologies have yet to be adopted by the general medical community.⁴⁸  Technological breakthroughs do not always translate into tools that are ready for real-world use. AI has certainly made strides in adult medicine, particularly in inpatient and ICU settings,^49,50  but its use in pediatric medicine lags behind.⁵¹  AI health care applications, like most other tools and medical interventions, must undergo rigorous testing before being adapted into everyday clinical practice.⁴⁸  AI use in neonatal and pediatric critical care may lag behind other fields of medicine because of uncertainties surrounding the safety and efficacy of these tools and techniques among particularly vulnerable populations like extremely preterm infants. There is a dearth of research on the use of AI in low- to middle-income countries in which many neonates and children require intensive care treatment. The impact of AI in such settings is an important area for future investigation.

Applying the TRL system to the articles in our review, few were in the implementation phase or deployed into an ICU setting. The AI models in these advanced phases were focused on narrow tasks, which is consistent with the original European Commission report, revealing that higher TRLs are achievable for AI technologies that are focused on specific capabilities.⁴²  Because it is difficult to predict which technologies will become an integral part of everyday care in the NICU and PICU, the TRL system may provide some insight into which AI models are making headway into pediatric critical care.

In majority of studies, researchers described AI algorithms that outperformed conventional modalities, and may have indirectly impacted health outcomes. Whether these AI modalities will lead to actual changes in health outcomes for patients is unknown. Future AI research will need to use prospective, experimental study designs to assess the impact of AI on outcomes (eg, disease burden, length of stay, readmissions, and mortality), and use reliable, validated measures and implementation frameworks.

It was difficult to assess the AI models in our review because of marked heterogeneity in reporting. For instance, in 1 study, researchers measured AI model performance on the basis of the AUROC quantitative outcome while others used recall. Metrics like AUROC may be deemed superior^52–54 ; however, they require that researchers have extensive technical knowledge to provide appropriate analysis.^8,9,55  Although it is reasonable for some AI models to use different performance metrics, this level of heterogeneity in reporting makes the comparison and evaluation of AI algorithms challenging. Moreover, the translation of these algorithms into clinical practice is also impaired because of heterogeneity in the algorithms used, the data used for training, and a lack of spectral or geographical validation. AI models need to be comparable when analyzing data from the same target population and when using metrics to assess overall AI performance. Otherwise, medical decisions made by using inappropriate AI algorithms could lead to poor patient outcomes.

AI technologies must be able to decipher and translate gigabytes of clinical data to function as effective decision support systems and aid intensivists in real-time. Notably, though, the subject of this review is focused on intensive care, many of these challenges may be broadly applicable across pediatric inpatient settings. A significant challenge in applying AI technologies is that the inpatient data needed for translation and training are typically complex, heterogenous across institutions⁵⁰  and largely unstructured.^{9,39,50,56–58}  Although digital data has become more easily accessible for analysis, it is not regularly sampled or accurately measured and recorded. Unstructured text data, like clinical notes in most inpatient settings, necessitate the use of NLP.⁵⁹  Unfortunately, annotating large chunks of text to train any model is expensive and time consuming.⁵⁹  In our review, there was only 1 study in which researchers used NLP, suggesting that this technology remains a more nascent tool among AI developers in settings of neonatal- and pediatric critical care.

Structuring, annotating, and analyzing unstructured data is often cumbersome. Very few studies^60,61  developed researchers have been able to develop AI-based approaches for extracting unstructured data from the EMR. This poses a dilemma in adopting AI into hospital-based practice because physicians rely on unstructured EMR data to understand complex disease processes. In addition, AI models trained on unstructured and/or potentially biased input data might also generate misleading and biased outputs and negatively affect patient care.^62–64  The studies in our review neither check for bias in data nor discuss specific techniques used to handle unbalanced or biased data, which contributes to the lack of reproducibility of results.

Algorithms trained on structured data can make AI input more user-friendly, ensure the extraction of quality data, and improve AI performance.⁶⁵  Future directions should explore a concept known as “representation learning.” In representation learning, AI models automatically learn features and use a predictive model to produce an abstract representation of individual patient data for clinical data extraction.^66,67  Apart from the technical limitations, adopting AI into neonatal and pediatric intensive care can also be hindered by several human factors and system-integration barriers. Future studies should focus on factors that influence clinician-trust in AI and explore the impact of AI on clinical workflow. The use of AI over time may also induce cognitive biases among providers. For instance, clinicians may begin to blindly trust the information generated by the AI system (because of its consistent and highly efficient performance) or reject AI output data without ultimately considering its effect on outcome (because of past experiences). From a quantitative perspective, AI trained on a particular subset of patients may be biased and produce results that hold good only for that particular patient demographic. To address such issues, future researchers will need to implement the aid of computer scientists, human factors engineers, and medical experts.

In addition to the barriers surrounding AI and data extraction from the EMR, AI algorithms are often unable to decipher the complexities and comorbidities involved in pediatric inpatient care. AI algorithms use all potential and available signals to achieve the best possible performance. Unreliable confounders, however, may be included and, as a result, impair the algorithm’s ability to generalize to new data sets.^9,39,68  It is, thus, necessary to understand the input data on which an AI has been trained and compared. In studies where algorithms were compared with physicians, the assumption that every physician is a gold standard may not necessarily be true. Several studies had a retrospective research design, meaning that researchers used historically labeled data to train and test algorithms (supervised learning). Importantly, AI's utility can only be realized through prospective studies because AI performance is likely to worsen when encountering real-world data that differs from algorithm training.

Finally, most studies simplified pediatric complications into binary classification types: assigning an individual to 1 of 2 categories (eg, disease versus no disease) by measuring a series of attributes. However, such an approach neglects that a patient might suffer from multiple comorbidities, each having a different severity or interdependency. Given that barriers to adopting AI are seldom discussed in the methodology of these studies, we did not include a summary or analysis of these factors in our review. This is another important area for future investigation.

This review has important limitations that should be considered. Our scope was limited to the last 10 years; coupled with the fact that AI technology and its conceptual definition, are rapidly evolving, it is possible that all relevant articles may not have been captured. To assess AI's impact on health outcomes, we used the EIT Health joint report, which was designed with adult health care settings in mind. This categorization system, however, remained applicable to our patient population in a majority of the focus areas. We excluded non-English articles, which may have impaired our ability to describe the use of AI in an even broader population. Finally, we were unable to conduct a meta-analysis because of the heterogeneity in study designs and outcomes measured.

There is growing literature supporting the use of AI technologies to improve pediatric health outcomes; however, it will be some time before AI will achieve widespread implementation into everyday pediatric hospital practice. Barriers remain to using supervised AI models for inpatient data. Input training on supervised models does not allow for effective application of AI in real-world scenarios (eg, complex critical illnesses with comorbidities), and AI typically has difficulty processing unstructured data for clinical data extraction. Although there is a need for more research investigating the use of AI technologies, particularly around handling unstructured data and unsupervised analysis, the potential benefits of AI in pediatric hospital-based care are yet to be fully realized.