Thank you so much for the opportunity to present the data science articles for this year. And I think it builds really nicely on the data science pieces that were woven into the last three talks, and especially Nancy's description of how we can better predict outcomes for patients and integrate those outcomes to target therapies for patients in the ICU. So I have no disclosures. Articles in pediatric data science have been exploding over the last decade. There's been a seven-fold increase in publications related to pediatric data science, and there were nearly 500 publications just in this last year. So the articles that I'll highlight today will show you how data science is already integrated into your research, as well as your clinical practice, and again, what we can look towards for the future. To select the articles for this year, I first started with a broad PubMed search using pediatric critical care, the data science search terms on the right, and then the year 2023. And this resulted in 441 articles. After screening out for titles, I ended up with 180 articles, reviewed those abstracts, and ended with 73. And I won't present all of those today. But I did group those 73 articles into broad themes. And you can see the majority of the articles were related to machine learning applications, categorizing omics, predicting clinical outcomes, some of which we've already heard about, using unsupervised learning techniques to look at different phenotypes of patients and trajectory modeling, and incorporating high-frequency or continuous physiologic data into our models. There was one article in bias in data science, one article that focused on informatics. And then given that data science as a field is growing, there were many articles focusing on reviews and perspectives pieces. So taking those 73, I looked at their methods, the quality of the methods, as well as the type of modeling that was performed, the maturity in the papers, so whether the model had been validated or implemented, the reporting quality, and then the broad context and themes for each article. I ranked and scored those articles, and I ended up with seven key articles. And so I'll provide a high-level overview of those. So the first that we'll review is focusing on how machine learning tools can help us identify new patterns in omics data. This article used machine learning to examine gene transcripts and molecular pathways for children with acute hypoxemic respiratory failure to determine the gene expression signatures of mild versus moderate or severe pediatric ARDS in the first 24 hours after diagnosis. The authors analyzed two pediatric data sets. The first was an ARDS data set, and the second was sepsis-associated acute hypoxemic respiratory failure. And given class imbalances in the data sets, they used stability selection using bootstrap with replacement to select the optimal number of common genes associated with mild versus moderate or severe ARDS. Genes common to both data sets were selected for pathway analysis. Stability selection resulted in 185 common genes between the two data sets. And differential gene regulation was seen in the severe phenotypes, specifically those with a PF ratio of less than 200. The top three pathways with differential gene regulation were metabolic, ribosomal, and coronavirus disease pathways, primarily immune and inflammatory pathways. The study identified gene networks and signaling pathways that are involved in differentiating pediatric ARDS severity, and it harnessed machine learning to find patterns in the omics data that would have been really difficult to identify with traditional statistical approaches. Identifying new patterns in common diseases is important to further our understanding of the pathobiology and the heterogeneity of these diseases, and it's a very important step as we work towards precision medicine. So next we'll see how machine learning tools can support the prediction of clinical outcomes and how those predictions can augment decision making at the bedside. The hypothesis of this study was that patients with acute neurological injury are at high risk for morbidity and mortality, and that harnessing nonlinear machine learning analytics would better predict patient outcomes when compared to traditional regression models. The authors performed a secondary analysis from a cohort of over 10,000 patients from seven Capcorn institutions, and in their sub-analysis, patients were selected if they had a suspicion of acute neurological injury on admission. Multiple models were derived, which you can see listed in the middle of the figure. The logistic regression model incorporated physiologic data via the PRISM-3 score and demographic and clinical variables, and then the advanced machine learning models were fed raw physiological variables and categorical data. Bootstrapped train and validation sets were used to prevent overfitting, and after the models were trained, feature importance analysis was performed. Finally, an off-the-shelf FHIR-based application was used to demonstrate what bedside implementation could look like. The top two figures demonstrate the receiver operating characteristic curves for mortality on the left and new morbidity or mortality on the right, and the bottom two curves show the precision recall for the same outcomes. All models performed better in predicting mortality compared to new morbidity or mortality, and the best-performing models were random forest and support vector machine. The feature analysis for the random forest model are shown here, and you can see the most important features leading to prediction of the outcome, which are a low GCS score, a low temperature, non-reactive pupils, certain markers of coagulation, and blood pressure metrics. Here's a sample demonstration of a real-time web interface. EHR patient data was entered into the random forest model to predict mortality alone in the top graph and morbidity or mortality in the bottom graph, and importantly, the Shapley waterfall plots that are generated and displayed augment the prediction score so that you can see which features are leading to that prediction score for each individual patient. So this study demonstrates that machine learning models can improve the prediction of clinical and complex outcomes. Clinical decision support tools that maximize interoperability, like the FHIR-based application shown here, can accelerate bedside implementation, and furthermore, implementation can be enhanced by improving the end user's experience of the tool, so by systematically identifying and describing which features are contributing most to the prediction. Next I will review two papers that harness machine learning tools to learn new phenotypes and trajectories. The first study's aim was to test the association between clinical instability and mortality. Within-patient clinical instability was measured in three-hour blocks by examining clinical deterioration and clinical improvement. In this figure, the white boxes represent deaths and the gray boxes represent survivors. Over nearly 8,400 admissions with 312 deaths, the authors showed that there's a greater maximum clinical deterioration and a greater maximum clinical improvement among non-survivors, and an increase in patient volatility was associated with an increased risk for mortality, particularly in the first 24 hours of PICU admission. So incorporating more dynamic modeling, like clinical instability or patient volatility, may improve our ability to predict outcomes such as mortality. Next, the objective of this retrospective multi-center study was to analyze organ dysfunction trajectories of children with sepsis-associated MODS to better describe sepsis phenotypes and to determine if those phenotypes were associated with a heterogeneity of treatment effect. Across 13 US PICUs, the authors performed trajectory-based modeling to identify phenotypes based on the type, the severity, and the progression of organ dysfunction in the first 72 hours of PICU admission. They trained a random forest classifier, which was then applied to two external validation sets to assess reproducibility. And the phenotype that was associated with the highest organ dysfunction was group 2, which was termed persistent hypoxemia, encephalopathy, and shock. This phenotype was associated with a significantly higher length of stay, persistence of organ dysfunction, and in-hospital mortality when compared to the other sepsis-associated phenotypes. Next, in a propensity score-matched analysis, treatment effect was assessed between patients with the high-risk phenotype, a persistent hypoxemia, encephalopathy, and shock, and the other sepsis phenotypes. And those with the high-risk phenotype were more likely to have an improved outcome if they received hydrocortisone or albumin in the first 24 hours compared to the other group. To summarize, trajectory-based phenotypes may identify groups of patients who have a higher risk of poor outcomes, and this phenotyping, again, may help drive precision-based therapies. The next study incorporates continuous physiological data from bedside monitors, which is an important step for predicting outcomes in real time. In this pilot study, the authors aim to predict pediatric in-hospital cardiac arrest within three hours of onset. The models were trained on 162 unique features, which included demographic data, heart rate variability characteristics from ECG waveforms, vital sign metrics, including arterial blood pressure metrics, and medications. The models were fine-tuned using results from the validation set, and performance was evaluated using a holdout test set. Pressure performance was assessed using Shapley values. Here are the receiver operating characteristic and precision recall curves demonstrating performance of each of the models. The model with the best performance was XGBoost with an area under the receiver operating characteristic curve of 0.97 and under the precision recall curve of 0.8. And here's an example of a risk score for in-hospital cardiac arrest generated from the XGBoost model. This is for a five-month-old patient starting five hours prior to the cardiac arrest event. On the left, you can see the top predictors, the vital signs, and the heart rate variability metrics over that time frame. And on the right is the average risk score over time. The risk score for this patient is approximately 0 at five hours prior to the arrest. It then increases rapidly in the fourth hour prior to arrest, and then remains relatively constant in the three hours prior to arrest. The conclusions from this study are that incorporation of waveform data can improve model performance. But it's really important that we balance performance with actionability. If the model performs really well, but we only have 30 minutes of lead time to intervene, that's probably not enough. And finally, including feature analysis can help with interpretability and implementation. So next, we'll review a study that illustrates bias in physiological data. This study examined pulse oximetry in children according to race as an indirect way to measure pulse oximetry accuracy by skin pigmentation. Data was recorded for patients undergoing cardiac catheterization. Pulse oximetry, or SpO2, was captured at one-minute intervals and was able to be linked to arterial oxygen saturation, or SaO2, measurements, which were obtained and measured via co-oximetry. 774 patients were included in the study, and 26% of whom were identified as black or African-American. The first SaO2 measurement for each patient that was obtained was matched with an SpO2 value within one minute. And then multivariable logistic regression, or linear regression, excuse me, was performed to adjust for clinical characteristics. The authors found that 12% of black or African-American patients, compared to 4% of white patients, had SpO2 levels suggesting normoxemia, but in fact, their SaO2 values reported that they were actually hypoxemic. So the important conclusion from this article is that our current pulse oximetry technology likely overestimates arterial oxygen saturation in patients with increased skin melanin, and it therefore should be used with caution when training prediction models, as it can introduce bias into the model. And finally, we'll consider prediction models in the context of implementation and dissemination. In this scoping review of supervised machine learning applications published between 2000 and 2022, the Pediatric Data Science and Analytics, or PEDAL, subgroup of Polisi reviewed 141 published articles. And if you're not familiar with PEDAL, please check out the website in the upper right corner and consider joining us. PEDALs were reviewed for following methodologic guidelines, reporting quality, the outcomes of interest, if validation was performed, and if implementation was achieved. There was a broad range of methodologic approaches used across the studies, with logistic regression being the most common. The majority of studies assess mortality or ICU transfer as the primary outcome. Importantly, what we found is that over half of these prediction models have not been implemented, even on follow-up with the study authors. So the conclusion from this scoping review are that there are a variety of prediction models that are published in the literature, but there's varying consistency and quality in the development, the validation, and the implementation, as well as in the reporting. Given the low rates of implementation, data scientists should consider an earlier focus on user-centered design strategies that can get these models to the bedside. So to summarize, data science and pediatric critical care medicine crosses all domains of research. Machine learning can better synthesize complex and linear data, and it can result in an improved prediction of clinical outcomes. It can identify new phenotypes. It can help us recognize evolving trajectories of our patients, and it can help us select which patients would benefit from which treatments. It's crucial that we study how bias can be integrated unintentionally into our prediction models, as well as understand the ethical applications of machine learning and augmented intelligence. And finally, we must consider implementation of models early and flip the paradigm, thinking about how we're going to use these models at the bedside, rather than an afterthought after they're already published. Thank you very much.

pediatric data science

machine learning

precision medicine

clinical outcomes

ethical implications