AI is Knocking at the CICU Door

Hollywood has led us to believe that artificial intelligence (AI) belongs in the world of Michael J. Fox and a DeLorean. But AI is no longer confined to the silver screen—nor is it something to fear. Indeed, AI is already pervasive. From students using ChatGPT to fudge their essays to driverless cars, AI has carved a niche that is here to stay. The pediatric cardiac intensive care unit (CICU) is no exception, with AI applications from quality improvement to outcome prediction. If it hasn’t already, AI is making its way to a CICU near you.

An array of terminology is thrown around in the race to the top of the AI mountain. At least a basic understanding of AI is now critical, not just for researchers but for clinical practice too. AI refers to the ability of a computer to mimic human thought and capabilities. Machine learning (ML) is a pathway to AI: technologies and algorithms that allow for pattern identification, decision making, and ongoing improvement based on data and experience.

AI will by no means seize control of the CICU à la the Terminator, but it will solve riddles that have flummoxed us for years. Consider surgery cancellations, the bane of every children’s hospital. Cancellations waste resources and frustrate families and staff alike. Machine learning has been successfully employed to deepen our understanding of the etiology of cancelations and to identify patients at risk^1,2. Such models provide actionable insights, allowing for targeted interventions. Models have been built to predict which patients are likely to cancel for specific reasons such as nil per oral (NPO) violation or “no show.” Geospatial models, using community-level data on socioeconomic disparities, identify neighborhoods with children at higher cancelation risk. The authors demonstrated similarities between neighborhood risk profiles of two cities in distinct regions of the United States. Generalizable ML models could be applied in concert with quality improvement efforts to decrease lost operating room time for CICUs everywhere.

AI may also provide answers for acute kidney injury (AKI) after cardiac surgery. Significant prior research, from observational studies to bench research with biomarkers, has identified risk factors for post-operative acute kidney injury (AKI) but the picture remains incomplete. Recently, ML was used to create a prediction model for cardiac surgery-associated AKI³. In this study, a variety of—admittedly esoteric sounding— ML algorithms such as random forests, artificial neural networks, and extreme gradient boosting were applied to create multiple models. These models were subsequently double-checked with data sets from a different hospital. Interestingly, the authors compared their best ML model (extreme gradient boosting) with a “classic” regression model to demonstrate the superiority of ML over a conventional approach.

AI’s hottest ticket may be just around the corner: the “holy grail” of predicting cardiac arrest. ML presents tantalizing opportunities for enhancing both cardiac arrest research and clinical practice in the form of clinical decision support tools. For example, Ruiz et al. used routinely-collected data from the electronic health record to develop an extreme gradient boosting model to predict significant clinical deterioration, including cardiac arrest, emergency intubation, and the need for ECMO cannulation⁴. The model achieved an area under the receiver operating characteristic curve of 0.92 at 4 hours prior to deterioration and 0.82 at 8 hours, suggesting that a good proportion of events can be predicted far enough in advance to give a realistic opportunity for intervention. Now it’s the responsibility of implementation science to incorporate such ML models in clinical informatics tools to provide early warning to clinicians at the bedside.

But let’s not get ahead of ourselves. AI is neither panacea nor talisman for CICU patients and clinicians. Nothing comes free in either medicine or AI. While the predictions of classic regression are easy to understand, the same cannot be said for more complicated and sophisticated ML algorithms. The “black box phenomenon” occurs when an AI system has internal workings that lack transparency to the user, and it can pose troubling ethical conundrums. For example, is it right to act on an algorithm’s recommendation of invasive interventions to prevent decompensation if the clinician doesn’t understand why? Moreover, AI is also dependent on the quality of the data used to train it. Biased data leads to biased models, which can perpetuate or even worsen pre-existing healthcare disparities particularly if the algorithm lacks transparency⁵.

The future, though, is bright if we are careful to address the concerns upfront. Pediatric CICU patients and clinicians are already benefiting from AI. There remain many questions to be answered, and care must be taken to ensure ethical application. But at the end of the day, all things considered in balance, AI may prove itself as a valued member of the healthcare team.

References

1. Liu L, Ni Y, Beck AF, et al. Understanding Pediatric Surgery Cancellation: Geospatial Analysis. J Med Internet Res 2021;23(9):e26231.
2. Liu L, Ni Y, Zhang N, Nick Pratap J. Mining patient-specific and contextual data with machine learning technologies to predict cancellation of children’s surgery. Int J Med Inform 2019;129:234–41.
3. Luo X-Q, Kang Y-X, Duan S-B, et al. Machine Learning-Based Prediction of Acute Kidney Injury Following Pediatric Cardiac Surgery: Model Development and Validation Study. J Med Internet Res 2023;25:e41142.
4. Ruiz VM, Goldsmith MP, Shi L, et al. Early prediction of clinical deterioration using data-driven machine-learning modeling of electronic health records. J Thorac Cardiovasc Surg [Internet] 2021; Available from: http://dx.doi.org/10.1016/j.jtcvs.2021.10.060
5. Seyyed-Kalantari L, Zhang H, McDermott MBA, Chen IY, Ghassemi M. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med 2021;27(12):2176–82.