An Analysis of AI Research in the ICU

Written by William Hsieh, PharmD 2025

January 9, 2024

An Analysis of AI Research in the ICU

Artificial Intelligence (AI) has been widely acknowledged as a transformative force by various technology companies, including OpenAI and Meta. However, many people might overlook its application in healthcare. A growing issue in ICU patient care remains the finite number of staff available. Daily life-critical decisions must be made, and healthcare providers must be able to make these essential decisions with limited time. Sikora et al. discuss the importance of optimizing how critical care is conducted regarding critical care pharmacists. However, not all critically ill patients may receive care from critical care pharmacists due to a lack of personnel or time. As a result, positive patient outcomes may not always be achievable. Several studies have looked into the use of AI to tackle these challenges in this rapidly evolving field.

To define the terms, AI refers to the science of developing intelligent software that can mimic human cognitive abilities. Machine learning, which refers to software's capacity to learn without explicit programming, is responsible for this. This can be achieved via several methods, such as Random Forest, XGBoost, Bayesian networks, and support vector machines. With these methods (and more), studies have found that AI excels at analyzing large volumes of data to pinpoint trends. This skill is very relevant when examining the use of statistics in healthcare, which the authors' NEJM editorial "Looking back on the millennium in medicine" lists as one of the 11 most significant advancements in medical science over the previous 1000 years. 

Hunter et al. dive into the role of AI within medical statistics by testing various AI models and have concluded that medical statisticians should embrace AI. Although current models still require a ‘human-in-the-loop’ design, AI has shown great promise in tackling large data sets and narrowing them down into more specific features. Applications of these methodologies for ICU usage have been discussed by Giannini et al., where machine learning algorithms were tasked with predicting severe sepsis and septic shock from patient data. Multiple trials of machine learning algorithms with real-time EHR data sets demonstrated the feasibility of prediction. However, the alerts generated did not significantly alter patient outcomes, potentially due to alert fatigue. Thus, more algorithmic training would be required to improve efficacy. Nonetheless, it demonstrates strong potential for critical care usage of AI.

In terms of AI usage in the ICU, Nguyen et al. emphasized the ability of AI to handle large volumes of data and perform complex pattern analysis. The study noted that AI algorithms are showing promise in sepsis prediction, achieving early and accurate diagnoses that outperform traditional methods. However, Nguyen et al. emphasized how AI tools still require careful interpretation from a clinical perspective, considering factors like training data bias, comparison to baseline models, and outcome-focused evaluations. Major issues within the AI models stem from how they are often trained on datasets that may not adequately represent diverse populations, particularly minority groups. If these datasets lack relevant social factors such as ethnicity or socioeconomic status, there is a risk that these minorities may be neglected or receive biased healthcare outcomes. This could worsen existing social and cultural inequalities, as AI systems may perpetuate or amplify biases present in the training data.

Mohammad et al. detail the usage of machine learning models to help validate a medication regimen complexity scoring tool for critically ill patients. The study employed the Medication Regimen Complexity-Intensive Care Unit (MRC-ICU) scores as a specific measurement tool. A total of six machine learning models were evaluated in this study, namely K-Nearest Neighbors (KNN), Naive Bayes (NB), random forest, support vector machines (SVM), neural networks, and logistic classifiers (LC). LC had the best overall performance and accuracy. The inclusion of MRC-ICU scores significantly improved mortality predictions, highlighting the importance of the MRC-ICU score and the APACHE III score in determining patient outcomes. As a result, these findings suggest that using the MRC-ICU score and other traditional prognostic tools can improve predictions of death, which could change how resources are allocated and how patients are cared for in critical care settings. However, limitations such as the small sample size, the need for more detailed data, and the study's single-center nature may affect its external validity. Mohammad et al. help establish the foundation for incorporating more comprehensive data into ML models to predict patient outcomes and optimize resource allocation.

Overall, there has been a large volume of research on the usage of AI in critical care; however, many areas still require improvements. AI algorithms are often prone to false positives and require sensitive fine-tuning. At the current time of writing, 2023, these algorithms cannot be ethically used to diagnose patients without human guidance. There have yet to be any significant legal regulations set for these programs; however, that may change with the growing independence of software from human practitioners. Incorporating AI in intensive care units presents a promising avenue for enhancing patient care and resource management. However, challenges such as data biases and the need for diverse representation persist. AI's role in ICU settings is evolving, necessitating a balance between technological and human expertise. Despite the challenges, the potential of artificial intelligence (AI) to significantly transform healthcare efficiency and improve patient outcomes is considerable and continuously expanding.

References:

1. Newsome AS, Murray B, Smith SE, et al. Optimization of critical care pharmacy clinical services: A gap analysis approach. American Journal of Health-System Pharmacy. 2021;78(22):2077-2085. doi:https://doi.org/10.1093/ajhp/zxab237

2. Looking Back on the Millennium in Medicine. New England Journal of Medicine. 2000;342(1):42-49. doi:https://doi.org/10.1056/nejm200001063420108

3. Hunter DJ, Holmes C. Where Medical Statistics Meets Artificial Intelligence. The New England Journal of Medicine. 2023;389(13):1211-1219. doi:https://doi.org/10.1056/nejmra2212850

4. ‌Giannini HM, Ginestra JC, Chivers C, et al. A Machine Learning Algorithm to Predict Severe Sepsis and Septic Shock. Critical Care Medicine. 2019;47(11):1485-1492. doi:https://doi.org/10.1097/ccm.0000000000003891

5. Nguyen D, Ngo B, vanSonnenberg E. AI in the Intensive Care Unit: Up-to-Date Review. Journal of Intensive Care Medicine. Published online September 28, 2020:088506662095662. doi:https://doi.org/10.1177/0885066620956620

6. ‌Nguyen D, Ngo B, vanSonnenberg E. AI in the Intensive Care Unit: Up-to-Date Review. Journal of Intensive Care Medicine. Published online September 28, 2020:088506662095662. doi:https://doi.org/10.1177/0885066620956620