Concise Critical Appraisal: Artificial Intelligence and the ICU Patient

Artificial intelligence (AI) is a hot topic in technological advancements. The recent advancements in AI systems such as ChatGPT have greatly contributed to the increasing popularity and accessibility of such systems. This popularity boom has been matched by enhancements in computing power and accessibility to diverse, abundant, and complex critical care data. Intensive care unit (ICU) clinicians continue to have difficulties anticipating deterioration, handling high heterogeneity, and providing early interventions before patients decompensate. AI and/or machine learning could have an immense impact on ICU patients. Predictive AI models use patient care data to aid clinical decision-making by informing ICU resource allocation and recognizing morbidity and mortality outcomes earlier.

ICU AI programs could replace traditional monitoring systems and clinical risk assessment tools with computers that use multidimensional and multidomain data patterns. With the recent success of large language models such as ChatGPT, AI language models can significantly decrease barriers to conducting AI research using electronic health record (EHR) data. In turn, AI is expected to allow ICU clinicians to seamlessly extract data, interpret clinical information, and immediately obtain the information they need. ICU clinicians may soon be able to quickly tackle common ICU issues, such as disorders of volume status and acid-base, sepsis, acute respiratory distress syndrome, cardiorespiratory instability, and surgical complications.

AI has been used to enhance ICU patient care in several ways, including early prediction of ICU readmission and 30-day mortality. Kessler et al recently sought to develop an AI model that predicted cardiovascular ICU readmission.¹ They used the open-access Medical Information Mart for Intensive Care (MIMIC-III) dataset to develop an AI modeling approach that was superior to machine learning approaches. A subset of 12,797 patients were selected from the MIMIC-III dataset using input value types recommended by experienced cardiac surgeons and intensivists. These values included patient weight, age, temperature, arterial line blood pressure, heart rate, oxygenation, creatinine, blood pH, potassium, sodium, hematocrit, white blood cell counts, bicarbonate, and bilirubin. Patients were then classified as either returning (readmitted to the ICU and died) or non-returning (discharged and alive).

The AI model learned from 80% of the final dataset and focused on maximizing the area under the precision-recall curve (AUPRC). The investigators found that not only did their AI model perform well in a balanced accuracy, it also outperformed and was less computationally complex than other models. They then developed an optimized decision threshold to improve their model’s precision. Using other ICU datasets focused on cardiovascular-related events from 8127 patients from 40,663 ICU stays, they found that their model was superior and focused on making the results more easily understood by humans. The study had several limitations, including the inability to test the AI model’s output against clinically validated ICU risk stratification tools.

Another area that displays promise is AI’s ability to predict post-surgical complications. Globally, at least 7 million people have complications following surgery each year, including at least 1 million deaths.² Although clinicians are routinely developing machine learning tools and algorithms aimed at curtailing these events, many have been associated with inherent setbacks, including intense complexity, interobserver variability, and lack of external validation. It is hoped that advancements in AI and machine learning will help to overcome these setbacks and provide clinicians with immediate data that are accurate, accessible, straightforward, and reliable.

For example, Lee et al collected data from 454,404 patients undergoing non-cardiac surgeries and found that their model outperformed subjective predictive models in predicting 30-day mortality while using only 12 to 18 clinical variables.³ The authors also postulated that adding intraoperative variables to their model would allow for an even more accurate prognostic prediction, which could be useful for ICU clinicians caring for postoperative patients.

Several other dynamic models have been described. Chen et al used a random forest classification dynamic model that predicted deterioration 90 minutes ahead of the crisis.⁴ Yoon et al designed a model that used a normalized dynamic risk score trajectory with a random forest model to predict tachycardia 75 minutes before it developed.⁵ Wijnberge et al described a machine learning-derived early warning system that can predict hypotension during elective noncardiac surgery.⁶ Joosten et al used a machine learning model to reduce intermediate- and high-risk surgery hypotension to 1.2% versus 21.5% using conventional methods.⁷ VentAI, a reinforcement learning algorithm, was used in mechanically ventilated patients and improved adjustments in tidal volume and positive end-expiratory pressure with target outcomes of 90-day and ICU mortality.⁸

Well-designed predictive models will have immense utility in creating decision support tools that can aid ICU clinicians in improving patient care. AI developers are collaborating with medical experts to use clinical data to develop problem-solving approaches. Language models will aid in ICU research by simplifying how researchers formulate research questions. Researchers commonly use language models to extract patient data from the EHR. AI could use sophisticated language models to greatly streamline capturing patient data for analysis while correcting for set inclusion and exclusion criteria. Moreover, researchers could simply communicate with the AI tool about how they want to expand their study or adjust for concerns about certain confounders and the EHR. Streamlining the medical research pipeline using AI will largely benefit ICUs that seek to pinpoint inefficiencies, errors, and potential areas for improvement. There is still work to be done to overcome obstacles to AI before it will become a core component of an ICU clinician’s workflow.

References

1. Kessler S, Schroeder D, Korlakov S, et al. Predicting readmission to the cardiovascular intensive care unit using recurrent neural networks. Digit Health. 2023 Jan 9;9:20552076221149529.
2. Weiser TG, Regenbogen SE, Thompson KD, et al. An estimation of the global volume of surgery: a modelling strategy based on available data. Lancet. 2008 Jul 12;372(9633):139-144.
3. Lee SW, Lee HC, Suh J, et al. Multi-center validation of machine learning model for preoperative prediction of postoperative mortality. NPJ Digit Med. 2022 Jul 12;5(1):91.
4. Chen L, Ogundele O, Clermont G, Hravnak M, Pinsky MR, Dubrawski AW. Dynamic and personalized risk forecast in step-down units. Implications for monitoring paradigms. Ann Am Thorac Soc. 2017 Mar;14(3):384-391.
5. Yoon JH, Mu L, Chen L, et al. Predicting tachycardia as a surrogate for instability in the intensive care unit. J Clin Monit Comput. 2019 Dec;33(6):973-985.
6. Wijnberge M, Geerts BF, Hol L, et al. Effect of a machine learning-derived early warning system for intraoperative hypotension vs standard care on depth and duration of intraoperative hypotension during elective noncardiac surgery: the HYPE randomized clinical trial. JAMA. 2020 Mar 17;323(11):1052-1060.
7. Joosten A, Rinehart J, Van der Linden P, et al. Computer-assisted individualized hemodynamic management reduces intraoperative hypotension in intermediate- and high-risk surgery: a randomized controlled trial. Anesthesiology. 2021 Aug 1;135(2):258-272.
8. Peine A, Hallawa A, Bickenbach J, et al. Development and validation of a reinforcement learning algorithm to dynamically optimize mechanical ventilation in critical care. NPJ Digit Med. 2021 Feb 19;4(1):32.