February 2023 - Artifical Intelligence

February 2023
EAST Monthly Literature Review

"Keeping You Up-to-Date with Current Literature"
Brought to you by the EAST Manuscript and Literature Review Committee

This issue was prepared by EAST Multicenter Trials Committee Members Mira Ghneim, MD, MS and Paul Chestovich, MD, FACS.


Thank you to Haemonetics for supporting the EAST Monthly Literature Review.


In This Issue: Artificial Intelligence

Scroll down to see summaries of these articles

Article 1 reviewed by Mira Ghneim, MD, MS
Executive summary of the artificial intelligence in surgery series. Loftus TJ, Vlaar AP, Hung AJ, Bihorac A, Dennis BM, Juillard C, Hashimoto DA, Kaafarani HM, Tighe PJ, Kuo PC, Miyashita S, Wexner SD, Behrns KE. Surgery. 2022 May;171(5):1435-1439.

Article 2 reviewed by  Paul Chestovich, MD, FACS
Electronic health record machine learning model predicts trauma inpatient mortality in real time: A validation study. Mou Z, Godat LN, El-Kareh R, Berndtson AE, Doucet JJ, Costantini TW. J Trauma Acute Care Surg. 2022 Jan 1;92(1):74-80.

Article 1
Executive summary of the artificial intelligence in surgery series. Loftus TJ, Vlaar AP, Hung AJ, Bihorac A, Dennis BM, Juillard C, Hashimoto DA, Kaafarani HM, Tighe PJ, Kuo PC, Miyashita S, Wexner SD, Behrns KE. Surgery. 2022 May;171(5):1435-1439.

Artificial intelligence (AI) has the potential to augment surgical care and improve patient outcomes. However, the surge of published literature has garnered mixed reviews regarding the safety, efficacy, validity, and clinical implementation of AI in surgery. Therefore, this multi-disciplinary group of authors, with an expertise in AI, developed a series focused on AI in surgery in collaboration with the Editors of Surgery. The executive summary of the series is presented in this current review. The authors’ aim was to summarize the available evidence regarding the current applications, knowledge gaps, and future implications of utilizing AI in surgery.

The utility of AI in surgery can be divided into three broad categories 1) applications in predictive analytics and decision support 2) applications for intraoperative and technical skill assessment 3) applications for semi-autonomous augmentation of surgical care delivery. AI application in research has shown promise in each category. This includes multi-objective optimization in perioperative decision making related to post-operative opioid use. Reinforcement in surgical decision making through reduction in mortality of patient’s where resuscitation was guided by the clinician’s actions and an associated data driven recommendations algorithm. Perioperative risk stratification through highly complex machine learning algorithms that are accessible through mobile device applications. Machine learning has also been used to predict surgeon experience based on robotic instrument kinematic data. Microrobots have been used to perform simple proof-of-concept tasks such as suturing defects in the stomach wall. When it comes to AI in trauma, the current primary advantage is in geo-mapping and its ability to potentially mitigate some of the trauma related health-care inequities and maximize resource utilization in trauma systems development. This can be accomplished through AI generated trauma-patient geospatial injury data that allow the identification of optimal locations to establish or potentially relocate trauma systems to cater to the trauma patients’ needs. As for the current knowledge and technological gaps in AI, those include the utility of machine learning in surgical systems where data are limited, virtual reality and machine learning in surgical training, and guidance of intraoperative management.

In conclusion, what does the future hold for AI in surgery? There is promising evidence that AI is the way of the future in surgery as leaders in the field push to transition AI from the preclinical phase to the clinical phase. This however is contingent on the ability to standardize data sources, advance model interpretability, carefully monitor outcomes, maintain attention to ethical and legal issues regarding algorithm fairness and the potential for errors, and continue to preserve the bedside and human intuition in the decision-making process.

Article 2
Electronic health record machine learning model predicts trauma inpatient mortality in real time: A validation study. Mou Z, Godat LN, El-Kareh R, Berndtson AE, Doucet JJ, Costantini TW. J Trauma Acute Care Surg. 2022 Jan 1;92(1):74-80.

Previous studies evaluating machine learning and artificial intelligence algorithms for use in clinical care have limited clinical application because they are separated from the electronic health record and only are available retrospectively. This study evaluates the Epic Deterioration Index (EDI) in a level 1 trauma center.  This study seeks to incorporate and validate the EDI in using trauma patients.
 
The EDI is a proprietary model from Epic and incorporates 125 patient variables abstracted from the EHR and generates a composite risk score for patient deterioration. The EDI ranges from 0-100 and is calculated every 20 minutes after patient arrival and is made immediately available to the clinical treating team.
 
For this study, a retrospective analysis was performed by evaluating mortality prediction at the maximum EDI at 6, 12, 18 and 24 hours from admission.  Unplanned ICU admission was evaluated using EDI slope over 24 hours. They found that the maximum EDI at 6 hours had an AUROC of 0.91 and increased up to 0.98 at 24 hours after admission. The slope of the EDI curve had AUROC of 0.85 compared with ISS 0.89 and NISS 0.91. For outcome of unplanned ICU admission, the EDI slope was the best predictor ending 4 hours before ICU admission. AUROC was 0.66, and the max EDI was not found a good predictor of unplanned ICU admission. Mortality performance threshold of EDI was analyzed, and the ideal threshold of 80 was identified as the point at which sensitivity decreases with minimal increase in specificity. 

The EDI holds two primary advantages over traditional prediction models for outcomes: frequent updates of input parameters and access to a large pool of data. The EDI demonstrated success at predicting mortality by utilizing the 24h max EDI, and unplanned ICU admission through the EDI slope over the first 24 hours.

Machine learning algorithms generally offer improvement in clinical decision-making capability of clinicians by increasing the data inputs used to make decisions. Humans utilize an average of 6 data points to arrive at a decision, while an artificial intelligence algorithm can use dozens or hundreds of inputs. This study is the first to evaluate the EDI in a real-time study to predict patient mortality and clinical deterioration in a trauma center.