Evaluation of machine learning models as decision aids for anesthesiologists.

Machine Learning (ML) models have been developed to predict perioperative clinical parameters. The objective of this study was to determine if ML models can serve as decision aids to improve anesthesiologists' prediction of peak intraoperative glucose values and postoperative opioid requirements. A web-based tool was used to present actual surgical case and patient information to 10 practicing anesthesiologists. They were asked to predict peak glucose levels and post-operative opioid requirements for 100 surgical patients with and without presenting ML model estimations of peak glucose and opioid requirements. The accuracies of the anesthesiologists' estimates with and without ML estimates as reference were compared. A questionnaire was also sent to the participating anesthesiologists to obtain their feedback on ML decision support. The accuracy of peak glucose level estimates by the anesthesiologists increased from 79.0 ± 13.7% without ML assistance to 84.7 ± 11.5% (< 0.001) when ML estimates were provided as reference. The accuracy of opioid requirement estimates increased from 18% without ML assistance to 42% (p < 0.001) when ML estimates were provided as reference. When ML estimates were provided, predictions of peak glucose improved for 8 out of the 10 anesthesiologists, while predictions of opioid requirements improved for 7 of the 10 anesthesiologists. Feedback questionnaire responses revealed that the anesthesiologist primarily used the ML estimates as reference to modify their clinical judgement. ML models can improve anesthesiologists' estimation of clinical parameters. ML predictions primarily served as reference information that modified an anesthesiologist's clinical estimate.

Machine learning approach to predict postoperative opioid requirements in ambulatory surgery patients.

Opioids play a critical role in acute postoperative pain management. Our objective was to develop machine learning models to predict postoperative opioid requirements in patients undergoing ambulatory surgery. To develop the models, we used a perioperative dataset of 13,700 patients (≥ 18 years) undergoing ambulatory surgery between the years 2016-2018. The data, comprising of patient, procedure and provider factors that could influence postoperative pain and opioid requirements, was randomly split into training (80%) and validation (20%) datasets. Machine learning models of different classes were developed to predict categorized levels of postoperative opioid requirements using the training dataset and then evaluated on the validation dataset. Prediction accuracy was used to differentiate model performances. The five types of models that were developed returned the following accuracies at two different stages of surgery: 1) Prior to surgery-Multinomial Logistic Regression: 71%, Naïve Bayes: 67%, Neural Network: 30%, Random Forest: 72%, Extreme Gradient Boost: 71% and 2) End of surgery-Multinomial Logistic Regression: 71%, Naïve Bayes: 63%, Neural Network: 32%, Random Forest: 72%, Extreme Gradient Boost: 70%. Analyzing the sensitivities of the best performing Random Forest model showed that the lower opioid requirements are predicted with better accuracy (89%) as compared with higher opioid requirements (43%). Feature importance (% relative importance) of model predictions showed that the type of procedure (15.4%), medical history (12.9%) and procedure duration (12.0%) were the top three features contributing to model predictions. Overall, the contribution of patient and procedure features towards model predictions were 65% and 35% respectively. Machine learning models could be used to predict postoperative opioid requirements in ambulatory surgery patients and could potentially assist in better management of their postoperative acute pain.

Hyper-G: An Artificial Intelligence Tool for Optimal Decision-Making and Management of Blood Glucose Levels in Surgery Patients.

BACKGROUND: Hyperglycemia or high blood glucose during surgery is associated with poor postoperative outcome. Knowing in advance which patients may develop hyperglycemia allows optimal assignment of resources and earlier initiation of glucose management plan. OBJECTIVE: To develop predictive models to estimate peak glucose levels in surgical patients and to implement the best performing model as a point-of-care clinical tool to assist the surgical team to optimally manage glucose levels. METHODS: Using a large perioperative dataset (6,579 patients) of patient- and surgery-specific parameters, we developed and validated linear regression and machine learning models (random forest, extreme gradient boosting [Xg Boost], classification and regression trees [CART], and neural network) to predict the peak glucose levels during surgery. The model performances were compared in terms of mean absolute percentage error (MAPE), logarithm of the ratio of the predicted to actual value (log ratio), median prediction error, and interquartile error range. The best performing model was implemented as part of a web-based application for optimal decision-making toward glucose management during surgery. RESULTS: Accuracy of the machine learning models were higher (MAPE = 17%, log ratio = 0.029 for Xg Boost) when compared with that of the linear regression model (MAPE = 22%, log ratio = 0.041). The Xg Boost model had the smallest median prediction error (5.4 mg/dL) and the narrowest interquartile error range (-17 to 24 mg/dL) as compared with the other models. The best performing model, Xg Boost, was implemented as a web application, Hyper-G, which the perioperative providers can use at the point of care to estimate peak glucose levels during surgery. CONCLUSIONS: Machine learning models are able to accurately predict peak glucose levels during surgery. Implementation of such a model as a web-based application can facilitate optimal decision-making and advance planning of glucose management strategies.

Association between acute phase perioperative glucose parameters and postoperative outcomes in diabetic and non-diabetic patients undergoing non-cardiac surgery.

BACKGROUND: The relationship between acute phase perioperative hyperglycemia and postoperative outcome is poorly understood. METHODS: Retrospective cohort study of diabetic and non-diabetic adult patients undergoing non-cardiac surgery. Mean glucose and glycemic variability during the intraoperative and immediate postoperative periods were compared to length of stay, 30-day mortality, and postoperative complications. RESULTS: . DIABETIC PATIENTS (N = 1096): Higher glycemic variability was associated with longer hospital length of stay (0.32 day per 10 mg/dL) and greater 30-day mortality risk (OR = 1.42). Higher mean glucose (OR = 1.07) and glycemic variability (OR = 1.11) were associated with higher risk of complications. NON-DIABETIC PATIENTS (N = 1012): Both higher mean glucose (0.29 day per 10 mg/dL) and higher glycemic variability (0.68 day per 10 mg/dL) were associated with longer hospital length of stay. Both higher mean glucose (OR = 1.13) and higher glycemic variability (OR = 1.21) were associated with greater risks of complications. CONCLUSIONS: Poor acute phase perioperative glycemic control is associated with poor outcome, but differently in diabetic and non-diabetic patients suggesting different glycemic management strategies for the two patient groups.

Explainable machine-learning predictions for the prevention of hypoxaemia during surgery.

Although anaesthesiologists strive to avoid hypoxemia during surgery, reliably predicting future intraoperative hypoxemia is not currently possible. Here, we report the development and testing of a machine-learning-based system that, in real time during general anaesthesia, predicts the risk of hypoxemia and provides explanations of the risk factors. The system, which was trained on minute-by-minute data from the electronic medical records of over fifty thousand surgeries, improved the performance of anaesthesiologists when providing interpretable hypoxemia risks and contributing factors. The explanations for the predictions are broadly consistent with the literature and with prior knowledge from anaesthesiologists. Our results suggest that if anaesthesiologists currently anticipate 15% of hypoxemia events, with this system's assistance they would anticipate 30% of them, a large portion of which may benefit from early intervention because they are associated with modifiable factors. The system can help improve the clinical understanding of hypoxemia risk during anaesthesia care by providing general insights into the exact changes in risk induced by certain patient or procedure characteristics.

Decisional practices and patterns of intraoperative glucose management in an academic medical center.

OBJECTIVE: To understand the decisional practices of anesthesia providers in managing intraoperative glucose levels. DESIGN: This is a retrospective cohort study. SETTING: Operating rooms in an academic medical center. PATIENTS: Adult patients undergoing surgery. INTERVENTION: Intraoperative blood glucose management based on an institutional protocol. MEASUREMENTS: Glucose management data was extracted from electronic medical records to determine compliance to institutional glucose management protocol that prescribes hourly glucose measurements and insulin doses to maintain glucose levels between 100 to 140mg/dL. Effect of patient and surgery specific factors on compliance to glucose management protocol was explored. MAIN RESULTS: In 1903 adult patients compliances to hourly glucose measurements was 72.5% and correct insulin adjustments was 12.4%. Insulin was under-dosed compared to the prescribed value by a mean of 0.85U/h (95% CI 0.76-0.95). Multivariate analysis showed that compliance to hourly glucose measurements decreased with increasing length of the procedure (OR=0.92 per hour, 95% CI 0.89-0.95) but increased with ASA status codes (OR=1.25 per ASA unit, 95% CI=1.06-1.49). Greater compliance to correct insulin adjustment was found in diabetic patients compared with non-diabetic patients (OR=1.31, 95% CI 1.09-1.55). On average, providers administered progressively more insulin with an additional 0.11U/h (95% CI=0.00-0.21] for every additional 10kg/m(2) of BMI and 0.20U/h (95% CI=0.01-0.39) less in diabetic patients than in non-diabetic patients. With the above practice pattern, the mean±SD of glucose level was 158±36mg/dL. Hypoglycemic (<60mg/dL) incident rate was 0.1% (9/8301 measurements) while hyperglycemic (>180mg/dL) incident rate was 28%. Glucose levels were within the target range (100-140mg/dL) only 28% of the time. CONCLUSIONS: Low compliance and considerable variability in initiating and following institutional glucose management protocol were observed.

The Effect of Intraoperative Blood Glucose Management on Postoperative Blood Glucose Levels in Noncardiac Surgery Patients.

BACKGROUND: Postoperative hyperglycemia has been associated with poor surgical outcome. The effect of intraoperative glucose management on postoperative glucose levels and the optimal glycemic threshold for initiating insulin are currently unknown. METHODS: We performed a retrospective cohort study of surgery patients who required intraoperative glucose management with data extracted from electronic medical records. In patients who required glucose management, intraoperative glucose levels and insulin therapy were compared against postoperative glucose levels during 3 periods: first postoperative level within 1 hour, within the first 12 hours, and 24 hours of the postoperative period. Logistic regression models that adjusted for patient and surgical factors were used to determine the association between intraoperative glucose management and postoperative glucose levels. RESULTS: In 2440 patients who required intraoperative glucose management, an increase in mean intraoperative glucose level by 10 mg/dL was associated with an increase in postoperative glucose levels by 4.7 mg/dL (confidence interval [CI], 4.1-5.3; P < 0.001) for the first postoperative glucose measurement, 2.6 mg/dL (CI, 2.1-3.1; P < 0.001) for the mean first 12-hour postoperative glucose, and 2.4 mg/dL (CI, 2.0-2.9; P < 0.001) for the mean first 24-hour postoperative glucose levels (univariate analysis). Multivariate analysis showed that these effects depended on (interacted with) body mass index and diabetes status of the patient. Both diabetes status (regression coefficient = 12.2; P < 0.001) and intraoperative steroid use (regression coefficient = 10.2; P < 0.001) had a positive effect on elevated postoperative glucose levels. Intraoperative hyperglycemia (>180 mg/dL) was associated with postoperative hyperglycemia during the first 12 hours and the first 24 hours. However, interaction with procedure duration meant that this association was stronger for shorter surgeries. When compared with starting insulin for an intraoperative glucose threshold of 140 mg/dL thus avoiding hyperglycemia, initiation of insulin for a hyperglycemia threshold of 180 mg/dL was associated with an increase in postoperative glucose level (7 mg/dL; P < 0.001) and postoperative hyperglycemia incidence (odds ratio = 1.53; P = 0.01). CONCLUSIONS: A higher intraoperative glucose level is associated with a higher postoperative glucose level. Intraoperative hyperglycemia increases the odds for postoperative hyperglycemia. Adequate intraoperative glucose management by initiating insulin infusion when glucose level exceeds 140 mg/dL to prevent hyperglycemia is associated with lower postoperative glucose levels and fewer incidences of postoperative hyperglycemia. However, patient- and procedure-specific variable interactions make the relationship between intraoperative and postoperative glucose levels complicated.

Intraoperative blood glucose management: impact of a real-time decision support system on adherence to institutional protocol.

Poor perioperative glycemic management can lead to negative surgical outcome. Improved compliance to glucose control protocol could lead to better glucose management. An Anesthesia Information Management System based decision support system-Smart Anesthesia Manager™ (SAM) was used to generate real-time reminders to the anesthesia providers to closely adhere to our institutional glucose management protocol. Compliance to hourly glucose measurements and correct insulin dose adjustments was compared for the baseline period (12 months) without SAM and the intervention period (12 months) with SAM decision support. Additionally, glucose management parameters were compared for the baseline and intervention periods. A total of 1587 cases during baseline and 1997 cases during intervention met the criteria for glucose management (diabetic patients or non-diabetic patients with glucose level >140 mg/dL). Among the intervention cases anesthesia providers chose to use SAM reminders 48.7 % of the time primarily for patients who had diabetes, higher HbA1C or body mass index, while disabling the system for the remaining cases. Compliance to hourly glucose measurement and correct insulin doses increased significantly during the intervention period when compared with the baseline (from 52.6 to 71.2 % and from 13.5 to 24.4 %, respectively). In spite of improved compliance to institutional protocol, the mean glucose levels and other glycemic management parameters did not show significant improvement with SAM reminders. Real-time electronic reminders improved intraoperative compliance to institutional glucose management protocol though glycemic parameters did not improve even when there was greater compliance to the protocol.

A novel computerized fading memory algorithm for glycemic control in postoperative surgical patients.

BACKGROUND: Hyperglycemia is commonly encountered in critically ill patients and is associated with increased mortality and morbidity. To better control blood glucose levels, we previously developed a new computerized fading memory (FM) algorithm. In this study we evaluated the safety and efficacy of this algorithm in surgical intensive care unit (SICU) patients and compared its performance against the existing insulin-infusion algorithm (named VA algorithm) used in our institution. METHODS: A computer program was developed to run the FM and VA algorithms. Forty eight patients, who were scheduled to have elective surgery, were randomly assigned to receive insulin infusion on the basis of either the FM or VA algorithm. On SICU admission, an insulin infusion was either continued from the operating room or initiated when the glucose level exceeded the target level of 140 mg/dL. Hourly blood glucose measurements were performed and entered into the computer program, which then prescribed the next insulin dose. The randomly assigned algorithm was applied for the first 8 hours of SICU stay, after which the VA algorithm was used. The number of episodes of hypoglycemia (glucose <60 mg/dL) and excessive hyperglycemia (>300 mg/dL) were noted. Additionally, the time required to bring the glucose level within target range (140 ± 20 mg/dL), the number of glucose measurements within the target range, glycemic variability, and insulin usage were analyzed and compared between the 2 algorithms. RESULTS: Patient demographics and starting glucose levels were similar between the groups. With the existing VA algorithm, 1 episode of severe hypoglycemia was observed. Three patients did not reach the target range within 8 hours. With the FM algorithm no hypoglycemia occurred, and all patients achieved the target range within 8 hours. Glycemic variability measured by the SD of mean glucose levels was 28% (95% confidence interval, 14% to 39%) lower for the FM algorithm (P < 0.001). The FM algorithm used 1.1 U/h less insulin than did the VA algorithm (P = 0.043). CONCLUSION: The novel computerized FM algorithm for glycemic control, which emulates physiologic biphasic insulin secretion, managed glucose better than the existing algorithm without any episodes of hypoglycemia. The FM algorithm had less glycemic variability and used less insulin when compared to the conventional clinical algorithm.

The problem of artifacts in patient monitor data during surgery: a clinical and methodological review.

Artifacts are a significant problem affecting the accurate display of information during surgery. They are also a source of false alarms. A secondary problem is the inadvertent recording of artifactual and inaccurate information in automated record keeping systems. Though most of the currently available patient monitors use techniques to minimize the effect of artifacts, their success is limited. We reviewed the problem of artifacts affecting patient monitor data during surgical cases. Methods adopted by currently marketed patient monitors to eliminate and minimize artifacts due to technical and environmental factors are reviewed and discussed. Also discussed are promising artifact detection and correction methods that are being investigated. These might be used to detect and eliminate artifacts with improved accuracy and specificity.

Monitoring cerebral oxygenation in a pediatric patient undergoing surgery for vascular ring.

Regional cerebral oxygenation can be monitored using near-infrared spectroscopy (NIRS). Inadequacy of collateral cerebral circulation and regional cerebral ischemia during cardiac and vascular surgery may be detected by the use of NIRS monitoring. We report a 2-year-old child who underwent surgical repair of vascular ring and subclavian reimplantation, where use of NIRS helped in early detection and timely intervention to prevent prolonged cerebral ischemia.