Multiobjective optimization challenges in perioperative anesthesia: A review.

Physicians use perioperative decision-support tools to mitigate risks and maximize benefits to achieve the most successful outcome for patients. Contemporary risk-assessment practices augment surgeons' judgement and experience with decision-support algorithms driven by big data and machine learning. These algorithms accurately assess risk for a wide range of postoperative complications by parsing large datasets and performing complex calculations that would be cumbersome for busy clinicians. Even with these advancements, large gaps in perioperative risk assessment remain; decision-support algorithms often cannot account for risk-reduction therapies applied during a patient's perioperative course and do not quantify tradeoffs between competing goals of care (eg, balancing postoperative pain control with the risk of respiratory depression or balancing intraoperative volume resuscitation with the risk for complications from pulmonary edema). Multiobjective optimization solutions have been applied to similar problems successfully but have not yet been applied to perioperative decision support. Given the large volume of data available via electronic medical records, including intraoperative data, it is now feasible to successfully apply multiobjective optimization in perioperative care. Clinical application of multiobjective optimization would require semiautomated pipelines for analytics and reporting model outputs and a careful development and validation process. Under these circumstances, multiobjective optimization has the potential to support personalized, patient-centered, shared decision-making with precision and balance.

Markov chain evaluation of acute postoperative pain transition states.

Previous investigations on acute postoperative pain dynamicity have focused on daily pain assessments, and so were unable to examine intraday variations in acute pain intensity. We analyzed 476,108 postoperative acute pain intensity ratings, which were clinically documented on postoperative days 1 to 7 from 8346 surgical patients using Markov chain modeling to describe how patients are likely to transition from one pain state to another in a probabilistic fashion. The Markov chain was found to be irreducible and positive recurrent, with no absorbing states. Transition probabilities ranged from 0.0031, for the transition from state 10 to state 1, to 0.69 for the transition from state 0 to state 0. The greatest density of transitions was noted in the diagonal region of the transition matrix, suggesting that patients were generally most likely to transition to the same pain state as their current state. There were also slightly increased probability densities in transitioning to a state of asleep or 0 from the current state. An examination of the number of steps required to traverse from a particular first pain score to a target state suggested that overall, fewer steps were required to reach a state of 0 (range 6.1-8.8 steps) or asleep (range 9.1-11) than were required to reach a mild pain intensity state. Our results suggest that using Markov chains is a feasible method for describing probabilistic postoperative pain trajectories, pointing toward the possibility of using Markov decision processes to model sequential interactions between pain intensity ratings, and postoperative analgesic interventions.

Teaching a Machine to Feel Postoperative Pain: Combining High-Dimensional Clinical Data with Machine Learning Algorithms to Forecast Acute Postoperative Pain.

BACKGROUND: Given their ability to process highly dimensional datasets with hundreds of variables, machine learning algorithms may offer one solution to the vexing challenge of predicting postoperative pain. METHODS: Here, we report on the application of machine learning algorithms to predict postoperative pain outcomes in a retrospective cohort of 8,071 surgical patients using 796 clinical variables. Five algorithms were compared in terms of their ability to forecast moderate to severe postoperative pain: Least Absolute Shrinkage and Selection Operator (LASSO), gradient-boosted decision tree, support vector machine, neural network, and k-nearest neighbor (k-NN), with logistic regression included for baseline comparison. RESULTS: In forecasting moderate to severe postoperative pain for postoperative day (POD) 1, the LASSO algorithm, using all 796 variables, had the highest accuracy with an area under the receiver-operating curve (ROC) of 0.704. Next, the gradient-boosted decision tree had an ROC of 0.665 and the k-NN algorithm had an ROC of 0.643. For POD 3, the LASSO algorithm, using all variables, again had the highest accuracy, with an ROC of 0.727. Logistic regression had a lower ROC of 0.5 for predicting pain outcomes on POD 1 and 3. CONCLUSIONS: Machine learning algorithms, when combined with complex and heterogeneous data from electronic medical record systems, can forecast acute postoperative pain outcomes with accuracies similar to methods that rely only on variables specifically collected for pain outcome prediction.

Of rough starts and smooth finishes: correlations between post-anesthesia care unit and postoperative days 1-5 pain scores.

OBJECTIVE: The goal of this project was to explore the association between post-anesthesia care unit (PACU) pain scores recorded within the first and second hour of the end of surgery with maximum and median pain scores recorded on postoperative days (PODs) 1 through 5. DESIGN: This study was a retrospective cohort study of clinically documented pain scores in a mixed surgical population. SETTING: This study was set in a single tertiary-care teaching hospital over a 1-year time period. PATIENTS: All patients were adult patients undergoing a single, non-ambulatory, non-obstetric surgical procedure. MEASURES: Pain scores, measured using the numerical rating scale, from PODs 0 through 5 were obtained from an integrated data repository. Kendall's Tau-b correlations were then calculated between maximum pain scores occurring within each of the two PACU time periods and maximum and median pain scores in each of the five ensuing PODs. RESULTS: A total of 349,797 pain scores from 8,332 patients were reviewed. Correlations between maximum pain score by time period demonstrated a significant and high correlation at Tau-b = 0.86, between 1-hour PACU pain scores and 2-hour PACU pain scores. However, the correlation of maximum pain scores recorded in the PACU with those recorded on PODs 1 through 5 was significantly lower, ranging from 0.19 to 0.27. The correlation of maximum PACU pain score with median pain scores recorded on PODs 1 through 5 ranged from 0.22 to 0.29. The correlation structures of the PODs 1 through 5 median pain scores may be consistent with an autoregressive pattern. CONCLUSIONS: Maximum scores measured within the PACU likely reflect a set of circumstances distinct from those experienced on PODs 1 through 5.

Use of machine-learning classifiers to predict requests for preoperative acute pain service consultation.

OBJECTIVE: The purpose of this project was to determine whether machine-learning classifiers could predict which patients would require a preoperative acute pain service (APS) consultation. DESIGN: Retrospective cohort. SETTING: University teaching hospital. SUBJECTS: The records of 9,860 surgical patients posted between January 1 and June 30, 2010 were reviewed. OUTCOME MEASURES: Request for APS consultation. A cohort of machine-learning classifiers was compared according to its ability or inability to classify surgical cases as requiring a request for a preoperative APS consultation. Classifiers were then optimized utilizing ensemble techniques. Computational efficiency was measured with the central processing unit processing times required for model training. Classifiers were tested using the full feature set, as well as the reduced feature set that was optimized using a merit-based dimensional reduction strategy. RESULTS: Machine-learning classifiers correctly predicted preoperative requests for APS consultations in 92.3% (95% confidence intervals [CI], 91.8-92.8) of all surgical cases. Bayesian methods yielded the highest area under the receiver operating curve (0.87, 95% CI 0.84-0.89) and lowest training times (0.0018 seconds, 95% CI, 0.0017-0.0019 for the NaiveBayesUpdateable algorithm). An ensemble of high-performing machine-learning classifiers did not yield a higher area under the receiver operating curve than its component classifiers. Dimensional reduction decreased the computational requirements for multiple classifiers, but did not adversely affect classification performance. CONCLUSIONS: Using historical data, machine-learning classifiers can predict which surgical cases should prompt a preoperative request for an APS consultation. Dimensional reduction improved computational efficiency and preserved predictive performance.