Recognizing surgical patterns

Abstract

In the Netherlands, each year over 1700 patients die from preventable surgical errors. Numerous initiatives to improve surgical practice have had some impact, but problems persist. Despite the introduction of checklists and protocols, patient safety in surgery remains a continuing challenge. This is complicated by some surgeons viewing their own work as an artistic manoeuver whose workflow cannot be captured. However, safeguarding patient safety is also a hospital's management responsibility and no longer only in the surgeon's hands. In spite of the inherent variations, surgeries of the same kind produce similar data, and are usually performed in similar workflows. Surgery is characterized by a peri-operative pipeline of pre-, intra- and post-operative processes. To both reduce errors and improve efficiency, the workflow in the peri-operative pipeline should be designed and planned as effectively as possible in terms of flow of patients and allocation of scarce resources such as operating rooms, instruments and personnel. Currently, planning is done on a very basic level, without using real-world data to learn and improve efficiency. Fortunately, there is lot of available, but unexploited data about surgical interventions that can be used for this purpose. The aim of this thesis is to use acquired and registered peri-operative data to support hospital management to improve safety and efficiency in surgery. The method of assessing safety and efficiency in surgery for individual patients needs to be tailored to each patient. As a result generalization of the results is difficult. We discuss how pattern recognition (PR) provides tools for the assessment of surgical outcome for individual patients. It also allows for handling of outliers and does not impose the same restrictions on data collection procedures as for randomized controlled trials. We show that PR is a pragmatic next step towards data intensive operating rooms with evidence based support for surgeries. Below the techniques as proposed in this thesis are brie y described. To support pre-operative planning of surgeries, assessment of surgical complexity is needed beforehand in order to prepare and possibly avoid complications and delays. This complexity assessment can also aid surgeons in decisions regarding how to proceed with the surgical procedure, for instance by taking extra precautions or making a referral to a more experienced surgeon when a complex surgery is predicted. We show how to use readily available patient data to predict surgical complexity. Classifiers are trained and evaluated using readily collected data from patients undergoing laparoscopic cholecystectomy (LAPCHOL). It is shown that complexity of LAPCHOL surgeries can be predicted in the pre-operative stage with accuracy up to 83% using an LDC or SVM classifier. We also derived the set of features that are relevant for predicting complexity including inflammation, wall thickening, sex and BMI score. To realize intra-operative safety and efficiency goals in surgery, hospitals are searching for autonomous systems for monitoring the surgical workflow in the operating room (OR). In this thesis we propose an autonomous registration technique for the OR. Registering the time of use of surgical instruments and the sequence in which they are used enables us to detect the surgical steps, including the duration of each step. By deploying this as a real-time system, dynamic support for the surgical team and dynamic planning of patients can be performed. For monitoring the usage of surgical instruments, signals from sensors which can detect video, motion and RFID tags can be used. For the application in the OR, it is necessary that these sensors are designed to meet the requirements of the OR environment, specifically with respect to sterilization and non-intrusiveness. We propose a tracking system to detect and track instruments in endoscopic video using biocompatible and sterilization-proof colour markers. The system tracks single and multiple instruments in the video. The output of the tracking tool is a log file with an identifier of the instrument used and the duration of its use for each entry. These instrument logs are then used for workflow mining and outlier detection in surgery. We derived a surgical consensus from multiple surgery logs using global multiple sequence alignment. We showed that the derived consensus conforms to the main steps of laparoscopic cholecystectomy as described in best practices. Using global pair-wise alignment, we showed that outliers from this consensus can be detected using the surgical log. These outliers are commonly simple variations in the execution of the surgical procedure, but can also represent serious complications or errors. To improve post-operative efficiency, accurate predictions of patients' length of stay (LOS) in the postanesthesia care unit (PACU) may lead to cost savings and a number of other efficiency benefits. We propose to use available perioperative data to predict the PACU LOS, using the features case demographics, intra-operative parameters, medications, patient co-morbidities, and surgeon. A linear regression method was used along with ordinary least square regression and `least absolute shrinkage and selection operator' (LASSO-) regression. A forward feature selection approach was then used to identify and rank factors that impact PACU LOS. We showed that PACU LOS can be predicted by perioperative factors with an improvement of 12-18 minutes compared to using the mean baseline. If this prediction is updated with online information, mainly by monitoring post-operative oxygen saturation, future work could lead to real-time LOS algorithms based on peri-operative factors to predict, manage and possibly intercept anticipated, prolonged PACU LOS. This thesis has proposed and demonstrated the application of pattern recognition tools to log, assess and predict surgical workflow parameters. Work in this thesis did not directly contribute to reduce errors and safety in the OR. However, the tools developed in the thesis can be used to support standardization of surgical workflow to both reduce errors and support surgical planning. Moreover, the proposed techniques for the operating room can be used in other medical domains such as the intensive care unit with only small contextual modifications.