Abstract
Introduction
The operational process significantly impacts surgical outcomes; however, the difficulty of capturing intraoperative data and the time cost associated with its analysis have hindered improvements to these processes. Recent advances in artificial intelligence and computer vision have provided an opportunity to address this by automating data capture and analysis. This has the potential to provide unique insights into surgical skill acquisition, operative technical skill assessment, outcome monitoring and credentialing. Despite this potential, most work in the field has yet to influence clinical practice. To achieve this, clinician engagement is needed to quantify the operative process better and understand how the clinical context impacts it. Greater clinician engagement will facilitate the generation of the high-volume structured datasets and the clinically derived targets for automation required to enable iterative operative process improvement needed to realise benefits for patients and clinicians.
Methods
This thesis aims to achieve this by using laparoscopic cholecystectomy as a model operation to:
1. Perform operative video analysis to define the operative process, including the impact that operative technical difficulty and surgeon experience have on this process.
2. Combine synoptic operative note data with video analysis to determine the impact of the operative process on clinical outcomes.
3. Use these insights to facilitate the development of clinically informed computer vision tools for intraoperative analysis.
Results
Operative process analysis confirmed that operative technical difficulty significantly impacts the operative process as quantified using a standard operative process model. The frequency and indications for the artery first technique were quantified, leading to its characterisation as an adjunct in difficult dissection. Analysis of synoptic operation note data revealed that the operative process followed at 2Christchurch hospital contributes to a low rate of bile duct injury (0.04% (2)). Structured synoptic operation note data informed by video analysis can be used to optimise intraoperative decision-making and perioperative management using a process-orientated approach. Having established the validity of process modelling and related it to patient care, these findings were used to evaluate and contextualise the utility of two separate computer vision models. This generated a novel metric, proportion of anatomy visualisation, for operative skill assessment.
Discussion
Operative video analysis coupled with structured process modelling is a powerful tool for generating insights into surgical techniques. Pairing this granular analysis with structured data capture as part of patient care allows for the ready application of these insights in practice. Structured data capture as part of workflows when coupled with real-time outcome monitoring could allow for iterative process improvement. Furthermore, these insights can be used to generate novel clinically applicable targets for computer vision automation, as evidenced here by the capacity of proportional anatomy visualisation to stratify primary operator experience as a proxy for operative technical skill. Computer vision and machine learning have the potential to disrupt surgery significantly. However, to have utility, hospital systems need to facilitate the capture and analysis of structured data as part of the workflow. As the exponential evolution of artificial intelligence continues, embracing this change has the potential to improve outcomes for patients and surgeons significantly.