Personalised, data-driven surgical management of pituitary tumours using artificial intelligence – a pilot study

Background: The gold standard treatment for most patients with symptomatic pituitary tumours is surgical excision via the endoscopic transsphenoidal approach (eTSA).

These operations are technically challenging, with significant heterogeneity in surgical techniques used, which may result in differing surgical outcomes.

Therefore, we generated an international consensus-driven analysis of the operative workflow (phases, steps, instruments, errors) in contemporary pituitary surgery.

This workflow analysis can be used to systematically segment operations, allowing granular assessment of surgical technique to drive improvement in outcomes.

Machine learning (ML), specifically artificial neural networks (ANNs), can be used to analyse workflow automatically and accurately. We have previously published a proof-of-concept for this technology using 50 anonymised eTSA videos.

However, linking operative video workflow analysis to the wider clinical context (e.g. post-operative data) is, to our knowledge, entirely novel. This presents an opportunity to generate a personalised, data-driven approach to each individual’s care.

Aims: We aim to build on our existing ML model for detecting surgical workflow in endoscopic pituitary adenoma resection by integrating post-operative outcome data – to enable outcome prediction based on operative videos, and outcome-driven improvement in surgical technique.

Methods: Through our earlier work, we have established an early project infrastructure (including ethical approval) and built an interdisciplinary study team in the UK’s largest pituitary centre (performing ~ 150 operations/year).

We will refine our existing operative workflow analysis ML model using a larger video dataset and more granular data labelling (e.g. annotating surgical instruments).

This ANN consists of an intra-operative variable input layer (video data) and an output layer predicting the operative steps.

A second ANN will be linked to the output of the first, and will integrate post-operative data from an internal prospective database, to predict post-operative outcomes.

For all networks an iterative approach will be used to optimise network hyperparameters, and evaluation will include: accuracy, F1 score, averaged receiver-operating characteristic analysis, and comparison against logistic regression. We estimate at least 100 cases would be required for an initial exploratory analysis.

How the results of this research will be used: The integration of post-operative data into intra-operative workflow analysis using AI has the potential for significant impact on surgically treated cancers.

This includes data-driven post-operative outcome modelling tailored specifically to each case, and the outcome-driven optimisation of operative technique.

If this pilot study using pituitary surgery as an exemplar is successful, this technology will be progressed through further study (ideally via doctoral fellowship).