Reducing post-operative infections : the development and validation of artificial intelligence-predictive model to inform shared decision making

Abstract

Background: Healthcare systems worldwide generate sizeable patient-related health data. There is growing interest amongst clinicians and healthcare staff in how this can be used to support patient care. One example is the use of predictive analytics in determining the risk of developing a particular complication, such as an infection post-surgery. Objective: To develop an artificial intelligence (AI) model to predict the likelihood of post-operative infection in surgical patients, while also exploring clinicians’ and patients’ perceptions on using AI decision support tools more broadly to inform shared decision making. Methods: This PhD programme of work involved a number of different stages. The literature was systematically reviewed for AI models that could inform clinical decision making with regard to post-surgical infection, and a candidate list of positive predictor variables extracted. A prognostic AI-model was developed to predict the risk of infection, and any inherent biases identified. Another systematic review was conducted to understand how clinicians and patients perceive using AI decision aids in shared decision making, and semi-structured interviews were carried out with clinicians to explore how to improve the clinical utility of AI decision support tools. Results: Nine steps were identified for developing AI-predictive models; the first six steps were applied in the development and evaluation of our model. Nineteen predictors were used. The ensemble model displayed high performance in training (sensitivity: 85.3%, specificity: 74.6%, AUC: 88.6%) and internal validation (sensitivity: 96.9%, specificity: 74.1%, AUC: 85.5%). Patients and clinicians raised concerns about AI model interfaces, in general, and their impact on clinical/patient conversations. Several suggestions were made on how to improve the model’s clinical application. Conclusion: This study provided a deeper understanding of how AI-predictive models can guide shared decision making. Future work should concentrate on improving the user inclusivity of these tools and reducing the risk of inherent bias that could potentially mislead clinical decision making.