Exploring a Prediction Model Based on Patient Record Data in Carotid Stenosis Risk Assessment.

Abstract

Introduction: Digitalised patient records represent a large potential source of real-world data. Nevertheless, confidentiality and data protection has made big data extraction from patient journals impossible in the past. Future options for free text reading by artificial intelligence might enable data extraction while maintaining confidentiality. In turn this could enable improvement in risk predictions. However, it remains unclear whether free-text data constitute an appropriate data source for this purpose, or whether this data source can be incorporated into a clinically applicable risk prediction tool using an appropriate statistical model. We aim to investigate this for stroke risk in patients with carotid stenosis.

Methods: A test dataset of patients with carotid stenosis, based on free text data, was manually established. Parameters of choice were extracted from digital patient records to assess their performance in the prediction of individual risk profile. The dataset is applied to evaluate a proposed statistical model including time dependency of specific variables and multiple endpoint analysis, comparing this to the results from a traditional Cox regression analysis.

Results: 392 symptomatic and 471 asymptomatic patients with carotid stenosis were included in the analysis, where respectively 49 repeat events and 56 first stroke events attributed to the carotid stenosis occurred. The suggested extended regression model shows a statistically significant risk of a high degree carotid stenosis lesion on the risk of an attributable stroke, which was not demonstrated by use of a traditional Cox regression analysis. For a stenosis of 70-99%, asymptomatic patients with a right sided carotid stenosis had a hazard ratio (HR) 4.98 for a right sided event, and for left sided carotid stenosis a HR 5.23 for a left sided event. Symptomatic patients were found to have a HR of 2.25 for a recurrent event, as many received protective treatment. This is suggested to imply improved risk prediction by an advanced statistical model.

Discussion: In this article a description of the data set and the methods behind our model choice has been the main emphasis, in order to evaluate its suitability for this purpose. Analysis of model performance by concordance analysis and Brier score has been published in a separate article.