Stroke Outcome Prediction

Jeremy Voisey, Terry Quinn, Keith Goatman

The aim is to predict 30 and/or 90 day level of disability, as measured by the modified Rankin Score, following an ischaemic stroke, with some form of confidence estimate (e.g. a probability distribution). Rather than just a single model, the intention is to have a suite of models that will use the data available at different points in the care pathway. Previous studies have generally predicted a binary outcome (e.g. good vs. poor outcome rather than the actual modified Rankin Score). In addition, these models usually require specific features all be available in order to make a prediction. We aim to develop a more flexible software-based approach that is (a) more specific in its predictions, (b) less constrained on availability of features, and (c) allows users to see an indication of confidence in the prediction.

Existing stroke outcome models, while being reasonably accurate are not in everyday clinical use. This is due to the fact that they require specific features, that are not readily available. In addition, the results are not sufficiently reliable to be of practical use and to justify the time taken to calculate the predictions. 1. Given sufficient input features and training data, a machine learning algorithm can achieve significantly better accuracy than state-of-the-art models in the literature[1]. 2. The accuracy of current models is insufficient for routine clinical decision support[2]. However, we propose that given a measure of confidence in the prediction, those predictions with a high degree of confidence could be suitable for clinical decision making.

Data A suite of models for predicting the modified Rankin Score has previously been trained using data from three studies carried out in Glasgow (ATTEST, POSH and WYETH). However, there were insufficient subjects available from these studies to train and validate a strong model. We would therefore like to request two sets of subjects from the VISTA archive: 1. The same cohort used in Quinn TJ et al. Validating and comparing stroke prognosis scales[3]. This data will only be used for the final validation of the models. 2. Data from subjects with similar features/measurements to those in (1) above, but which were not included in that cohort (and who were also not from the ATTEST, POSH or WYETH studies). This data will be used to refine the existing model, prior to final evaluation. To this end, we would like to request as many eligible datasets as are available. Analysis Statistical models to be considered include logistic regression, proportional odds, partial proportional odds, random forest and boosted decision trees. Different models will be trained on data available at different points in the care pathway. To ensure robustness, statistical methods will be used to allow predictions to be made in the event of missing data. The models will be integrated into a software package that will ensure ease of use. The software will determine which model or weighted ensemble is most suitable. The models will be externally validated using the same cohort as Quinn et al. in “Validating and comparing stroke prognosis scales”[3].

-

Funding is in place for the project from Canon Medical Research Europe Ltd.

References and Appendices [1] G. Ntaios, M. Faouzi, J. Ferrari, W. Lang, K. Vemmos, and P. Michel, “An integer-based score to predict functional outcome in acute ischemic stroke: the ASTRAL score.,” Neurology, vol. 78, no. 24, pp. 1916–22, Jun. 2012. [2] C. Counsell and M. Dennis, “Systematic Review of Prognostic Models in Patients with Acute Stroke,” Cerebrovasc. Dis., vol. 12, no. 3, pp. 159–170, 2001. [3] T. J. Quinn, S. Singh, K. R. Lees, P. M. Bath, and P. K. Myint, “Validating and comparing stroke prognosis scales,” Neurology, vol. 89, no. 10, pp. 997–1002, 2017.