Introduction Endovascular thrombectomy (ET) is currently a standard of care in acute ischemic stroke (AIS). However, data from major trials provided conflicting evidence on the efficacy of ET in the elderly (80 years and older). Different selection criteria used across different trials have resulted in variable functional outcomes, thus questioning the use of ET in the elderly. Here, we applied machine learning algorithms to patient selection for ET to determine the optimal candidates for ET and to maximize the efficacy of ET in the elderly population using retrospective and prospective data.
Methods We retrospectively reviewed the records for AIS patients age 80 and older who underwent ET between January 2013 and July 2017 at the Medical University of South Carolina. Data on patient demographics, ASPECT scores, pre-stroke and admission deficits, comorbidities, and the use of IV-tPA were specifically collected since these variables are readily available at the time of decision for ET. Outcomes were collected at 90 days using the mRS scores. Following data normalization, we used a supervised approach to train 3098 classifiers including parametrized decision tree, association rule, Bayes based, and function estimation based algorithms and identified the best classifier to predict 90 day outcomes. Each algorithm was executed in multiple iterations while altering algorithm parameters to cover all reasonable combinations. We then used the classifier that returned the best results to predict outcomes in prospectively enrolled patients between July and December 2017 for validation of prediction.
Results A total of 110 patients were retrospectively reviewed and used to train the classifiers. Among the 3098 tested classifiers, we identified a custom classifier SPOT (Stroke Prognosis in Octogenarians undergoing thrombectomy) based on an M5P decision tree as the best performing classifier with mean absolute error of 0.83. We then compared the predictive success of SPOT to that of common predictors of outcomes after stroke showing a higher area under the curve for SPOT (0.918) compared to admission NIHSS (0.82), pre-stroke mRS (0.68), or ASPECT (0.58). SPOT showed 88.5% sensitivity and 89.2% specificity in predicting functional independence (mRS 0–2) at 90 days. The overall rates of functional independence among all patients was 20%; however, among those predicted by SPOT to have functional independence, 70% had an mRS 0–2 at 90 days compared to 2.5% in patients predicted by SPOT not to have functional independence at 90 days. We then applied SPOT on a prospective cohort of 13 patients (age 80 or older) undergoing ET at our center. The overall actual rate of functional independence was 53% in the 13 patients, compared to 72% in patients predicted by SPOT to have functional independence, and 16.7% in patients predicted by SPOT not to have functional independence.
Conclusion Using machine learning to rapidly predict outcomes of patients eligible for ET may help optimize selection criteria in populations not well studied in clinical trials, and improve outcomes of intervention. Future validation of SPOT using data from multiple centers may support prospective use of SPOT for patient selection in future trials using ET in the elderly.
Disclosures A. Alawieh: None. F. Zaraket: None. A. Chatterjee: None. A. Spiotta: 1; C; Penumbra, Pulsar Vascular, Microvention, Stryker. 2; C; Penumbra, Pulsar Vascular, Microvention, Stryker.
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.