The Finnish Prehospital Stroke Scale Detects Thrombectomy and Thrombolysis Candidates-A Propensity Score-Matched Study.

BACKGROUND: Prehospital stroke triage is challenged by endovascular treatment for large vessel occlusion (LVO) being available only in major stroke centers. Conjugate eye deviation (CED) is closely related to LVO, whereas common stroke signs (face-arm-leg-speech-visual) screen stroke. We hypothesized that combining CED with common stroke signs would yield a prehospital stroke scale for identifying both LVO and stroke in general. METHODS AND RESULTS: We retrospectively analyzed consecutive patients (n = 856) with prehospital Code Stroke (recanalization candidate). The National Institutes of Health Stroke Scale (NIHSS) and computed tomography were administered to patients on arrival. Computed tomography angiography was performed on patients with NIHSS score of 8 or greater and considered to benefit from endovascular treatment. With random forest analysis and deviance analysis of the general linear model we confirmed the superiority of the NIHSS "Best Gaze" over other NIHSS items in detecting LVO. Based on this and commonly used stroke signs we presented the Finnish Prehospital Stroke Scale (FPSS) including dichotomized face drooping, extremity weakness, speech difficulty, visual disturbance, and CED. FPSS detected LVO with a sensitivity of 54%, specificity of 91%, positive predictive value of 48%, negative predictive value of 93%, and likelihood ratio of 6.2. CONCLUSIONS: Based on CED and universally used stroke signs, FPSS recognizes stroke in general and additionally, LVO as a stroke subtype comparably to other scales intended to detect LVO only. As the FPSS items are dichotomized, it is likely to be easy for emergency medical services to implement.

Extensive phenotype data and machine learning in prediction of mortality in acute coronary syndrome - the MADDEC study.

Objective: Investigation of the clinical potential of extensive phenotype data and machine learning (ML) in the prediction of mortality in acute coronary syndrome (ACS). Methods: The value of ML and extensive clinical data was analyzed in a retrospective registry study of 9066 consecutive ACS patients (January 2007 to October 2017). Main outcome was six-month mortality. Prediction models were developed using two ML methods, logistic regression and extreme gradient boosting (xgboost). The models were fitted in training set of patients treated in 2007-2014 and 2017 (81%, n = 7344) and validated in a separate validation set of patients treated in 2015-2016 with full GRACE score data available for comparison of model accuracy (19%, n = 1722). Results: Overall, six-month mortality was 7.3% (n = 660). Several variables were found to be significantly associated with six-month mortality by both ML methods. The xgboost scored the best performance: AUC 0.890 (0.864-0.916). The AUC values for logistic regression and GRACE score were 0.867(0.837-0.897) and 0.822 (0.785-0.859), respectively. The AUC value of xgboost was better when compared to logistic regression (p = .012) and GRACE score (p < .00001). Conclusions: The use of extensive phenotype data and novel machine learning improves prediction of mortality in ACS over traditional GRACE score. KEY MESSAGES The collection of extensive cardiovascular phenotype data from electronic health records as well as from data recorded by physicians can be used highly effectively in prediction of mortality after acute coronary syndrome. Supervised machine learning methods such as logistic regression and extreme gradient boosting using extensive phenotype data significantly outperform conventional risk assessment by the current golden standard GRACE score. Integration of electronic health records and the use of supervised machine learning methods can be easily applied in a single centre level to model the risk of mortality.