Big Data/Machine Learning/AI
Importance of Incorporating Imaging Data into Machine Learning Algorithms for Predicting Post-Stroke Cognitive Decline: Results from the EMMA study Vinivius de Camargo* Vinícius de Camargo Carine Savalli Alessandra Baccaro Isabela M Benseñor Paulo A Lotufo Yuan-Pang Wang Alessandra C Goulart Alexandre Chiavegatto Filho
After a stroke, patients may experience long-term cognitive decline. Predicting cognitive decline in the early stages can help delay its progression. This study aimed to predict cognitive decline one year after a stroke, in patients from a Brazilian community-based cohort (EMMA study). To assess the cognitive status at baseline and after one year, we used the Modified Telephone Interview for Cognitive Status (TICS-M), validated in Brazilian Portuguese, and the outcome was defined as a decrease of at least 10% in TICS-M from baseline. The machine learning algorithm XGBoost was trained with variables collected at baseline, including demographics, comorbidities, clinical, laboratory variables. Also, neuroimaging variables captured by 3 tesla magnetic resonance (thickness, area and intracranial volume (ICV) in the right and left hemisphere) 1-3 months after stroke (subacute phase). After hyperparameter tuning, we evaluated the predictive performance of the model in a repeated 5-fold cross-validation, using the area under the receiver operating characteristic curve (AUROC). The contribution of variables to outcome prediction was assessed using SHAP (Shapley Additive exPlanations) values. A total of 66 patients with a neuroimaging evaluation were included in this analysis. Results indicated robust performance in predicting the decrease of at least 10% in TICS-M score after one year (AUC=0.78, SD=0.16). Imaging exam variables were assessed with greater importance through SHAP, with emphasis on ICV and average thickness of the right hemisphere. These variables were more relevant to model discriminate patients and predict cognitive decline after a stroke. According to our results, variables obtained in image exams can be the key to predict the long-term cognitive decline after a stroke. Future studies should incorporate more detailed information about image exams.