Big Data/Machine Learning/AI
Machine learning to predict long-term post-stroke cognitive decline in a Brazilian stroke cohort (EMMA study) Carine Savalli* Carine Savalli Alessandra Baccaro Maria C Escalante-Rojas Marianna G H S J Leite Isabela M Benseñor Paulo A Lotufo Yuan-Pang Wang Alessandra C Goulart Alexandre Chiavegatto Filho
Following a stroke, cognitive decline is a common long-term disability that affects patients. To advise on preventive and intervention strategies, it is necessary to assess the risk of progressive cognitive decline in patients during the initial stages. This study aimed to evaluate the cognitive status of stroke patients from a Brazilian community-based cohort (EMMA study) in the sub-acute phase of stroke (1–3 months), in a follow-up assessment at 6 months (n=100), 1-year (n=103), and 2-years (n=57). The Modified Telephone Interview for Cognitive Status (TICS-M), which has been validated in Brazil, was used to assess cognitive status. Two outcomes were analyzed for each follow-up: an absolute decrease in the TICS-M score or a decrease of at least 10% from baseline. The machine learning algorithm XGBoost was trained using various variables collected at baseline, including demographics, comorbidities, clinical and laboratory variables. After hyperparameter tuning, we evaluated the predictive performance of each model in a repeated 5-fold cross-validation, using the area under the ROC curve (AUC) as the primary metric. The highest level of performance was achieved in predicting a decrease of at least 10% in TICS-M score within one year after the stroke. When considering the outcome of any absolute decrease in the TICS-M score, we found AUC values of 0.60 (SD=0.09), 0.53 (SD=0.12), and 0.40 (SD=0.21), for the 6-month, 1-year, and 2-years follow-up periods, respectively. When considering the decrease of at least 10% in TICS-M, we found an AUC of 0.59 (SD=0.13), 0.66 (SD=0.12), and 0.53 (SD=0.19) for the 6-month, 1-year, and 2-years follow-up, respectively. According to this post-stroke cognitive decline model based only on routine clinical variables, the TICS-M tool presented a low to moderate performance for screening any change in cognition status. Future studies should incorporate other relevant variables and train other machine learning algorithms.