Artificial Intelligence to Predict Functional Status after Acute Stroke Symptoms From Wrist Accelerometry Devices

BackgroundFunctional outcomes after stroke are commonly assessed via modified Rankin Scale (mRS). However, mRS is subject to patient and assessor biases and is impractical to collect in many cases, limiting its impact on post-stroke care. Artificial intelligence (AI) applied to wrist-worn triaxial accelerometry (WWTA) device data can objectively characterize post-stroke functional status and related changes. MethodsWe used patient data from REACH Stroke-Sleep, a study investigating WWTA-derived measures of sleep, physical activity, and recurrent stroke risk among patients with acute stroke symptoms. We determined moving accelerometry averages and vector sums over four time windows (minute, hour, day, week). We trained a tree-based (random forest; RF) and deep learning (LSTM) model to predict individual 6-month mRS scores and differences between 1- and 6-month mRS scores. We used 5-fold cross validation, modeled each outcome as binary exact-match between actual-predicted values, and determined area under the receiver-operating curve (AUROC), sensitivity, precision, negative predictive value, and F1 scores for both models. For mRS score differences, we determined mean absolute error (MAE) and standard deviation (SD). ResultsWe identified 362 patients in REACH Stroke-Sleep, of whom 302 (83.4%) had a 1-month mRS score, 251 (69.3%) had a 6-month mRS score, and 191 (52.8%) had both. Patients wore devices for median 41.0 (IQR 34.4-44.0) days. For all outcomes, RF models (6-month AUROC 0.81, 95%CI 0.74- 0.89; 1-6 month mRS AUROC 0.82, 95%CI 0.76-0.90) outperformed LSTM models (6-month AUROC 0.63, 95%CI 0.55-0.71; 1-6 month mRS AUROC 0.53, 95%CI 0.45-0.61). RF models (MAE 0.37, SD 0.12) outperformed LSTM (MAE 0.87, SD 0.48) for predicting 1-6 month mRS difference, modeled as a non-binarized outcome. ConclusionsWe found that AI predicted short-term mRS and mRS changes after acute stroke symptoms from WWTA data with moderate performance. Future studies are warranted to investigate whether multimodal data can improve performance with the goal of developing objective, automatable functional status assessments.