Abstract: BACKGROUND: Most patients with post-stroke aphasia still have basic communication deficits 1 year after onset, highlighting an urgent need for accurate prognostic prediction tools to guide clinical rehabilitation decisions.
OBJECTIVE: To construct a machine learning-based model for predicting language function prognosis at discharge in patients with post-stroke aphasia, aiming to improve prediction accuracy.
METHODS: Clinical data were collected from 245 patients with post-stroke aphasia admitted to the Rehabilitation Hospital Affiliated to Fujian University of Traditional Chinese Medicine from July 1, 2022 to July 1, 2025, with an aphasia quotient change ≥ 6 points as the outcome indicator. The study cohort was randomly divided into training (n=171) and test (n=74) sets in a 7:3 ratio. Predictive factors were screened using recursive feature elimination. Six machine learning algorithms (logistic regression, random forest, decision tree, support vector machine, Gaussian naïve Bayes, and extreme gradient boosting classifier) were used to construct models. Internal validation was performed using the bootstrap method. Model performance was evaluated using receiver operating characteristic curves, calibration curves, and Shapley additive explanations analysis.
RESULTS AND CONCLUSION: The language function improvement rate among the 245 patients with post-stroke aphasia was 69.80%. Ten factors, including age, female sex, education level, anomic aphasia, non-fluent aphasia, global aphasia, and baseline total score of the Chinese functional communication profile, were selected as predictive factors. The Gaussian naïve Bayes model performed best in the test set, with an area under the curve of 0.71 and an F1 score of 0.83. The calibration curve showed good consistency between predicted probabilities and actual outcomes (Brier score=0.19). Shapley additive explanations analysis identified anomic aphasia (0.24), advanced age (0.17), female sex (0.08), baseline total score of the Chinese Functional Communication Profile (0.08), and non-fluent aphasia (0.07) as key risk factors. These findings indicate that the Gaussian naïve Bayes prediction model based on machine learning can effectively identify language functional prognosis risks at discharge in patients with post-stroke aphasia, providing decision support for individualized rehabilitation interventions.

Key words: post-stroke aphasia, machine learning, prognostic prediction, Gaussian na?ve Bayes, Chinese Functional Communication Profile