.png){kind=icon}
.png){kind=icon}
Return
Predicting the prognosis of language impairment outcomes in patients with post-stroke aphasia: a systematic review
.png){kind=icon}
PRE
AI
H
L
J
X
S
B
J
DOI：10.1080/02687038.2026.2649848.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
BackgroundTraditional statistical models often struggle to capture the nonlinear interactions in post-stroke aphasia (PSA) recovery. Machine learning (ML) offers a robust alternative, yet a synthesis of its performance and the efficacy of stage-specific predictors is lacking.MethodsWe systematically searched seven databases through September 2024 for studies developing or validating ML models for PSA prognosis. We assessed methodological quality using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Predictors, algorithms, and performance metrics were synthesized across recovery stages to evaluate clinical utility.ResultsTwelve studies yielding 29 best-performing models were analyzed. Initial severity emerged as the primary predictor across all recovery stages. A distinct shift in predictive focus was observed, where acute-stage models relied on clinical and demographic data, whereas chronic-stage models prioritized neuroimaging markers such as white matter integrity and functional connectivity. In terms of performance, acute-stage models achieved high discriminative power with an AUC of 0.87-0.91, and chronic-stage models reached classification accuracies between 0.67 and 0.93. Traditional algorithms such as Support Vector Machines and Random Forests dominated the field, while deep learning applications remained limited. However, 19 models were rated at high risk of bias in more than two PROBAST domains, with the analysis domain being most affected due to low events-per-variable ratios and inappropriate participant exclusion.ConclusionsMachine learning provides a promising framework for PSA prognosis, supporting early screening and long-term goal setting. However, the high risk of bias and lack of external validation limit current clinical generalizability. Future research should prioritize external validation and expand outcome measures beyond general aphasia quotients to specific domains like naming and comprehension.
Keywords:
Post-stroke aphasia
prognosis prediction
rehabilitation outcome
machine learning
Journal
IF:
2.1
Papers:
253
Citations:
4.4K
Organization
Cited Papers