Supervised prediction of post-stroke upper limb motor recovery with uncertain knowledge graph and large language model

 

 Do you not understand, prediction is completely useless for stroke survivors? It does nothing to get them recovered. There are a lot of mentors and senior researchers that need to be re-educated on the purpose of stroke research. The only goal in stroke is 100% recovery; not biomarkers, prediction, prognosis or other useless shit! I'd fire all of you for incompetence!

Supervised prediction of post-stroke upper limb motor recovery with uncertain knowledge graph and large language model

• Tianxing Wu, 
• Xixi Wu, 
• Jian Li, 
• Hangyu Sha, 
• Haofen Wang, 
• Ya Chen, 
• Yali Jiang, 
• Li Liu, 
• Yue Lan & 
• Xiao Lu 

Abstract

Purpose

Accurate prediction of post-stroke upper limb motor recovery is crucial for developing stroke rehabilitation strategies. However, existing methods often focus on static predictions at a fixed post-stroke time point and fail to utilize unstructured textual data within electronic health records (EHRs). To overcome both limitations, we propose a new task which is to dynamically predict the recovery outcomes at variable time points based on the most recent EHR, and aim to address this task by leveraging the text in EHRs with uncertain knowledge graph (UKG) and large language model (LLM).

Methods

The proposed method is developed using EHRs and their corresponding recovery outcomes at different follow-up time points (i.e., 3, 6 and 12 months post-stroke). We first transform EHRs comprising unstructured textual data, categorical data, and numerical data, into vector representations. Textual data enriched by UKG and LLM are embedded using a pretrained language model, and subsequently concatenated with the encoded one-hot representation of categorical data, and normalized numerical data. We then leverage a decision tree based model for feature selection, grouping features into distinct priority levels. We finally perform supervised learning to predict recovery outcomes, with a comparative evaluation of multiple models.

Results

Extensive experiments demonstrate the superiority of the proposed method in different evaluation metrics, achieving 92.5% accuracy in dynamic predictions. After feature selection, the best model (i.e., Gradient Boosting Classifier) achieves a 7.9% increase in accuracy compared to the models using all features.

Conclusion

Our study demonstrates that integrating UKG with LLM enables flexible and reliable post-stroke upper limb motor recovery prediction. By allowing dynamic predictions based on clinical needs and providing clear feature selection guidance, the proposed method has the potential for practical application to support the development of personalized rehabilitation strategies.