Strokeformer: A novel deep learning paradigm training transformer-based architecture for stroke prognosis prediction

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Stroke is a leading cause of death and disability globally.
• Accurate stroke prognosis is crucial for determining thrombolytic therapy eligibility.
• Existing deep learning models struggle with small, imbalanced medical datasets, leading to overfitting and performance degradation.

Purpose of the Study:

• To propose a novel deep learning model, Strokeformer, for accurate stroke prognosis prediction.
• To address the challenges of overfitting and performance degeneration in medical datasets.
• To enhance clinical decision-making for thrombolytic therapy.

Main Methods:

• Developed Strokeformer, a novel transformer-based model with intra- and inter-feature interaction modules.
• Employed a pretraining and fine-tuning strategy on large-scale, class-balanced datasets followed by downstream, imbalanced datasets.
• Validated the model on 20 public OpenML datasets and two private clinical stroke prognosis datasets.

Main Results:

• Strokeformer significantly outperformed existing comparison models on stroke prognosis prediction tasks.
• The pretraining and fine-tuning method effectively prevented overfitting on small, imbalanced datasets.
• The model demonstrated promising empirical results on real-world stroke prognosis data.

Conclusions:

• Strokeformer offers a robust solution for stroke prognosis prediction, particularly with challenging medical data.
• The proposed training paradigm is effective in mitigating overfitting in deep learning models for medical applications.
• While interpretability remains a challenge, Strokeformer shows potential for clinical decision support in stroke management.