An Integrated Multi-omics prediction model for stroke recurrence based on L_net transformer layer and dynamic...

Area of Science:

• Neurology
• Biomedical Informatics
• Data Science

Background:

• Stroke recurrence poses a significant threat, increasing mortality and disability.
• Existing research faces challenges in multi-omics data integration and noise reduction from MRI.
• A reliable multi-omics dataset for stroke recurrence prediction is lacking.

Purpose of the Study:

• To develop a high-performance multi-omics prediction model for stroke recurrence.
• To address challenges in feature extraction from MRI and integrate diverse omics data.
• To create a robust dataset for stroke recurrence prediction.

Main Methods:

• Compiled MRI and clinical data from 737 stroke patients into the PSTSZC dataset.
• Introduced the Integrated Multi-omics Prediction Model for Stroke Recurrence (MPSR), incorporating ResNet, Lnet-transformer, LSTM, and dynamically weighted DNN.
• Developed a novel Lnet regularization layer for MRI noise reduction and a dynamic weighting mechanism for omics data integration.

Main Results:

• The MPSR model outperformed seven single-omics and six state-of-the-art multi-omics models.
• MPSR achieved superior performance with accuracy, AUROC, specificity, and sensitivity of 0.96, 0.97, 1, and 0.94, respectively.
• The model effectively reduced noise from MRI data and integrated multi-omics information.

Conclusions:

• MPSR is the first high-performance multi-omics prediction model for stroke recurrence.
• The model demonstrates potential as a clinical tool for diagnosing stroke recurrence predisposition.
• This work provides a robust framework for multi-omics data analysis in predicting disease recurrence.