Deep Learning for Dynamic Prognostic Prediction in Minimally Invasive Surgery for Intracerebral Hemorrhage: Model...

Area of Science:

• Neurology and Neurosurgery
• Artificial Intelligence in Medicine
• Medical Data Science

Background:

• Intracerebral hemorrhage (ICH) prognosis after minimally invasive surgery (MIS) is dynamic and challenging to predict with traditional models.
• Existing models struggle with multi-time point, imbalanced, and missing clinical data.
• There is a need for dynamic prognostic models that can adapt to evolving patient states.

Purpose of the Study:

• To develop and validate a dynamic prognostic model for ICH patients undergoing MIS.
• To predict short-term and long-term survival and functional outcomes using multi-time point data.
• To address limitations of existing models in handling complex clinical data.

Main Methods:

• Retrospective collection of data from 287 ICH patients undergoing MIS at multiple time points.
• Development of a MultiStep Transformer model for simultaneous prediction of 30-day and 180-day survival, and 180-day functional outcomes (mRS 0-3).
• Evaluation using five-fold cross-validation, ROC curves, calibration curves, DCA, and attributable value analysis.

Main Results:

• The MultiStep Transformer model demonstrated superior predictive efficacy compared to traditional scores and other deep learning models.
• Achieved high AUROCs for 30-day survival (0.87), 180-day survival (0.85), and 180-day functional outcome (0.75).
• Decision curve analysis confirmed significant clinical utility across various threshold probabilities.

Conclusions:

• The MultiStep Transformer model effectively utilizes imbalanced data for dynamic prognosis prediction in ICH patients post-MIS.
• This model offers a novel, powerful tool for individualized prognosis assessment, improving patient care.
• The study highlights the potential of advanced AI models in neurosurgical outcome prediction.