Deep transformation models for functional outcome prediction after acute ischemic stroke.

Area of Science:

• Medical Informatics
• Machine Learning
• Biostatistics

Background:

• Interpretable models with high prediction performance are crucial for medical applications, especially when handling semi-structured data (tabular and imaging).
• Deep Transformation Models (DTMs) offer a solution by allowing specification of deep neural networks for different input modalities.

Purpose of the Study:

• To apply and compare Deep Transformation Models (DTMs) for distributional regression in predicting functional outcomes in stroke patients.
• To assess the trade-off between interpretability and flexibility in model building using semi-structured data.
• To evaluate the predictive utility of tabular clinical data versus brain imaging data.

Main Methods:

• Trained and compared several DTMs, including baseline-adjusted models, on a dataset of 407 stroke patients.
• Utilized statistical principles for model building to balance interpretability and flexibility.
• Assessed the relative importance of tabular and imaging data modalities for predicting ordinal functional outcomes.

Main Results:

• DTMs provide interpretable effect estimates and achieve state-of-the-art prediction performance.
• Tabular clinical data was more effective for predicting functional outcomes than imaging data alone.
• Combining both data modalities did not substantially improve prediction performance.

Conclusions:

• DTMs offer a powerful and interpretable approach for analyzing semi-structured medical data.
• The study highlights the utility of tabular data over imaging data for predicting stroke outcomes.
• DTMs have the potential to support clinical decision-making by providing reliable and interpretable predictions.