Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Precision ICU Resource Planning: A Multimodal Model for Brain Surgery Outcomes

Maximilian Fischer, Florian M. Hauptmann, Robin Peretzke, Paul Naser, Peter Neher, Jan-Oliver Neumann, Klaus Maier-Hein

TL;DR

This work tackles ICU admission prediction after brain surgery, a high-cost decision with imbalanced data, by introducing a multimodal framework that fuses pre/post-operative clinical data with imaging features extracted from T1 MRI. The approach combines traditional tabular models (XGBoost) and deep learning, with dynamic fusion via Dynamic Affine Feature Map Transform (DAFT) and brain foundation models to better integrate modalities. Results show that naive fusion of imaging latents with tabular data often harms performance, while DAFT using 3D SSL latent representations achieves the strongest performance (F1 ≈ 0.41, ROC-AUC ≈ 0.76), particularly when incorporating both pre- and post-operative data. The findings demonstrate the value of multimodal fusion and pretrained image features for predicting ICU needs in imbalanced neurosurgical cohorts, informing future research toward clinically actionable ICU resource planning.

Abstract

Although advances in brain surgery techniques have led to fewer postoperative complications requiring Intensive Care Unit (ICU) monitoring, the routine transfer of patients to the ICU remains the clinical standard, despite its high cost. Predictive Gradient Boosted Trees based on clinical data have attempted to optimize ICU admission by identifying key risk factors pre-operatively; however, these approaches overlook valuable imaging data that could enhance prediction accuracy. In this work, we show that multimodal approaches that combine clinical data with imaging data outperform the current clinical data only baseline from 0.29 [F1] to 0.30 [F1], when only pre-operative clinical data is used and from 0.37 [F1] to 0.41 [F1], for pre- and post-operative data. This study demonstrates that effective ICU admission prediction benefits from multimodal data fusion, especially in contexts of severe class imbalance.

Precision ICU Resource Planning: A Multimodal Model for Brain Surgery Outcomes

TL;DR

This work tackles ICU admission prediction after brain surgery, a high-cost decision with imbalanced data, by introducing a multimodal framework that fuses pre/post-operative clinical data with imaging features extracted from T1 MRI. The approach combines traditional tabular models (XGBoost) and deep learning, with dynamic fusion via Dynamic Affine Feature Map Transform (DAFT) and brain foundation models to better integrate modalities. Results show that naive fusion of imaging latents with tabular data often harms performance, while DAFT using 3D SSL latent representations achieves the strongest performance (F1 ≈ 0.41, ROC-AUC ≈ 0.76), particularly when incorporating both pre- and post-operative data. The findings demonstrate the value of multimodal fusion and pretrained image features for predicting ICU needs in imbalanced neurosurgical cohorts, informing future research toward clinically actionable ICU resource planning.

Abstract

Although advances in brain surgery techniques have led to fewer postoperative complications requiring Intensive Care Unit (ICU) monitoring, the routine transfer of patients to the ICU remains the clinical standard, despite its high cost. Predictive Gradient Boosted Trees based on clinical data have attempted to optimize ICU admission by identifying key risk factors pre-operatively; however, these approaches overlook valuable imaging data that could enhance prediction accuracy. In this work, we show that multimodal approaches that combine clinical data with imaging data outperform the current clinical data only baseline from 0.29 [F1] to 0.30 [F1], when only pre-operative clinical data is used and from 0.37 [F1] to 0.41 [F1], for pre- and post-operative data. This study demonstrates that effective ICU admission prediction benefits from multimodal data fusion, especially in contexts of severe class imbalance.

Paper Structure

This paper contains 6 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Dataset
  3. Materials and methods
  4. Experiments
  5. Results
  6. Discussion

Figures (2)

  • Figure 1: Left: Brain slice of a random subject of the cohort. Right: Overview of six tabular feature distributions.
  • Figure 2: Left: F1-scores of different scenarios. The bar in the lighter color and dashed lines present the performance when only pre-operative data is used. Right: ROC-AUC curve of the GBT-baseline and our best performing method for only pre- and for pre- and post-operative data.