CSRA: Controlled Spectral Residual Augmentation for Robust Sepsis Prediction
CSRA: Controlled Spectral Residual Augmentation for Robust Sepsis Prediction
Authors: Honglin Guo, Rihao Chang, He Jiao, Weizhi Nie, Zhongheng Zhang, Yuehao Shen Date: 2026-04-16 Paper ID: openalex:2604.14532
Summary
CSRA addresses the difficulty of short-window sepsis prediction by introducing a controlled spectral residual augmentation framework. By grouping clinical variables and applying adaptive spectral perturbations, the method generates structured and clinically plausible trajectory variations. The model is trained end-to-end with an anchor consistency loss and controller regularization, leading to consistent performance gains across various clinical settings, including constrained data availability and challenging forecast horizons.
Key Contributions
- Proposed CSRA, a controlled data augmentation framework that improves sepsis prediction performance in short-window and limited-data clinical settings.
- Introduced input-adaptive residual perturbation in the spectral domain for generating clinically plausible multi-system ICU time-series variations.
- Demonstrated empirical superiority over baselines, reducing regression MSE by 10.2% and MAE by 3.7% on the MIMIC-IV dataset while maintaining effectiveness on the ZiGongICUinfection dataset.
Open Questions & Future Work
Key Concepts
- controlled-spectral-residual-augmentation: A time-series augmentation framework that performs adaptive residual perturbation in the spectral domain, controlled by a unified objective to maintain clinical plausibility.
Archivist Review
The concept of Controlled Spectral Residual Augmentation is approved as a reusable method for domain-aware time-series data augmentation. MIMIC-IV is included as a core, widely-used clinical dataset. The open question regarding irregular time-series sampling is retained as it highlights a specific, significant gap for frequency-domain approaches in clinical applications, while the other open question was rejected as too generic.
Approved Concepts
- Controlled Spectral Residual Augmentation: Introduces a controlled frequency-domain approach to data augmentation that preserves physiological plausibility in time-series forecasting.
Approved Open Questions
- Spectral Augmentation for Irregular Time-Series: Addressing irregular sampling is critical for the practical deployment of clinical time-series models in real-world EHR settings.
Rejected Candidates
- [open_question] Adaptive Variable Dependency Modeling (
adaptive-variable-dependency-modeling) - generic: This is a general challenge in multivariate time-series modeling rather than a specific unresolved problem unique to spectral augmentation.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.14532
- paper_id
- 2604.14532
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
- MIMIC-IV
- concept_slugs
- controlled-spectral-residual-augmentation
- dataset_slugs
- mimic-iv
- skill
- TimeSeriesSkill
- created_at
- 2026-04-19T05:45:26Z