CLaI: Collaborative Learning and Inference for Low-Resolution Physiological Signals: Validation in Clinical Event...

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Critical Care Medicine

Background:

• Machine learning (ML) in clinical data often requires high-resolution data, unavailable in most Intensive Care Units (ICUs).
• Existing ML methods struggle with class imbalance, a common issue in clinical datasets.
• Routinely collected, low-resolution physiological time series data in ICUs remain underutilized for advanced ML tasks.

Purpose of the Study:

• Introduce and validate Collaborative Learning and Inference (CLaI), a novel ML method for event detection and prediction.
• Utilize learned latent representations of multivariate physiological time series, leveraging patient similarities.
• Enable ML interpretability through case-based reasoning for clinical event analysis.

Main Methods:

• Developed CLaI to analyze low-resolution physiological time series by learning from patient similarities.
• Evaluated CLaI on predicting intracranial hypertension (KidsBrainIT dataset) and sepsis (MIMIC-IV dataset).
• Compared CLaI against classification-based and sequence-to-sequence benchmarks, including additional experiments on sepsis detection, class imbalance robustness, and seizure detection (CHB-MIT dataset).

Main Results:

• CLaI effectively detects and predicts events using low-resolution physiological time series data.
• The method demonstrates robustness to class imbalance, consistently achieving competitive performance.
• CLaI achieved the highest F1 score across evaluated tasks, including intracranial hypertension and sepsis prediction.

Conclusions:

• CLaI offers a novel approach for analyzing routinely collected ICU data, overcoming limitations of high-resolution data requirements.
• The method successfully leverages patient similarity for improved event detection and prediction in critical care settings.
• CLaI provides a powerful, interpretable ML tool for critical care event analysis, enhancing clinical decision-making.