Causal inference for time series datasets with partially overlapping variables.

Area of Science:

• Biomedical Informatics
• Causal Inference
• Health Data Science

Background:

• Observational healthcare data offers rich insights for causal inference.
• Incomplete and non-standardized variables across patient datasets present significant challenges.
• Existing causal inference methods struggle with missing data, reducing analytical power or generalizability.

Purpose of the Study:

• To develop a novel method for causal inference from time series data with partially overlapping variables.
• To address the challenges posed by missing data in large-scale observational healthcare datasets.
• To improve the accuracy and applicability of causal discovery in complex health data.

Main Methods:

• Propose Causal Model Combination for Time Series (CMC-TS).
• Leverage partial variable overlap between datasets to reconstruct missing information.
• Iteratively correct errors and reweight inferences using shared data.

Main Results:

• CMC-TS demonstrated superior performance on simulated data, achieving the lowest false discovery rate and highest F1-score.
• Evaluation on real-world neurological intensive care unit (ICU) stroke patient data identified fewer implausible and more plausible causes of adverse events.
• The method effectively handles missing variables by utilizing overlapping information across patient records.

Conclusions:

• CMC-TS enables robust causal inference from patient data with partially overlapping variable sets.
• This approach enhances the utility of observational healthcare data for discovering causal relationships.
• The findings suggest a pathway for more effective utilization of complex, real-world health datasets.