Causal Machine Learning for Heterogeneous Treatment Effects from Routine Hospital Records
Causal Machine Learning for Heterogeneous Treatment Effects from Routine Hospital Records
Description
Details
Context and Problem Statement
Predictive hospital models answer associational questions, while treatment decisions require causal reasoning. Observational treatment histories are confounded because prior clinical decisions already influenced outcomes. This project focuses on rigorous estimation and robustness analysis rather than automated prescribing.
Research Question
To what extent are heterogeneous treatment-effect estimates stable when identification assumptions and the magnitude of unmeasured confounding are varied?
Proposed Approach
Specify a causal graph with clinical input, estimate heterogeneous effects with meta-learners, causal forests, and propensity-score methods, and perform sensitivity analyses quantifying the strength of hidden confounding required to overturn conclusions.
Expected Contribution
A practical robustness protocol for observational causal analysis in hospital data.
Expected Prototype
A reproducible analysis library producing subgroup effect reports with sensitivity bounds for research use.
Datasets
MIMIC-IV, eICU, public trial data where available, and anonymized local hospital data under agreement.
Challenges
Unmeasured confounding, indication bias, incomplete records, over-interpretation risk, and ethical data governance.
Research Question
Innovation
Expected Deliverable
Technologies
Required Skills
- Causal Inference and Advanced Statistics
- Structured Clinical Data Processing
- Scientific Programming in R or Python
- Collaboration with Clinicians
Datasets
- MIMIC-IV (PhysioNet, controlled access)
- eICU Collaborative Research Database
- Local hospital data under formal agreement