EHR Risk Blog
Practical articles on EHR risk prediction, clinical machine learning methods, and U.S. research/policy context. Software is not a medical device and must not guide patient care. Statistics are cited from public sources—verify primary literature before reuse.
Start here (tutorials)
How to prevent data leakage in clinical AI
Index time, temporal splits, and leakage audits for EHR risk models.
How to build an EHR risk prediction model
Docker quickstart with teaching data, calibration, and SHAP.
How it works A–Z
Full workbench walkthrough with screenshots.
Workbench vs ad-hoc notebooks
When a shared leakage-aware loop beats a one-off notebook.
Methods crisis & chronic disease
Label leakage & ICD codes (JAMA 2025)
40.2% of reviewed MIMIC same-admission studies used post-discharge ICD features.
The calibration gap (Brier & ECE)
Why ranking metrics are not enough for trustworthy risk scores.
Chronic disease burden & honest risk models
CDC chronic disease context + leakage-safe methods.
Market growth vs methods quality
Spending surges; pilot-to-production still fails without hygiene.
U.S. policy & research ecosystem
ONC SAFER Guides 2025 & AI transparency
Inventory, oversight, monitoring—methods alignment for teaching.
FDA AI/ML device clearances in 2025
Record clearances; research software is not a device pathway.
HHS AI Strategic Plan (Jan 2025)
Trust, access, workforce—and reproducible research tooling.
AHRQ digital healthcare & AI safety research
Evidence-based AI methods for safety-relevant evaluation.
PCORI AI/ML methods for CER
Funding advances science; open infrastructure helps labs execute.
NIH All of Us & reproducible ML
Large linked data still needs leakage-safe, reproducible methods.