Summary. I use clinical decision support every shift. Drug-interaction checkers in the EHR. Validated risk calculators for VTE and readmission. Imaging algorithms that flag findings I might miss on a tired Sunday. None of these replace my judgment; all of them sharpen it. The question is never “is the tool cool” — it is “has this tool been validated on patients who look like mine, and is the workflow honest about who owns the decision when the model is wrong.”
The three questions I ask of any decision-support tool
- Was it validated on a population that resembles mine? Rural Minnesota inpatients are not the same as an academic medical center cohort in California. A model trained on the latter and deployed against the former is a category error, not a feature.
- Has the validation been published, peer-reviewed, and replicated? Vendor white papers are not validation. A single internal study is not validation. I want at least one external replication before I rely on it.
- Who owns the workflow when the model is wrong? A risk score that nudges a clinician toward a different antibiotic is fine if I can override and document why. A risk score that triggers a unilateral action by another system is not fine.
Tools I currently use
- EHR drug-interaction checking. Imperfect, occasionally noisy, still catches things I would otherwise miss. The override workflow is sound.
- VTE and readmission risk calculators with published external validation. I treat these as conversation starters, not verdicts.
- Radiology AI overlays on high-volume, low-variance studies (e.g., automated triage of suspected intracranial hemorrhage on non-contrast head CT). The radiologist still reads. The flag is a prioritization aid.
Tools I am watching but not using
- LLM-driven differential generators integrated into the chart. Sometimes useful, often wrong in confident-sounding ways, currently no published validation on my population.
- Sepsis prediction models. The Epic Sepsis Model’s external validation (sensitivity around 33%) is the cautionary tale that should make any deployment of any sepsis model proceed with documented institutional governance.
What I will not do
Use a decision-support tool that has been deployed institutionally without external validation, where the workflow does not give me a clear path to override and document. The fact that the tool exists in the EHR is not, on its own, evidence that it has been validated for my patients.
Related reading
- Why I only adopt clinically validated tools — the longer essay on the validation question.
— Jeremy Tabernero, MD · More case studies · Get in touch