From Composable Forms to a Clinical Flywheel: A Contextual Module Recommender and Patient Monitoring in VibeRounds
In the previous post, we made the EMR composable. A case starts blank, and clinicians add element sets (disease contexts, care pathways, templates) as the patient's picture evolves.
That solves how structure is assembled. It leaves open who decides which structure is needed, and when.
Right now the clinician has to know what to add. In a busy clinic, nobody remembers which of fifty modules applies to a 62-year-old with diabetes, stage 3b CKD and a new foot ulcer. This post covers the next layer: a Contextual Module Recommender, and a patient-facing counterpart that extends the same modules to home monitoring.
The Problem: A Marketplace Nobody Has Time to Browse
Suppose element sets are published as standard FHIR R4 SDC Questionnaires, reviewed by an informatician and shared in a catalog. A catalog is still passive. A "form library" works only if the clinician searches it, and clinicians under time pressure don't.
The recommender turns the catalog into something that works at the point of care. It reads the live encounter state and surfaces the few modules that matter right now.
How the Recommender Works
|
|
|
|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
The split in step 3 follows the same rule as the rest of VibeRounds: deterministic logic decides what is allowed, and the LLM only helps with ordering and explanation.
Three Design Patterns That Keep It Trustworthy
1. Micro-modules, not monolithic forms
An 80-question specialty template is hard to reuse and hard to review. Small, focused modules (Acute Dyspnea Triage, Anticoagulation Bleeding Risk, Diabetic Foot Screen) can be combined. When a patient has overlapping problems, the recommender picks the two or three modules for that specific intersection. This is the same property we relied on in the trajectory post: element sets compose, templates don't.
2. Redundancy suppression
Before recommending anything, the recommender checks recent QuestionnaireResponse and Observation timestamps. If a PHQ-9 or smoking-status screen is already inside its valid window, it is suppressed. Clinicians trust a recommender that respects what is already charted, and ignore one that asks the same questions again.
3. A closed feedback loop
Every outcome (accepted, dismissed, modified) is logged using the CDS Hooks feedback pattern. The signals are practical:
- A high dismissal rate on a module tells the informatician to tighten its
useContextcriteria. - Frequent free-text additions suggest a missing field, and the LLM pipeline can draft a new micro-module for human review.
The catalog then improves from real use instead of drifting.
A Worked Example
Patient: 62-year-old man, Type 2 diabetes, CKD stage 3b, presenting with a foot ulcer.
|
|
|
|---|---|---|
|
|
|
|
|
|
|
|
|
An external EHR gets two clean CDS Cards. VibeRounds merges both modules into one pre-populated workspace with the relevant trajectories visible. The clinician sees two high-signal suggestions, not fifteen.
(The module names above are illustrative. Real triggers and thresholds would be defined and verified by a clinician or informatician, not inferred by the model.)
Extending the Same Module to the Patient
A well-designed SDC Questionnaire separates clinical meaning (codes, scoring, extraction rules) from presentation (wording, layout, schedule). That means one marketplace module can have two projections, so no separate patient-form engine is needed.
|
|
|
|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
What makes focused monitoring work
- Condition-specific micro-checklists. On discharge or diagnosis (heart failure, gestational diabetes, a post-antibiotic course), the recommender assigns a short 4 to 5 item module matched to the patient's regimen, not a generic health tracker.
- Adaptive questioning. With
enableWhen, a patient who reports "feeling fine, weight stable" finishes in a couple of taps. If weight jumps more than 2 kg in 48 hours or breathlessness is ticked, follow-up safety questions expand and the clinician is alerted. - Trajectory ingestion. Every submission goes through the same SDC extraction pipeline into the same longitudinal store. By the follow-up visit, the clinician's encounter form can already show a 14-day symptom and vital trend.
- Portable ownership. Patients can share a structured monitoring summary with an outside clinic through a SMART Health Link instead of a paper log.
(Alert thresholds such as the 2 kg example are examples only. Real thresholds must be set and approved clinically for each condition.)
The Ecosystem in One Picture
- Marketplace: LLM-drafted, deterministically validated, informatician-approved SDC modules, each with a clinician and a patient projection.
- Clinician side: context-triggered forms and comorbidity composition in VibeRounds, with the same modules reachable from external EHRs via CDS Hooks.
- Patient side: adaptive home check-ins whose structured results feed the clinical reasoning loop.
The loop is: catalog → recommend → compose → patient check-in → structured extraction → better recommendation.
What Still Has to Be Proven
An architecture is a hypothesis until it is tested. The risks I would measure first:
- Terminology and logic drift. Do spliced modules ever ask conflicting questions or write to the same resource differently?
- Governance load. Is reviewing an LLM-drafted module faster than building one by hand? If not, the catalog will not grow.
- Alert fatigue. Does the recommender actually reduce clicks and missed risks, or does it just add a new kind of clutter?
- Round-trip fidelity. Does one module keep its meaning from catalog to clinician view to patient check-in to FHIR extraction, across more than one EHR?
The plan is to pick one narrow scenario (for example Type 2 diabetes with CKD), define two composable modules, wire a basic CDS Hooks trigger, and measure time saved, suppression accuracy and extraction fidelity on synthetic cases before touching anything real.
Why This Matters
The first post made the EMR respond to a diagnosis. The second made it follow the patient over time. This one asks the EMR to anticipate what the clinician and the patient will need next, while keeping humans, standards and deterministic checks in charge of what is safe.
Previous posts: Beyond Static EMRs · From Diagnosis to Trajectory & Pathways
Project repository: VibeRounds-Clinical-Analytics-App
Live demo: https://viberounds-clinical-analytics-app.onrender.com/ (contact me for access: avi33tbtt@gmail.com)
Disclaimer: this is a design direction and prototype using synthetic data. It is not for clinical use.
No comments:
Post a Comment