Blogs

A patient needs an infusion and three sites are plausible: hospital outpatient department, ambulatory infusion suite, or home infusion. The safest answer may not be the cheapest answer. The cheapest answer may fail if nursing skill, monitoring, handling, or home support are not adequate.

Share: An infusion site-of-care decision is not just hospital versus home. Safety, monitoring, supplier capacity, patient burden, policy, and avoidable cost all interact. #Causality #Healthcare #SiteOfCare

The Problem Is Not Always One Thing

Site-of-care decisions are easy to flatten into a cost argument. That misses the clinical and practical pieces: reaction risk, monitoring need, drug handling, travel burden, home support, supplier capacity, payer policy, and out-of-pocket exposure. The better question is which site is low value for this case, not which site is cheapest in the abstract.

Why The Obvious Explanation Can Mislead

The obvious answer is to move the infusion to the lower-cost site. That may be right, but only if safety and treatment completion remain acceptable. A site-of-care decision is a decision under uncertainty, which is why this article uses an influence-style query after the graph is clear.

The Question For This Example

This example asks how low-value-site risk changes after observing a difficult infusion context, and what changes if an ambulatory site review is actually performed.

The target outcome is LowValueSiteDecision, which means low-value site decision in this teaching model.

A Small Model Of The Situation

The model is an influence-style teaching example. The BBN estimates risks for safety, treatment completion, avoidable cost, and low-value site choice. The Python SDK example shows how a decision-style comparison can score an action. It is not a clinical protocol or coverage policy.

The graph has 18 nodes:

• InfusionReactionRisk: Infusion reaction risk (low / high)
• MonitoringNeed: Monitoring need (low / high)
• DrugHandlingComplexity: Drug handling complexity (low / high)
• PatientMobility: Patient mobility (adequate / low)
• HomeSupportReliability: Home support reliability (adequate / low)
• TravelBurden: Travel burden (low / high)
• HospitalSiteAvailability: Hospital site availability (adequate / low)
• AmbulatorySiteAvailability: Ambulatory site availability (adequate / low)
• HomeInfusionSupplierAvailability: Home infusion supplier availability (adequate / low)
• NursingSkillMatch: Nursing skill match (good / poor)
• PayerSitePolicy: Payer site policy (clear / unclear)
• OutOfPocketExposure: Out-of-pocket exposure (low / high)
• MissedWorkBurden: Missed work burden (low / high)
• AmbulatorySiteReview: Ambulatory site review (yes / no)
• TreatmentCompletionRisk: Treatment completion risk (low / high)
• SafetyEventRisk: Safety event risk (low / high)
• AvoidableCostRisk: Avoidable cost risk (low / high)
• LowValueSiteDecision: Low-value site decision (no / yes)

What Is Sourced And What Is Synthetic

Medicare and CMS home-infusion materials support including home infusion services, equipment, monitoring, and coverage context. AHRQ supports considering patient safety by setting. NHIA’s literature review supports including cost differences across sites of care as a motivation, while the probabilities remain synthetic.

Home infusion services

Medicare and CMS materials support including home infusion equipment, services, caregiver training, and monitoring.

Safety by setting

AHRQ patient-safety resources support treating safety risk as setting-dependent.

Cost by site of care

NHIA’s review supports including cost differences across home, outpatient, and inpatient or hospital settings as a planning concern.

The Query The Graph Cannot Show By Itself

The visual graph lets readers inspect how evidence moves the target risk. The decision part is easier to express in code because we want to score a site-of-care choice, not just ask for one posterior probability.

Using the Darkstar Python SDK, a subscribed SDK user can load the same model, set the site-review choice, and compare the resulting low-value-site risk. The download includes this as illustrative SDK code, while the standalone query_example.py reproduces the article numbers without shipping the SDK.

import json

from darkstar import DiscreteModel

TARGET = "LowValueSiteDecision"
OBSERVED_EVIDENCE = {
    "InfusionReactionRisk": "high",
    "TravelBurden": "high",
    "HospitalSiteAvailability": "low",
    "PayerSitePolicy": "unclear"
}


def probability_after_setting(model, setting, evidence):
    changed_model = model.intervene(json.dumps(setting))
    query = {"nodes": [TARGET], "evidence": evidence}
    result = json.loads(changed_model.pquery(json.dumps(query)))
    return result[TARGET]["yes"]


with DiscreteModel(model_json) as model:
    no_review = probability_after_setting(
        model,
        {"AmbulatorySiteReview": "no"},
        OBSERVED_EVIDENCE,
    )
    reviewed = probability_after_setting(
        model,
        {"AmbulatorySiteReview": "yes"},
        OBSERVED_EVIDENCE,
    )

In plain English, the code asks: if the same patient and coverage context is present, how much does the low-value-site risk change when an ambulatory site review is actually performed?

The standalone query code gives this synthetic result:

Baseline low-value site decision: 29.2%
Observed scenario (high reaction risk, high travel burden, unclear policy): 37.2%
Action scenario (perform ambulatory site review): 28.2%
Same context with the action left unfavorable: 49.8%
Synthetic reduction versus the observed scenario: 9.1 percentage points

What We Learn By Observing The Situation

The observed scenario sets:

• Infusion reaction risk = high
• Travel burden = high
• Hospital site availability = low
• Payer site policy = unclear

The observed scenario sets high infusion-reaction risk, high travel burden, low hospital availability, and unclear payer site policy. In the synthetic model, those facts raise low-value-site risk because both safety and avoidable-cost pathways are stressed.

What Changes If We Set One Condition Differently

The action scenario sets:

• Ambulatory site review = yes

The action query sets ambulatory site review to yes. The model treats that review as a structured check, not a magic site switch. It lowers risk when it helps align site choice with clinical monitoring, capacity, payer rules, and patient burden.

What This Example Cannot Prove

The model cannot decide where a real patient should receive therapy, replace clinical judgment, or determine coverage. It is a teaching example for why a site-of-care choice needs both probability and utility thinking.

Try The Model

Open the companion Darkstar model: Open Infusion Site Of Care Influence BBN

Set MonitoringNeed to high and HomeSupportReliability to low, then compare the ambulatory review setting. The point is to see that the review matters most when the site choice is otherwise ambiguous.

Download the teaching package: Infusion Site Of Care Influence BBN ZIP

The ZIP includes the synthetic model, standalone query code, an illustrative SDK influence-style snippet, tests, and figures. It does not include the SDK itself, clinical protocols, patient data, payer contracts, or chart-generation code.

Sources

1. Medicare.gov: Home Infusion Therapy Services, Equipment, and Supplies
2. CMS Home Infusion Therapy Services
3. NHIA: Cost Savings, Home Versus Inpatient Infusion Therapy
4. AHRQ Patient Safety Resources by Setting