The Boring Part Is the Safety Part

I spent three decades hanging bags in oncology, hematology and immunology. Nobody there asks about the vitamin blend. They ask where the vial came from, what lot it was, who checked it, and what you did the one time a patient went pale on you.

That is the job. The interesting part of infusion care is maybe four percent of it; the rest is procurement, chain of custody, screening and documentation — the part that decides whether anyone gets hurt.

Where the drug comes from is a clinical decision

People treat purchasing as a finance function. In infusion it is a patient-safety function, and the history is not ambiguous.

In 2012 the New England Compounding Center shipped preservative-free methylprednisolone acetate contaminated with fungus. CDC’s final count was 753 cases and 64 deaths across 20 states [1]. Not a dosing error, not a nursing error — a sourcing failure, upstream of every clinician who touched it.

Federal law draws a line anyone buying injectables should understand. Under section 503A of the FD&C Act, a state-licensed pharmacy compounding for one patient is exempt from current good manufacturing practice requirements. Under 503B, a registered outsourcing facility is not [2]. Both are legal; they are not the same risk.

Our default is not to be in that conversation. We buy FDA-approved, manufacturer-labeled product through McKesson, a licensed pharmaceutical wholesale distributor, and vials from LiquiVida. That means the Drug Supply Chain Security Act governs the chain that hands us the box: authorized trading partners, package-level product identifiers, transaction information passed forward, and a defined process for identifying suspect product and notifying FDA [3]. So when a recall lands, “which lot, who got it” is a query — our inventory carries the lot number and expiration date as part of the receipt line’s identity, and two lots of one NDC are two rows.

It constrains the product too. We publicly do not use compounded GLP-1s. FDA has documented what that market produced: patients and prescribers administering five to twenty times the intended dose, from confusing milligrams with “units” on a syringe, with adverse events including pancreatitis [4].

Cold chain, receiving checks, single-dose vials never crossing patients [5] — table stakes, no more forgiving in a living room than in an infusion suite. Home infusion is not inherently more dangerous, but its risk is not evenly distributed: the best prospective data found 0.99 bloodstream infections per 1,000 catheter-days, driven almost entirely by central and multilumen catheters, parenteral nutrition and transplant [6]. We do none of those — our nurses place a peripheral IV and remove it before they leave.

What we screen with, and what we do not claim

We license two Wolters Kluwer databases, and they answer different questions.

Medi-Span is the drug-knowledge base — the patient axis. Drug–drug interactions, allergy and cross-sensitivity screening, disease-state cautions, dose-range checking against age and indication. It is the class of source health systems run their order-entry alerts on: one integrated system’s published work on dose alerts states plainly that “the Medi-Span drug dosing database provided the data for drug-dose decision support” [11].

Trissel’s is the chemistry axis — IV compatibility and injectable-drug stability: whether two things can share a bag or a Y-site, at what concentration, in what base fluid, for how long. Researchers who evaluated seven IV compatibility references concluded that “the highest-performing references... used the compatibility information provided in Trissel’s 2 database as their source of information” — while manufacturers' own labeling covered just 13% of the drug pairs [12].

You need both, and the chemistry half is not optional. In one prospective ICU study, 68.3% of drug request lists contained at least one possible incompatibility [13]: precipitation, occluded lines, lost potency. Nobody photographs that for a brochure.

The patient never free-types their way into a clinical decision. At intake, patients pick their conditions, medications and allergies from a searchable list of licensed concepts, and only a picked concept ID moves downstream. Text matching nothing does not quietly resolve to the nearest string — it is stored as unmatched, carried through verbatim, and lands on the physician’s review card marked unmatched — verify. Deciding clinical outcomes by comparing name strings is how you get hurt; 31 distinct concepts in that database are named “ACE Inhibitors.”

And the evidence is version-pinned like any other dependency — the one I’d underline. Licensed data is under NDA, so it sits outside normal source control, which for a while meant it shipped from whatever was on the deploy machine. Production ran month-stale evidence through August 2026 behind a green health check, and a development copy was later found missing 333 drug-interaction rows and every base-fluid verdict. Nothing looked broken; the physician’s evidence panel was just quietly thinner than it should have been. It now ships pinned by digest, changing only in a reviewed commit, and the running service reports whether all five licensed slices loaded. If they are absent the application does not start — an outage that returns a clean screen looks exactly like a healthy patient.

The AI is the first draft, not the decision

u-wellness is AI-native and a physician signs every plan. Both are true, and the second does less work than people assume. Physician review is the last line of defense; if it is the only one, you have built a system that hands a doctor unfiltered model output and calls the signature safety. Here is what stands between the model and Dr. Sohrabi.

The menu is filtered before the model sees it. A deterministic pre-filter runs the patient’s picked concept IDs against the licensed matrices and removes ingredients outright — contraindicated, not-recommended or extreme-caution disease matches, major interactions, any allergen match. The model is not asked to be careful about ketorolac in severe liver disease; ketorolac is not on its menu. Route matters too: a rule describing only oral exposure becomes a caution, not a silent drop — treating them alike once stripped saline from every patient with edema.

A deterministic chemistry floor runs after the design. One rulebook, where every rule quotes its source document, locator and verbatim sentence: per-session dose ceilings summed across bags, concentration and volume bounds, co-infusion verdicts from Trissel’s. Only three authorities may write in it — Trissel’s, McKesson ingredient specs, LiquiVida product documentation. Thirty-six legacy rules with no admissible document were deleted rather than grandfathered, including one I raised from memory. I was right about the drug and wrong about having a citation, and the citation is what ships.

The failure mode is refusal, not improvisation. A design that trips a hard block is rebuilt, up to three attempts, then the request fails loudly. If both model providers are down, the error propagates. There is no fallback protocol that bypasses contraindication logic — skipping screening is worse than an outage.

The same floor runs again at approval, against the identical rulebook, so approval is never looser than design. Dr. Sohrabi can modify, add or remove; he cannot approve past a hard block. And everything is logged — the exact prompt sent to the model is retained per patient, and every change writes an append-only row naming actor, before and after.

There is a regulatory shape to this. Congress excluded certain clinical decision support software from the device definition on the condition, among others, that it enable the professional “to independently review the basis for such recommendations... so that it is not the intent that such health care professional rely primarily on any of such recommendations” [7]; FDA finalized guidance on those criteria in January 2026 [8]. Showing the physician why — which rule, which document, which sentence — is not a UX nicety. It is the condition.

It also has to stay reviewable, because review degrades into noise. In one inpatient study clinicians overrode 73.3% of medication-related alerts, and only about 60% of those overrides were appropriate [9]. An alert stream nobody trusts is not a safety layer — which is the argument for hard deterministic removal over soft warnings wherever a document supports it.

What a drip bar doesn’t do

The model I keep getting asked about is the drip bar. Walk in, read a board, pick a bag. Immunity, Hangover, Glow. It is a menu, and it was written before anyone knew you existed.

The California Board of Pharmacy uses that word. Its policy statement on IV hydration in unlicensed clinics describes businesses that “offer patients a menu of pre-selected IV mixtures,” and says plainly that IV hydration at a clinic “is a medical treatment that requires an examination with an authorized prescriber before administration.” It adds that such clinics are “generally unregulated” in California, that the Board is “aware of incidents of harm,” and that patients harmed have “very little recourse” [14].

The examination is the part that goes missing. A 2025 JAMA Internal Medicine study reviewed 255 IV hydration spa websites and ran a secret-shopper call to 87 of them. 27.6% required a consultation with a licensed medical professional before treatment. 86.2% recommended a specific therapy for symptoms described over the phone by a stranger. Every site made claims for beneficial use; two cited a source [15].

That is the actual disagreement. Not that vitamins are dangerous — that a bag picked off a board is picked without your medications, your allergies or your dose tolerance anywhere in the room. Nobody screened them, because nobody collected them. What we do instead runs above.

Then there is the part after. In a telephone survey of 63 medical spas, 46% said a medical director or supervising physician would be notified in the event of a complication, and 39.7% had a number to call outside business hours [16]. In the JAMA sample, not one of the 255 sites listed an emergency protocol [15]. You walk out, and it is over.

That is backwards, and the outpatient data says so. In four primary care practices, 25% of patients had an adverse drug event. Twenty-eight percent were ameliorable — the injury was already underway and its severity could have been reduced — and 63% of those were attributed to the physician failing to respond to symptoms the patient had reported. The conclusion is one line: “Monitoring for and acting on symptoms are important” [17].

So we text at one hour and again the next morning, and what the patient reports changes the next plan. Clinical staff read the answers; it is not a satisfaction survey. The infusion is not the end of the engagement. It is the first point at which we know anything.

Scope

Our nurses work inside the California Nursing Practice Act, where a function overlapping medical practice requires a standardized procedure developed jointly by nursing, medicine and administration [10]. Injections are long-settled nursing practice. Deciding what goes in the bag is not — which is why a physician does it.

None of this is exciting. It is lot numbers, concept IDs, quoted package inserts, and a system built to stop rather than guess. That is what I’d want in my arm.

— Michael Mott, RN, Head of Operations, u-wellness

References

1. Centers for Disease Control and Prevention. Multistate Outbreak of Fungal Meningitis and Other Infections — Case Count. Final update October 30, 2015. https://archive.cdc.gov/www_cdc_gov/hai/outbreaks/meningitis-map-large.html

2. U.S. Food and Drug Administration. FD&C Act Provisions that Apply to Human Drug Compounding (sections 503A and 503B). https://www.fda.gov/drugs/human-drug-compounding/fdc-act-provisions-apply-human-drug-compounding

3. U.S. Food and Drug Administration. FDA’s Implementation of Drug Supply Chain Security Act (DSCSA) Requirements. https://www.fda.gov/drugs/drug-supply-chain-security-act-dscsa/fdas-implementation-drug-supply-chain-security-act-dscsa-requirements

4. U.S. Food and Drug Administration. FDA alerts health care providers, compounders and patients of dosing errors associated with compounded injectable semaglutide products. https://www.fda.gov/drugs/human-drug-compounding/fda-alerts-health-care-providers-compounders-and-patients-dosing-errors-associated-compounded

5. Centers for Disease Control and Prevention. Preventing Unsafe Injection Practices. Last reviewed March 26, 2024. https://www.cdc.gov/injection-safety/hcp/clinical-safety/index.html

6. Tokars JI, Cookson ST, McArthur MA, Boyer CL, McGeer AJ, Jarvis WR. Prospective evaluation of risk factors for bloodstream infection in patients receiving home infusion therapy. Ann Intern Med. 1999;131(5):340–347. PMID 10475886. https://pubmed.ncbi.nlm.nih.gov/10475886/

7. 21 U.S.C. § 360j(o)(1)(E) — software functions excluded from the device definition. https://uscode.house.gov/view.xhtml?req=granuleid:USC-prelim-title21-section360j&num=0&edition=prelim

8. U.S. Food and Drug Administration. Clinical Decision Support Software — final guidance, issued January 6, 2026 and re-issued January 29, 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software

9. Nanji KC, Seger DL, Slight SP, et al. Medication-related clinical decision support alert overrides in inpatients. J Am Med Inform Assoc. 2018;25(5):476–481. PMID 29092059. https://pmc.ncbi.nlm.nih.gov/articles/PMC7646870/

10. California Board of Registered Nursing. An Explanation of the Scope of RN Practice Including Standardized Procedures, NPR-B-03 (rev. 01/2011). https://www.rn.ca.gov/pdfs/regulations/npr-b-03.pdf

11. Saiyed SM, Greco PJ, Fernandes G, Kaelber DC. Optimizing drug-dose alerts using commercial software throughout an integrated health care system. J Am Med Inform Assoc. 2017;24(6):1149–1154. PMID 28444383. https://pmc.ncbi.nlm.nih.gov/articles/PMC7651973/

12. Smith WD, Karpinski JP, Timpe EM, Hatton RC. Evaluation of seven i.v. drug compatibility references by using requests from a drug information center. Am J Health Syst Pharm. 2009;66(15):1369–1375. PMID 19635773. https://pubmed.ncbi.nlm.nih.gov/19635773/

13. Kayra T, Ferliçolak L, Ateş İ, Altıntaş ND, Onay-Besikci A. Clinical pharmacist-led prevention of intravenous drug incompatibilities in an adult ICU: a prospective observational study. BMC Pharmacol Toxicol. 2026. DOI 10.1186/s40360-026-01095-2. https://pmc.ncbi.nlm.nih.gov/articles/PMC13041401/

14. California State Board of Pharmacy. Policy Statement Related to Risks to Patients Receiving Intravenous Hydration in Unlicensed Clinics or Locations. https://www.pharmacy.ca.gov/about/intravenous_hydration_policy.pdf

15. Sivakumar A, Forman HP, Wang I, Lurie P, Ross JS. State Policies and Facility Practices of IV Hydration Spas in the US. JAMA Intern Med. 2025;185(12):1455–1461. DOI 10.1001/jamainternmed.2025.5028. PMID 41051745. https://pmc.ncbi.nlm.nih.gov/articles/PMC12501857/

16. Almukhtar RM, Aigen AR, Loyal J, Mishra V. Supervision Unveiled: Navigating the Supervision Landscape in Medical Spas. Dermatol Surg. 2024;50(10):954–957. DOI 10.1097/DSS.0000000000004248. PMID 38771947. https://pubmed.ncbi.nlm.nih.gov/38771947/

17. Gandhi TK, Weingart SN, Borus J, et al. Adverse drug events in ambulatory care. N Engl J Med. 2003;348(16):1556–1564. DOI 10.1056/NEJMsa020703. PMID 12700376. https://pubmed.ncbi.nlm.nih.gov/12700376/