Composable, Personalized Information for Home Enteral Nutrition

Keywords: home enteral nutrition · feeding tubes · patient education · health literacy · human-centered AI · medical safety

1Introduction

Enteral nutrition — feeding delivered directly into the stomach or small intestine through a tube — sustains people who cannot safely meet their nutritional needs by mouth. As care has shifted from the hospital to the home, a growing population now manages this themselves, every day, often for months or years. A gastrostomy (G), gastrojejunostomy (GJ), jejunostomy (J), or nasal (NG/NJ) tube is not merely a medical device; it reorganizes a person’s daily life around routines that must be performed correctly to avoid harm.

The people doing this work are usually not clinicians. They are patients and the family members who care for them — disproportionately older adults — who learned the essentials during a short, stressful window around discharge and were then sent home to practice. When a question comes up at 2 a.m. (“the tube won’t flush,” “there’s redness at the site,” “can I crush this pill?”), the resources within reach are rarely equal to the moment.

2The information gap

2.1 Fragmentation

Guidance is scattered across discharge paperwork, device and formula manufacturer documents, general-purpose search results, and online communities. The pieces are inconsistent, vary in quality, and place the burden of reconciliation on the reader least equipped to do it. Trust is hard to establish when every source looks different and none of them knows the reader’s situation.

2.2 The literacy mismatch

Much of what exists is written for clinicians. A substantial share of adults have limited health literacy, and clinical phrasing — “maintain patency,” “assess for aspiration” — does not translate into confident action at the bedside. The cost is not only comprehension but anxiety: uncertainty about a safety-critical task is itself a burden.

2.3 One answer for everyone

The right guidance depends on specifics — the type of tube, the formula, how long the person has had it, their age — and these specifics change the answer. A flush volume, a venting technique, or a medication caveat appropriate for one setup can be wrong for another. Generic content cannot make these distinctions, so it hedges, and hedged advice is hard to act on.

2.4 The stakes

Tube-care problems — clogs, dislodgement, aspiration, dehydration, and site infection — are common and consequential, and the transition from hospital to home is a recognized window of elevated risk. Information that is hard to find, hard to read, or not specific enough is not a minor inconvenience here; it is a safety issue.

2.5 Why existing tools fall short

Static articles and PDFs cannot personalize and are not built to be scanned or acted on under stress. Open-ended chatbots can personalize and converse, but they can also hallucinate, render unconstrained content, and produce confident, unsafe answers — an unacceptable failure mode in this domain. Neither approach treats medical safety as a first-class, enforceable property.

3Design principles

Lugano is built on a small set of commitments that the rest of the system serves:

• Safety first. Never diagnose; never render free-form markup from a model; keep disclaimers structural; make emergency pathways mandatory where they belong.
• Start from what we already know. The person’s profile — tube type, duration, formula, age, primary concern — is given to the system and used directly, so the reader is met as someone known, not surveyed.
• Plain language, in the second person, in familiar units. Specific and actionable (“flush with 1–2 oz (30–60 mL) of water”), not vague.
• Built for older adults. Large type, high contrast, generous spacing, large touch targets, clear hierarchy.
• Composition over prose. Information takes the shape of its content — a sequence becomes steps, a comparison becomes two columns — so it can be scanned and trusted.

4The Lugano approach: composable cards

4.1 The block model

The unit of information is a card: an ordered sequence of typed blocks — a step-by-step procedure, a checklist, a side-by-side comparison, a table, a callout, a symptom triage guide, a calculator, a nutrition-facts panel, and others. Each block type is rendered by a single, reviewed interface component. Critically, the AI produces structured data that must conform to a fixed schema for each block — it never produces markup or HTML. This design choice eliminates an entire class of safety and injection risks: there is no path by which model output becomes executable or arbitrarily-styled content on the page.

4.2 Generation grounded in the profile

When a person asks a question, the system answers from their profile, assembling a card whose blocks are chosen to fit the shape of the answer. Every generated card is validated against strict schemas before it is stored or shown; output that fails structural or quality checks is rejected rather than displayed. Personalization is therefore data-driven and auditable, not improvised in a prompt.

4.3 Safety as an invariant

Because blocks are typed, safety rules can be enforced by construction rather than hoped for in wording. Symptom and triage blocks must include both an emergency pathway and a “contact your care team” pathway. Calculators present reference ranges as informational, never prescriptive, and direct the reader to their clinicians. Medical disclaimers are part of the structure of a health card, not a footnote that can be omitted. These are properties of the system, checked the same way every time.

4.4 The personal board

Cards are not disposable chat replies; they accumulate into a durable, personal board the patient or caregiver returns to and curates, pinning the takeaways that matter to them. Over time the board becomes a tailored manual for one person’s care, rather than a generic booklet that fits no one exactly.

4.5 From question to card

Behind a single card is a sequence of stages, each a checkpoint rather than a leap of faith. The person’s question is first joined with what the system already knows about them — tube type, formula, duration, primary concern — to form an enriched query. That query is grounded: the system draws on the open web and on Lugano’s own growing corpus of community-created, personalized cards, so an answer reflects accumulated, real-world tube-care knowledge rather than a model’s general recall alone. From this a precise prompt is composed and routed to a leading language model through a model-agnostic layer, so the system is never bound to a single vendor and can adopt a stronger model as soon as one exists.

The model is required to answer not in free prose but in Lugano’s structured blocks. Every returned block is validated against its schema — anything malformed or unexpected is dropped rather than shown — and the content is sanitized so that no markup or executable text can reach the page. The draft then passes a quality gate: it is scored against an explicit checklist — is it specific, is it well-structured, does it cite sources, does it open with empathy, does it say when to seek help — and if it falls short on any point, the system rewrites it and checks again, automatically, before the reader ever sees it. A final medical-safety gate enforces the invariants above. Only then is the card delivered — and the reader can refine it further by answering a couple of curated, situation-aware questions.

5Why composition

The choice of composition over either static content or open conversation is deliberate. Structured blocks are scannable and actionable under stress, where a wall of prose is not. A constrained component set makes safety auditable and consistent — the same vetted triage component behaves the same way everywhere. Personalization becomes a property of data and schema rather than of a single fragile prompt. And the constraint that the model emits data, not markup, converts “please don’t render anything unsafe” from a hope into a structural guarantee. Conversation is a fine way to ask; it is a poor way to render high-stakes instructions.

6Understanding need: search as a privacy-preserving signal

To improve, a system has to learn where people actually struggle. Lugano classifies each question into a coarse topic — flushing, feeding, formula and nutrition, hydration, medications, site care, troubleshooting, supplies — and records only that derived type, never the raw text of the question. This is deliberate data minimization: we keep enough to see that “flushing” and “medications” dominate a person’s month, or that a topic is rising across the community, without retaining sensitive free text. The result is a feedback loop — from real questions to better coverage — that moves the corpus toward continual improvement while respecting the sensitivity of health questions.

This is the engine of Lugano’s compounding advantage. As more people use it, the corpus of community-created, personally-tailored cards grows, and the system’s picture of what different kinds of patients — by tube type, age, and concern — actually struggle with sharpens. That understanding feeds back into better answers, better personalization questions, and better cards, so the knowledge base widens with every card in a way no general-purpose model can match.

7Current status and evaluation

We are candid about stage. Lugano is early and design-led. Today the system gates every generated card on a combination of structural validity and a quality score, and we track coverage and engagement across the topics above. We have not conducted clinical outcome trials, and we make no claims of clinical efficacy; the claims in this paper are about information design and safety architecture, not patient outcomes. Our near-term evaluation plan centers on clinician review of generated content, comprehension testing with patients and caregivers, and instrumented measures of whether people can find and act on what they need.

8Roadmap

• Clinician-in-the-loop. Expert review and co-authored cards, so the corpus is shaped by professionals as well as by demand.
• More block types and controls. Broader coverage, reading-level controls, and multilingual support for the families who need them.
• Continuity with the care team. Lightweight longitudinal signals — routines, comfort, mood — that a patient can choose to share with their clinicians between visits.
• A community-and-evidence loop. Combining real questions, expert curation, and open discussion into knowledge that keeps getting sharper.

9Limitations and responsible use

Lugano provides educational information and is not a substitute for professional medical advice, diagnosis, or treatment. In an emergency, contact local emergency services or the care team directly. AI systems can err; the constrained, schema-validated component model reduces but does not eliminate that risk, and the human care team remains central to every decision. This is a working paper, not a peer-reviewed publication; any specific figures cited in future revisions will be sourced, and quantitative claims here are intentionally qualitative pending that work.

Selected sources informing this work

• American Society for Parenteral and Enteral Nutrition (ASPEN) — clinical guidelines and patient resources for enteral nutrition.
• The Oley Foundation — education and peer support for people on home enteral and parenteral nutrition.
• U.S. Department of Health and Human Services and the Agency for Healthcare Research and Quality (AHRQ) — health-literacy research and plain-language guidance.
• Device and formula manufacturers’ instructions for use (IFU) for enteral feeding tubes and products.

A future revision will add formal, numbered citations and quantitative figures. This version is a living draft.