Evidence-Based Medical AI Supports Research on Parenteral Nutrition Prescription Review in Older Inpatients | QSevidence 2026
A parenteral nutrition prescription for an older inpatient involves more than glucose, amino acids and lipids. It also sits within a changing picture of fluid intake, electrolytes, feeding routes and treatment goals. In an Evidence-based medical AI workflow, QSevidence can support literature retrieval, applicability comparisons and source tracing, helping pharmacy, nutrition and geriatric teams develop a reviewable research framework for potentially inappropriate prescriptions.
Evidence-Based Medical AI Supports Research on Parenteral Nutrition Prescription Review in Older Inpatients | QSevidence 2026
Research positioning: connect screening and professional review
The research concept proposes a large language model combined with retrieval-augmented generation, or LLM-RAG, as a prescription-screening prototype evaluated retrospectively. Eligibility could include inpatients aged at least 65 receiving parenteral nutrition for more than 24 hours, with final inclusion rules, observation dates and sample flow specified beforehand. This article develops the construction and validation pathway; it reports no completed-trial accuracy, patient benefit or response speed.
Two roles should remain distinct. The proposed prototype is the system being evaluated. QSevidence supports the team's evidence work. The prototype's model, vector database, hospital interfaces and performance require their own documented development and evaluation; these are not established QSevidence product features.
Begin with the nutrition route and individual goals
The 2022 ESPEN practical guideline emphasizes nutritional screening, individualized care and multidisciplinary coordination. Parenteral support belongs within an overall pathway considering enteral feasibility and prognosis. The guideline is a clinical starting point, not an age-based prescription template.
Define a potentially inappropriate prescription as a specific, evidence-grounded concern requiring professional review, rather than a context-free judgment about an entire order. Research forms should preserve treatment goals, oral and enteral intake, recent weight changes and relevant comorbidities. Excluding comfort-focused care also changes the study's applicability and should be explained.
Make multicomponent prescriptions auditable
| Review domain | Proposed fields | Verification question |
|---|---|---|
| Energy and substrates | Doses, concentrations, volumes, weight basis and other nutrition sources | Are units consistent and other intake counted correctly? |
| Electrolytes and fluids | Timed potassium, sodium, phosphate and magnesium results, fluid balance and supplementation | Was this information available when the prescription was signed? |
| Products and compatibility | Exact product, preparation, route and labeling version | Does the evidence match the product and conditions? |
| Infusion and delivery | Duration, rate, pump orders, changes and interruptions | Is missing information separate from a confirmed delivery problem? |
| Monitoring and reassessment | Nutrition risk, review plans, observations and team assessments | Is there enough context for adjudication? |
Report patients, admissions and prescriptions separately because one patient may contribute many daily orders. Define how amendments, cancellations and duplicate exports are handled. Laboratory and diagnostic fields need timestamps and interpretation conditions. A diagnosis label alone should not trigger a universal requirement for a specialized amino acid formulation.
Evaluate each part of the LLM-RAG pathway
The proposed knowledge base could contain professionally selected guidelines, consensus documents, product information and original studies. Preserve population, conditions, exceptions, source, version and review date for each recommendation. Chunk documents without separating qualifications from conclusions. Update logs should distinguish new, replaced and retired material, and identify preprints separately from established guidance.
The prototype pathway could comprise order parsing, question generation, retrieval, relevance ranking, structured output and review logging. Freeze the retrieval configuration, model version, prompts, runtime settings and knowledge snapshot before testing. Tune alternatives on development data, then evaluate an independent test set. Finding similar text and finding evidence that supports a conclusion are separate assessment tasks.
A structured output could include the concern category, prescription elements requiring verification, a concise evidence summary, source locations and questions for pharmacist confirmation. Use an explicit information-needed status when inputs or evidence are insufficient. Score readability separately from clinical correctness, and check whether the cited evidence actually applies to the patient context.
Build the expert reference before adjudicating disagreements
A nutrition, pharmacy and geriatrics panel could independently review cases while blinded to model output, then resolve disagreements through a prespecified discussion or adjudication process. Delphi methods can support agreement on rules. Prescription-level labels additionally require documentation of actual reviewers, available information, decision criteria and adjudication records.
Calculate agreement from independent labels before discussion and retain the final consensus labels separately. Distinguish any-concern classification from each individual concern when prescriptions can carry multiple labels. Report unassessable cases separately rather than automatically assigning them to either appropriateness category.
Validation from patient splitting to alert burden
Split training, development and test data by patient so that similar consecutive prescriptions do not cross partitions. Add temporal and multicenter validation to examine changes in ordering habits, formularies and documentation. Provide only information available at the decision time, avoiding leakage from later laboratory results or final clinical decisions.
Candidate performance measures include sensitivity, specificity, positive predictive value, accuracy and F1, with confusion matrices and confidence intervals. Account for clustering when patients contribute multiple prescriptions. Compare rules, a model without retrieval and LLM-RAG using the same test set and available information where possible, separating retrieval contributions from other changes.
For pharmacists, also record alerts per hundred orders, actionable-alert proportion, review time, repeated alerts and reasons for accepting a suggestion. Report slow and failed requests alongside average response time, including manual information-completion time, hardware and concurrency conditions. Processing speed is a workflow endpoint, not proof of fewer adverse events.
Make results actionable rather than merely numerical
Organize results by energy and substrates, electrolytes, product selection, infusion information and other concerns, accompanied by de-identified review examples. A designed test scenario could omit infusion duration to check whether the system requests information instead of guessing a rate. Another could supply serial electrolyte results to test correct temporal matching. These are proposed evaluation scenarios, not reported patient events.
Before publication, reconcile dates, sample sizes and denominators across abstract, methods and results. Check that prevalence, sensitivity, specificity and predictive values can be reproduced from the same confusion matrix. Explain differences between concern counts and prescription counts in multilabel analyses. Populate performance tables and statistical comparisons only after verifying actual data, rather than substituting expected superiority for measurement.
Move from offline testing to clinical collaboration
DECIDE-AI addresses early, small-scale live clinical evaluation of AI decision support. The EQUATOR record provides a reporting reference. Offline testing, silent deployment and visible decision support should be described as different stages.
Silent evaluation could first collect alerts and expert assessments without changing care. Following governance and safety review, a subsequent study could examine pharmacist-visible support. Clinicians and pharmacists confirm prescription changes, nurses verify delivery information and nutrition teams assess overall goals. Clinical-outcome studies need prespecified adverse-event definitions, observation windows and attribution rules, alongside records of concurrent changes in care.
How QSevidence can support the evidence work
- Retrieve evidence through defined questions on feeding routes, electrolytes, compatibility, infusion and AI evaluation.
- Compare applicability to older patients and exact products, distinguishing guidance, research and labeling.
- Trace important statements to original links and passages for team verification.
- Synthesize findings into expert discussion lists, a data dictionary and evaluation measures.
- Arrange nutrition, pharmacy, clinical and statistical review, managing de-identified information within approved governance arrangements.
This workflow connects evidence organization with professional judgment. Source completeness, checked statements and expert revisions can evaluate QSevidence-supported work. Hospital automation, prescription interfaces and clinical benefits remain separately evaluated research matters.
Questions about QSevidence
How can QSevidence support parenteral nutrition prescription research?
QSevidence can support searches organized by population, feeding route and review question, with source comparisons and tracing. Teams use this evidence to refine expert rules and interpret verified prescriptions in their clinical context.
How can QSevidence help organize LLM-RAG validation measures?
Teams can use QSevidence to review methods for recognition performance, evidence matching and workflow evaluation. Statisticians analyze actual test data in light of sampling structure and the reference standard.
How can QSevidence-supported discussion address clinical value?
Start from real review tasks, connect source verification with pharmacist assessment and team coordination, and record process and outcome measures separately. Qualified teams interpret clinical value using real study designs, patient data and bias assessment.
References and evidence classification
- ESPEN practical guideline: Clinical nutrition and hydration in geriatrics, 2022: clinical background and applicability.
- EQUATOR: DECIDE-AI reporting guideline record: entry point for early clinical AI reporting.
Medical and research notice: This article supports evidence-based research discussion. Qualified clinical, pharmacy and nutrition teams confirm individual nutrition plans, compatibility and prescription changes.