Table 2. The GIRISH framework: six steps for safe, effective AI use at the NICU bedside.

BNFc: British National Formulary for Children; BPD: bronchopulmonary dysplasia; BW: Birth weight; CRP: C-reactive protein; DOB: Date of birth; DOL: Day of life; FiO₂: Fraction of inspired oxygen; GA: Gestational age; GIRISH: Goal, Input, Role, Iterative refinement, Safety verification, and Human accountability; LLM: Large language model; MRN: Medical record number; PMA: Post-menstrual age; WBC: White blood cell count.

Step	What you do	Illustrative prompt	Clinical safety note
G – Goal	Define one bounded clinical task per interaction – a single, answerable question rather than “help me with this baby.”	“Generate a ranked differential for clinical deterioration in a 28-week infant, DOL 7, rising FiO2, CRP 48.”	One interaction, one goal. Broad, open-ended prompts return broad, unverifiable answers.
I – Input	Provide structured clinical context and anonymize rigorously before entering anything.	GA 28w, BW 950 g, DOL 7, FiO2 0.55 (was 0.38), CRP 48, WBC 18.4, temp 37.9℃, on ampicillin day 7. No name, DOB, or MRN.	Never enter identifiers into any LLM. In small NICUs, GA, BW, and DOL together may re-identify a patient – state what is unavailable (e.g., “echo not yet done”) rather than omit it.
R – Role	Tell the AI what role it is playing and explicitly limit its scope.	“Act as a neonatal evidence-retrieval assistant. Generate differentials and suggest investigations. Do not prescribe. Flag uncertainty and an evidence grade for every suggestion.”	Explicitly limiting the role reduces hallucination risk. An AI instructed not to prescribe is less likely to volunteer an uncaveated drug dose.
I – Iterative refinement	Do not accept the first response; interrogate it, and follow up with specific challenges.	After the AI ranks late-onset sepsis first, ask: “Can BPD be diagnosed at DOL 7?” It corrects itself: not applicable until 36 weeks PMA. Then ask for the evidence grade for adding Gram-negative cover.	If output contradicts clinical gestalt, do not default to the AI. Ask it to cite its source and state what it is uncertain about. Disagreement is information.
S – Safety verification	Verify every drug dose, antibiotic, and threshold against your formulary and unit protocol before acting.	Verify gentamicin dose (GA-specific, Neofax/BNFc). Check the local antibiogram for Gram-negative cover. Review the record for contraindications.	Hallucination red flags: a dose outside the GA-expected range; a guideline attributed to a non-existent or outdated source; internal inconsistency (e.g., ibuprofen despite documented oliguria); overconfident probability language without an evidence grade.
H – Human accountability	Document that a named clinician made the decision. The AI assisted; the clinician decided.	Record: “AI-assisted differential generation, GIRISH framework, [date]. Platform: [name]. Output verified against [formulary/guideline]. Decision by [Name], [Designation], [Time].”	In multidisciplinary settings, designate a GIRISH interaction lead. The AI has no accountability; the clinician does. Documentation should follow local institutional and medico-legal policy: as consumer LLMs are not approved medical devices, units should record AI assistance in line with their own information-governance and AI-use policies rather than adopting fixed wording uncritically (see “Regulatory context”).