Table 1. Five tiers of AI in neonatology and their bedside relevance.

Performance ranges are drawn from the cited source studies and are illustrative. Before any local use, calibration metrics (e.g., Brier score) and external validation in a comparable population should be sought in addition to AUROC.

AUROC: Area under the receiver operating characteristic curve; BPD: Bronchopulmonary dysplasia; CDSS: Clinical decision support system; CNN: Convolutional neural network; EEG: Electroencephalography; GIRISH: Goal, Input, Role, Iterative refinement, Safety verification, and Human accountability; HIE: Hypoxic-ischemic encephalopathy; LLM: Large language model; ML: Machine learning; MRI: Magnetic resonance imaging; RCT: Randomized controlled trial; RNN: Recurrent neural network; ROP: Retinopathy of prematurity.

AI tier	Examples	Neonatal application	Bedside relevance and current status
Tier 1: Rule-based CDSS	Protocol alerts; decision trees	Phototherapy thresholds; medication-dose alerts; ventilator alarm limits	Already built into monitors and pharmacy systems. Useful and validated – but it only knows the rules it was given.
Tier 2: Predictive ML	Random forest; XGBoost; logistic regression	Sepsis early-warning (AUROC 0.87–0.94); BPD risk (0.82–0.91); mortality scoring [12]	Being offered to NICUs now. Strong at risk ranking, but confirm it was validated in populations like yours; ask for calibration data, not just AUROC.
Tier 3: Deep learning	CNN; RNN; transformer models	ROP screening (>93% sensitivity) [7,8]; HIE MRI grading; EEG seizure detection (RCT-proven) [9]	The most mature AI in neonatology. i-ROP is externally validated across countries [7,8]; ANSeR improved seizure detection in an eight-center RCT [9]. These work with appropriate senior oversight.
Tier 4: Conversational AI/LLMs	ChatGPT; Gemini; Claude; Copilot; Perplexity; Consensus AI	Drug-dose queries; differentials; guideline summaries; parent-counseling drafts; discharge letters	The tier most commonly used informally at the bedside. High utility with a real risk of hallucination; the GIRISH framework is designed for this tier. Consensus AI searches peer-reviewed literature specifically and suits the evidence-retrieval role when a cited answer is required.
Tier 5: Agentic multimodal AI	Integrated vision-language and biosignal fusion	Real-time multiparameter monitoring; autonomous alert generation; agentic clinical support	Investigational and not yet validated for NICU deployment. Emerging governance frameworks should accompany any future clinical adoption.