Lessons from history guide the future of AI in medical decision-making

Neonatal intensive care units are increasingly becoming decision laboratories: clinicians must synthesize rapidly changing vital signs, lab results, imaging, and historical context—often under time pressure where errors can have irreversible consequences. In this environment, artificial intelligence (AI) is being positioned as a cognitive co-pilot, offering faster retrieval of relevant medical knowledge and the ability to detect data patterns that may be difficult for humans to notice in the moment.

In a new perspective published in Journal of Perinatology, Stevenson, Katz, Wong and colleagues argue that the core promise of AI in medicine is practical: narrowing the search space for clinicians by transforming scattered patient data into structured risk signals. Such systems can be trained to recognize associations between physiologic trajectories and outcomes, potentially supporting earlier intervention and more consistent decision pathways across shifts and institutions.

Yet the authors emphasize that decision-making is not purely an algorithmic exercise. For example, even when a model provides statistically accurate recommendations, the “best interest” of a specific infant may involve considerations that are not encoded in datasets—family values, goals of care, clinician judgment about uncertainty, and ethically grounded judgments about tradeoffs.

Importantly, the report highlights the limits of pattern recognition when applied to novel clinical contexts. AI can be vulnerable to dataset bias, variations in clinical practice, and distribution shifts—situations in which the patient data arriving today differs from the data the system learned from. In neonatal care, where conditions can evolve quickly and protocols differ, these gaps can matter.

There is also the question of accountability. The moral weight of clinical responsibility cannot be fully delegated to an analytical tool, because responsibility includes explaining decisions, responding to patient-specific circumstances, and ultimately accepting the ethical burden of action or inaction.

Rather than “replacing” physicians, the authors suggest AI should be understood as a scaffold for knowledge access and hypothesis generation—something that augments human reasoning while preserving human authority. In practice, this means careful validation, transparent performance reporting, and workflows that make AI suggestions auditable and contestable.

For critical-care medicine, the future likely lies in tightly coupled human–AI systems: models that flag risk, clinicians who integrate context and values, and institutions that continually monitor safety and equity. The take-home lesson is clear—AI may accelerate interpretation, but wisdom still belongs to the bedside.

AI in medical decision-making, with focus on neonatal intensive care.