Neo-InstructTime: Editing Deceleration Events in Neonatal Vital Signs Using Natural Language

Decelerations in preterm infant vital signs (heart rate and oxygen saturation) are critical biomarkers of adverse outcomes in the NICU. Synthesizing such events can address data scarcity and support counterfactual reasoning on their pathophysiological signatures. Although recent advances in time series editing allow fine-grained modification of time series toward target conditions while preserving others, existing methods rely on rigid feature vectors and lack control over editing strength. We propose InstructTime, the first instruction-based time series editor, which specifies edits in natural language and enables controllable editing strength. Conditioning on free-form text enables incorporating nuanced clinical details such as demographics, comorbidities, interventions, and individualized information from medical free-text data. We evaluate InstructTime on editing decelerations in neonatal vital signs through four research questions: (1) performance comparison to state-of-the-art editors; (2) robustness under rare-event prevalence; (3) the effect of individualized clinical context on editing quality; and (4) exploratory insights from counterfactual edits. Our results show that InstructTime can synthesize realistic decelerations, maintain robust quality under data scarcity, and provide exploratory insights into these clinically significant deceleration events in neonatal vital signs.