Abstract
Recognition of pain experienced by immature and/or critically ill newborn infants in the Neonatal Intensive Care Unit (NICU) remains a challenge despite the use of objective scoring systems that depend on physiological and behavioural parameters. We consider there is a need to identify pain using only physiological data streams. Data were collected from three preterm male infants (gestational age 27.25±0.95 weeks, birth weight 941.25±189.31 grams). Heart rate (HR), mean arterial pressure (MAP), respiratory rate (RR), and blood oxygen saturation (SpO₂) were considered for the Neural Network (NN) Input Vector. The NN output was set to identify a noxious stimuli pattern (NSP). Events such as ‘vascular access’ and ‘reintubation’ statistically coincided with the NSP in 100% of cases. This study showed a positive relation between nociceptive response and non-invasive physiological response, verifying the NN as an accurate tool for clinical decision support systems.
Introduction
Recognizing pain in critically ill or immature newborn infants is a significant challenge in the NICU. Because these patients cannot communicate, clinicians currently rely on objective scoring systems that combine physiological data with behavioral observations, such as facial expressions or crying. However, behavioral assessment is subjective, intermittent, and difficult to perform consistently in a busy clinical setting. There is a pressing need for an automated, continuous system that can identify pain using only the physiological data streams already being monitored by bedside equipment.
Methodology
The researchers utilized a neural network to process a multi-dimensional input vector consisting of four standard vital signs: Heart Rate (HR), Mean Arterial Pressure (MAP), Respiratory Rate (RR), and Blood Oxygen Saturation (SpO₂). The study analyzed data from three preterm male infants.
The NN was trained to identify a Noxious Stimuli Pattern (NSP), defined by specific physiological thresholds: HR ≥ 160 bpm, MAP ≥ 55 mmHg, RR ≥ 40 breaths/min, and SpO₂ < 90%. To validate the network, the authors correlated the NN’s detections with "artifact events"—documented clinical procedures known to be stressful or painful, such as reintubation, vascular access, and routine care.
Results
The neural network successfully identified noxious events by analyzing the combined patterns of vital signs. A key finding was that multivariate analysis is essential; a single parameter like high heart rate (HR ≥ 160) occurred in only 24.6% of the documented event occurrences. In contrast, the NN-defined NSP coincided with 100% of high-intensity events like vascular access and reintubation.
For lower-intensity or routine events, the NSP coincided with 52.67% of "routine care" occurrences. This suggests that while the model is highly effective at identifying intense noxious stimuli, routine care may not always trigger a full physiological stress response detectable by these specific thresholds. The successful identification of these events is visualized in the 3D data mapping shown in Figure 1.
Discussion and Conclusion
This study provides a proof-of-concept for a clinical decision support system that could alert medical staff to infant distress in real-time without requiring manual scoring. By successfully correlating the NN's output with documented clinical actions, the authors verify a positive relationship between nociceptive response and non-invasive physiological data.
The primary limitation is the small sample size (n=3). Future work should involve larger, more diverse cohorts and potentially training the network on raw clinical labels of pain determined by experts to discover more subtle physiological signatures of distress. Nonetheless, the current NN proves to be a robust tool for identifying significant noxious events in the NICU environment.
Summary updated