Introduction: Why We Need to Rethink Pain Measurement in the NICU
Current pain assessment tools for newborns were not specifically designed for extremely preterm infants, defined as those born before 28 weeks gestational age. This group represents some of the most fragile patients in neonatal care, making accurate pain measurement essential for guiding treatment and improving outcomes.
Extremely preterm infants often exhibit subtle and atypical pain responses due to neurodevelopmental immaturity. Conditions such as sepsis, sedation, or severe brain injury—like intraventricular hemorrhage (IVH)—can mask or alter typical pain behaviors. Consequently, behavioral pain scales developed for term or near-term infants might not provide reliable assessments in this population.
This uncertainty raises a key question: Are current pain scales systematically biased depending on gestational age or clinical condition? To explore this, we use a scenario-based validation approach. Rather than evaluating whether a scale “works” overall, we investigate its performance across distinct clinical contexts. Using modern psychometric techniques, we assess whether factors such as sedation, oxygen dependency, or sepsis alter the expression and measurement of pain, thereby revealing potential shortcomings of existing tools.
From this study, we expect to identify specific pain scale items that show sensitivity to clinical context, known as Differential Item Functioning (DIF). Items exhibiting such bias may require adjustment or stratification to ensure fair assessment. We also anticipate finding items with low discrimination ability, typically below 0.5, which may need revision or removal due to poor differentiation between pain levels. Additionally, we will examine temporal changes in item behavior—referred to as temporal drift—to determine if time-specific recalibration is necessary. Finally, we expect to observe divergent pain score trajectories across clinical subgroups, supporting the need for subgroup-specific interpretation.
Our study fills a critical gap by rigorously testing pain scales under real-world conditions specific to extremely preterm infants. This work lays the foundation for more accurate, fair, and clinically useful pain measurement in the NICU.
Methods
Study Population and Design
Our study focuses on extremely preterm infants—those born before 28 weeks gestational age—admitted to the Neonatal Intensive Care Unit (NICU). We employ a longitudinal, observational design to collect repeated behavioral pain assessments at multiple time points throughout their NICU stay. This approach allows us to observe how pain expression evolves over time.
Defining Clinical Scenarios: The Subgroups
To better understand how pain measurement performs amid clinical complexity, we categorize infants into five clinical subgroups. These groups are defined by conditions known to influence pain behavior:
| Label | Definition | Expected Measurement Effect |
|---|---|---|
| G1 | Stable: No comorbidities or infection | Reference group |
| G2 | Sepsis: Lab-confirmed sepsis with hemodynamic instability | Blunted behavioral expression |
| G3 | IVH: Severe intraventricular hemorrhage (Grade III-IV) | Suppressed pain expression |
| G4 | Sedation: Opioids administered for more than 12 hours | Blunted behavioral and physiological signs |
| G5 | Respiratory insufficiency: Oxygen dependency for more than 14 days | Elevated physiological baseline |
It is important to note that these subgroups are not mutually exclusive. Infants frequently belong to more than one subgroup at the same time, reflecting the real-world clinical challenges faced by caregivers and adding complexity to pain measurement.
Addressing Overlapping Clinical Profiles
Traditional grouping methods in Item Response Theory (IRT) assume mutually exclusive groups, which is insufficient given that infants often experience multiple concurrent conditions. To address this, we apply latent class and mixture modeling to identify hidden patient profiles beyond fixed categories. Multidimensional IRT models allow us to capture multiple latent traits affecting pain behavior simultaneously. Additionally, covariates and interaction terms enable us to model how overlapping clinical conditions combine to affect individual item responses. This flexible framework captures clinical heterogeneity more accurately and enhances the validity of pain measurement.
Scenario-Based Analysis Approach
We analyze raw ordinal pain scores rated on a scale from 1 to 5, organized in a long-format dataset that includes subject ID, time point, and individual item responses. Groupings are derived from clinical information extracted from electronic medical records. We assume that missing data are Missing At Random (MAR), allowing the use of standard imputation techniques.
The analysis proceeds through a three-step modeling strategy. First, a baseline IRT model, specifically a Two-Parameter Logistic Model, evaluates each item’s ability to discriminate between pain levels and determines its threshold (difficulty) across the entire sample, providing a general picture of item performance. Next, a multi-group IRT model tests for Differential Item Functioning across clinical subgroups, allowing item properties to vary by condition. For example, sedation may blunt visible behavioral pain signs without affecting the infant’s underlying pain experience, potentially biasing item responses if not accounted for. Identifying DIF is therefore critical to ensuring measurement accuracy and fairness. Finally, a longitudinal IRT model investigates how item properties shift over time, using growth modeling to detect changes in item functioning between early assessments (e.g., 24 hours after birth) and later time points (e.g., seven days). Understanding these temporal drifts guides decisions about whether items require time-specific recalibration.
Tools and Software
All analyses are performed using R (version 4.4.0), utilizing several specialized packages. The mirt package supports IRT modeling, lme4 is used for longitudinal mixed models, lavaan facilitates latent variable analysis, and brms enables Bayesian modeling extensions. The tidyverse suite is employed for efficient data processing and management, ensuring reproducible workflows.
Handling Multicenter Variation and Observer Training
Data collection spans multiple NICUs, introducing variability in assessment practices, sedation protocols, and observer expertise. To account for this heterogeneity, we include center as a random effect in growth models and test for center-level Differential Item Functioning. Additionally, standardized observer training modules are implemented across sites to minimize variability in pain assessment. While multicenter results are still forthcoming, our modeling framework anticipates and adjusts for these differences to ensure robustness.
Future Directions for Scale Development
Our findings will guide several improvements to the pain scale. One approach involves developing a modular design, combining a core set of items with context-specific submodules tailored to different clinical scenarios. We will consider group-specific thresholds for pain cutoffs that adjust for clinical conditions, as well as differential item weighting informed by subgroup test information. Bayesian scaling methods incorporating prior clinical knowledge may further enhance measurement precision.
Expected Outcomes and Interpretation
We anticipate that this work will result in a pain scale that is both more accurate and fair for extremely preterm infants. Adjusting for context-sensitive items will improve measurement validity. Recognizing temporal dynamics and subgroup-specific pain trajectories will enable clinicians to interpret scores more meaningfully and tailor pain management strategies accordingly.
Future Vision: Towards a Condition-Sensitive Pain Scale
Ultimately, we aim to develop a stratified, adaptive pain assessment tool that acknowledges the multidimensional and dynamic nature of pain in extremely preterm infants. This tool would combine modular components, context-aware scoring, and advanced psychometric methods to provide reliable and personalized pain measurement in the NICU.
Conclusion
Accurate pain assessment in extremely preterm infants remains an urgent, unmet need. By integrating sophisticated psychometric methods with scenario-based validation, our work reveals biases in current tools and proposes tailored solutions. Improved pain measurement will empower clinicians to reliably detect pain even when behavioral signs are suppressed, enabling personalized analgesic interventions and effective monitoring of pain trajectories. Ultimately, this will support better decision-making and improve outcomes for the NICU’s most vulnerable patients.
