Non Invasive Respiratory Support Ontology
- Non Invasive Respiratory Support (NIRS) Ontology is a formal, OWL-based framework that standardizes machine-interpretable concepts and relationships for modalities like HFNC, CPAP, and BiPAP.
- It supports semantic interoperability by unifying modality classifications, patient characteristics, therapy parameters, and clinical outcomes for enhanced acute care analysis.
- Evaluated through SPARQL queries and hypothetical patient scenarios, the ontology bridges bedside documentation and advanced computational reasoning in critical care.
Searching arXiv for recent work on the NIRS ontology and closely related non-invasive respiratory support papers. The Non Invasive Respiratory Support (NIRS) ontology is an ontological framework for acute care that formalizes machine-interpretable concepts, relationships, parameters, and outcomes associated with non-invasive respiratory support modalities such as HFNC, CPAP, and BiPAP/NIPPV. It was developed in Web Ontology Language (OWL) semantics with Protégé, extended with Semantic Web Rule Language (SWRL) rules for rule-based clinical reasoning, and evaluated through hypothetical patient scenarios, SPARQL queries, and data from the eICU Collaborative Research Database. Its stated purpose is to support knowledge representation in acute care, consistent documentation practices, integration into clinical data models, and analysis of NIRS outcomes (Islam et al., 26 Jul 2025).
1. Domain and rationale
The ontology addresses a combined clinical standardization and health informatics interoperability problem in acute care. In this setting, non-invasive respiratory support is not a single intervention but a heterogeneous domain spanning modalities, interfaces, therapy parameters, patient characteristics, and outcomes. The NIRS ontology is intended to unify these concepts into a formal representation so that they can be queried, reasoned over, and reused across clinical data environments (Islam et al., 26 Jul 2025).
The need for such formalization is evident in adjacent NIRS literature. In pediatric intensive care, HFNC is treated as a distinct non-invasive respiratory support modality whose initiation, observation window, and failure state are operationalized with explicit temporal anchors: “HFNC initiation,” “HFNC period,” and “HFNC trial,” with failure defined in that study as escalation to NIV or intubation within 24 hours of initiation (Pappy et al., 2021). In device-oriented work, non-invasive ventilation is further differentiated by support subclass and interface subclass, including CPAP, BiPAP, and helmet-based interface configurations with infection-control elements such as inlet and outlet viral filters and a PEEP valve (Khan et al., 2020). This heterogeneity provides a direct informatics justification for an ontology that can normalize terminology and encode relations explicitly rather than leaving them implicit in narrative documentation.
A further rationale arises from the distinction between therapeutic support, monitoring, and mechanistic physiology. The NIRS ontology is described as representing ventilation modalities, patient characteristics, therapy parameters, and outcomes (Islam et al., 26 Jul 2025). Closely related studies show that these categories interact with noninvasive monitoring modalities, EMR-derived prediction tasks, and respiratory mechanics models, but not all such concepts belong to the same ontological level. A plausible implication is that a formal NIRS ontology is needed not only to name therapies, but also to separate modality classes, event classes, parameter classes, and evidence provenance classes in a consistent way.
2. Semantic structure and representational design
The ontology was built with OWL semantics in Protégé and augmented with SWRL rules to enable rule-based clinical reasoning beyond hierarchical structures (Islam et al., 26 Jul 2025). In practical terms, this means that the ontology is not limited to subsumption hierarchies such as modality taxonomies, but can also express conditional knowledge that supports targeted inference.
The representational scale reported for the ontology is as follows:
| Ontology element | Reported value |
|---|---|
| Classes | 132 |
| Object properties | 12 |
| Data properties | 17 |
| Axioms | 882 |
| Annotations | 350 |
| Hypothetical patient clinical scenarios | 17 |
These counts indicate that the ontology is not merely a vocabulary list. The 882 axioms establish concept relationships, while the 350 annotations add descriptive definitions based on controlled vocabularies, thereby supporting terminology normalization and semantic clarity (Islam et al., 26 Jul 2025). The distinction between object properties and data properties is particularly important in a clinical ontology: object properties encode relations among entities such as therapies, patients, and outcomes, whereas data properties encode values associated with those entities.
The addition of SWRL rules is methodologically significant because acute-care reasoning often depends on combinations of patient state, therapy parameters, elapsed time, and outcomes rather than on class membership alone. The ontology’s architecture therefore combines declarative domain modeling with rule-based inference. This suggests a hybrid representation in which taxonomic structure and executable clinical logic coexist.
3. Clinical content model
The ontology is reported to formally represent domain-specific concepts including ventilation modalities, patient characteristics, therapy parameters, and outcomes (Islam et al., 26 Jul 2025). At minimum, this places HFNC, CPAP, and BiPAP/NIPPV within the ontology’s core modality layer, together with associated clinical descriptors and therapy-result relations.
Related NIRS studies clarify the kinds of entities that such a content model can accommodate. The pediatric HFNC prediction literature supplies ontology-ready event definitions with unusually explicit temporal granularity: “HFNC initiation” is defined as the start of HFNC treatment for a child not currently on HFNC, “HFNC period” as the 24 hours following a qualifying initiation, and “HFNC trial” as the analytic unit whose last HFNC period is the training target (Pappy et al., 2021). These constructs are valuable because they distinguish therapy start, observation window, analytic episode, and escalation outcome rather than collapsing them into a single notion of “HFNC use.”
Device-oriented work contributes a different layer of ontology content. A helmet-based non-invasive ventilator for COVID-19 is explicitly positioned as non-invasive respiratory support, more specifically non-invasive ventilation and more specifically noninvasive positive-pressure ventilation, with implemented support subclasses CPAP and BiPAP and an interface subclass of helmet-based interface (Khan et al., 2020). The same source enumerates ontology-ready hardware entities and relations, including blower, pressure sensor, differential pressure sensor, touch-screen, viral filters, PEEP valve, oxygen inlet, and buzzer, together with relations such as controls, drives, generates, monitors, detects, measures, and determines. This demonstrates that a NIRS ontology may need to represent not only therapies and outcomes but also interface types, device subsystems, and component-level functional relations.
Mechanistic respiratory modeling adds yet another content layer. In the theoretical model of the extremely preterm infant, CPAP is explicitly simulated as an increase in airway opening pressure from 0 to , while laryngeal braking is modeled as a 10-fold increase in expiratory upper airway resistance (Fix et al., 2018). The model introduces physiologic entities such as chest wall compliance, alveolar pressure , pleural pressure , end-expiratory lung volume, alveolar derecruitment, and time to failure. Although that work is not a formal ontology, it provides ontology-relevant distinctions between external support modalities, endogenous compensatory mechanisms, and pathophysiologic processes. A plausible implication is that a mature NIRS ontology can benefit from explicitly separating device-delivered support from intrinsic respiratory behaviors and from representing mechanistic variables as distinct from therapy labels.
4. Reasoning, querying, and validation
The NIRS ontology was evaluated by adding 17 hypothetical patient clinical scenarios and using SPARQL queries together with data from the eICU Collaborative Research Database to retrieve and test targeted inferences (Islam et al., 26 Jul 2025). This evaluation strategy is central to the ontology’s claimed utility: the ontology is not only a static schema but also a reasoning substrate whose adequacy can be tested against scenario-based logic.
According to the reported results, SPARQL queries successfully validated all test cases by retrieving appropriate patient outcomes. The paper gives a concrete example: a patient treated with HFNC for 2 hours due to acute respiratory failure may avoid endotracheal intubation (Islam et al., 26 Jul 2025). This example is important because it illustrates the ontology’s transition from concept organization to clinically meaningful inference. The target of inference is not merely a class label, but a possible outcome conditional on modality, indication, and therapy duration.
The use of SWRL rules “beyond hierarchical structures” also clarifies the type of knowledge encoded. Hierarchical ontologies can represent that HFNC is a non-invasive respiratory support modality, or that CPAP is a support subclass. They cannot by themselves express rule-like dependencies involving elapsed time, patient state, and likely outcomes. The reported SPARQL and eICU evaluations therefore serve as evidence that the ontology can support rule-based reasoning and therapy recommendations rather than only terminology harmonization (Islam et al., 26 Jul 2025).
From an informatics perspective, the validation strategy has two layers. The first is logical consistency and rule execution under hypothetical scenarios. The second is query-based retrieval against an external critical-care database. This suggests a design orientation toward operational computability rather than purely descriptive modeling.
5. Relationship to monitoring, prediction, and provenance
A NIRS ontology in acute care cannot be restricted to therapy names alone, because respiratory support increasingly interacts with continuous monitoring, predictive modeling, and multimodal physiological inference. Evidence for this broader requirement appears in work on electrical impedance tomography (EIT), where EIT is described as a noninvasive imaging modality that measures changes in regional bioimpedance and is used for monitoring regional ventilation distribution in critically ill patients (Strodthoff et al., 2020). That study is not primarily about HFNC, CPAP, or NIV/BiPAP, and its direct evidence is strongest for mechanically ventilated ICU patients, but it is highly relevant to ontology design for monitoring modality classes, physiological signal classes, and provenance relations.
The EIT study is especially important for differentiating observed, inferred, and reference variables. It explicitly distinguishes directly observed signals such as airway pressure , flow , volume , arterial blood pressure , and esophageal pressure from model-inferred outputs such as predicted absolute volume, predicted absolute flow, predicted normalized airway pressure, predicted normalized arterial blood pressure, and predicted absolute transpulmonary pressure (Strodthoff et al., 2020). It also provides the explicit physiological definition
for transpulmonary pressure. This distinction is ontology-critical because a clinical knowledge graph should not represent a parameter value without encoding whether it came from direct device measurement, invasive reference measurement, or model-based estimation.
The same issue appears in HFNC risk prediction from EMR data. In that study, an LSTM model generated a continuous prediction of HFNC failure, with prediction updated whenever a new measurement became available, and the outcome of interest was escalation to NIV or intubation within 24 hours of HFNC initiation (Pappy et al., 2021). Here the ontology-relevant entities include EMR-derived physiologic observations, laboratory results, intervention records, prediction events, and outcome labels. A plausible implication is that a NIRS ontology intended for advanced analysis should encode provenance not only for therapies and outcomes but also for predictions, timestamps, inclusion criteria, and study-specific outcome definitions.
Taken together, these adjacent studies suggest that the NIRS ontology can function as a bridge between bedside therapy representation and computational phenotyping. The direct claim established in the ontology paper is that it provides a foundation for consistent documentation practices, integration into clinical data models, and advanced analysis of NIRS outcomes (Islam et al., 26 Jul 2025). The monitoring and prediction literature indicates what those integrations may need to represent in practice: noninvasive monitoring modalities, hybrid multimodal inputs, time-series predictions, and explicit measured-versus-estimated parameter status.
6. Scope boundaries, interpretive cautions, and significance
The NIRS ontology unifies NIRS concepts into an ontological framework and demonstrates applicability through hypothetical patient scenarios and alignment with standardized vocabularies (Islam et al., 26 Jul 2025). That contribution is substantial, but its scope should be delimited carefully. The reported evaluation concerns logical reasoning with 17 hypothetical clinical scenarios and SPARQL-based retrieval, not direct prospective bedside deployment. The ontology therefore establishes representational and inferential feasibility within the stated evaluation frame.
Adjacent literature also shows why ontological scope discipline matters. The HFNC prediction study provides a pediatric ICU-specific operational definition of HFNC failure, and its own synthesis explicitly notes that such definitions are study operational definitions in a specific pediatric ICU population rather than universal clinical definitions (Pappy et al., 2021). The EIT study explicitly states that it is not primarily about named noninvasive respiratory support modalities such as HFNC, CPAP, or NIV/BiPAP, even though it is highly relevant for noninvasive monitoring and estimated-parameter provenance (Strodthoff et al., 2020). The preterm infant mechanics model is a theoretical open-loop model whose relations are best labeled as mechanistically modeled, simulation-supported, and hypothesis-generating rather than as direct clinical efficacy evidence (Fix et al., 2018). These distinctions are not peripheral; they are precisely the kinds of epistemic boundaries that a rigorous ontology should preserve.
A common misconception is to treat a NIRS ontology as interchangeable with a device taxonomy, a predictive model, or a physiology simulator. The evidence does not support that equivalence. The ontology described in (Islam et al., 26 Jul 2025) is a formal knowledge representation artifact; device papers contribute component and interface relations, predictive studies contribute temporal event and outcome definitions, and monitoring or mechanics studies contribute signal provenance and physiologic parameter semantics. A plausible implication is that the principal long-term value of the NIRS ontology lies in its ability to align these heterogeneous evidence types within a shared semantic structure while preserving their different validation statuses.
In acute care informatics, that alignment is consequential. It supports consistent documentation, standardized concept use, integration into clinical data models, and more advanced analysis of NIRS outcomes (Islam et al., 26 Jul 2025). More broadly, it provides a basis for representing not only what support was delivered, but also to whom, by which interface, under what parameters, with what temporal logic, and with what measured, inferred, or predicted consequences.
