Artificial intelligence (AI) Across Respiratory Care: Inhaler Systems, AERD Identification and Paediatric Asthma Risk Prediction

AI in respiratory medicine is no longer confined to a single task. Recent work spans the inhaler itself, the molecular signature of aspirin-exacerbated respiratory disease (AERD), and the emergency department pathway for children with asthma. The evidence is encouraging in places, but its practical message is clear: useful tools must fit the clinical problem, the data and the workflow.

Key points

Why this matters

Respiratory care is shaped by variation at every level. A well-designed medicine may still fail to reach the target airway if the aerosol, device and inhalation pattern do not align. A clinically important asthma subtype may remain unrecognised when confirmation requires a time-consuming challenge. A child with asthma may leave the ED without being identified for scarce specialist support, despite later returning with further acute-care needs.

The supplied sources examine these distinct problems through machine learning (ML). A narrative review considers inhaler drug–device combination systems from formulation development to manufacturing and smart monitoring. A primary AERD study investigates whether a nasal brushing can provide a molecular signal for a difficult diagnosis. AIRE-KIDS tests whether routinely collected information can support risk-aware paediatric asthma referral after an ED visit.

These are not competing versions of one technology. They are different applications of predictive methods to different parts of the respiratory pathway. Their shared lesson is that technical performance becomes clinically meaningful only when the data are representative, the output is understandable, and the next clinical action is appropriate.

Evidence at a glance

Study Setting/population Clinical question AI method Endpoint Main result Key limitation Practical relevance
Pulmonary drug-delivery review Evidence across inhaler formulation, devices, manufacturing and sensor-enabled monitoring Can AI improve inhaled-product performance across its lifecycle? Artificial neural networks, random forest, support vector machines, deep learning, digital twins and computational fluid dynamics-ML Aerosol performance and process measures Applications reported for emitted dose, fine-particle fraction, deposition and process optimisation Underlying evidence is heterogeneous; many systems are not mature for routine use May guide future development and monitoring strategies
AERD nasal mRNA study Two cohorts from National Jewish Health and Scripps Health Can a nasal expression signature identify AERD? 34-gene logistic regression/gradient-boosting classifier AERD classification 93% training accuracy; 83% independent-validation accuracy Small cohorts; validation task differed from training task Could become an adjunctive test after prospective validation
AIRE-KIDS Children with an asthma index ED visit at the Children’s Hospital of Eastern Ontario Can future asthma acute-care use be predicted using available data? LightGBM, XGBoost and fine-tuned open-source large language models One-year repeat ED use or admission; hospital admission Six-feature LightGBM model: area under the curve 0.712, F1 0.510 in temporal validation Single-centre retrospective evidence May support future targeted specialist referral

Study-by-study clinical interpretation

1. AI-enabled pulmonary drug delivery: an engineering-to-clinic continuum

The inhaler review is broad by design. It argues that inhaled medicines should be considered as integrated drug–device combination products, because formulation behaviour, device design and patient use jointly determine what dose is emitted and where it deposits. This matters for asthma, chronic obstructive pulmonary disease, cystic fibrosis, pulmonary hypertension and respiratory infections, all of which are identified in the review as inhalation-related treatment contexts.

The review describes the central delivery challenge succinctly: lung anatomy supports rapid absorption, but consistent deposition is difficult. It estimates that historically only 10–20% of the dose delivered from an inhaler reaches the lungs. Particle size, inspiratory flow, device resistance, powder cohesion, plume characteristics, airway anatomy and technique all contribute.

The technical targets help explain where AI may be useful. Mass median aerodynamic diameter indicates the midpoint of aerosol particle mass. The review describes particles larger than roughly 5 micrometres as more likely to deposit in the mouth or throat, while particles smaller than roughly 1 micrometre can be exhaled. Fine-particle fraction is the percentage of emitted dose at or below 5 micrometres. These measures, alongside emitted dose and deposition efficiency, provide model outputs that may be predicted from formulation and device inputs.

The review’s most practical point is that algorithm selection should follow the problem rather than fashion. Artificial neural networks can model complex nonlinear relationships but require well-organised data and have limited interpretability. Random forest models may identify important features and resist overfitting. Support vector machines may suit smaller datasets. Deep learning can extract features from images and sensor streams, but also depends on large annotated datasets. Computational fluid dynamics-ML models may simulate aerosol deposition and manufacturing processes but remain computationally intensive and need independent validation.

Examples in the review show why this is attractive. Muddle and colleagues reported R² values of 0.92 or higher when a 9-4-1 neural network predicted dry powder inhaler fine-particle fraction for salmeterol and salbutamol. Jeong and colleagues used scanning electron microscopy information from arformoterol–lactose formulations, reporting that multilayer perceptrons predicted emitted or delivered dose well, while convolutional neural networks better predicted fine-particle fraction and fine-particle dose. These findings support predictive modelling of formulation behaviour; they do not, by themselves, demonstrate improved clinical control.

The review also describes smart inhalers as a route towards capturing real-world use. Sensors may record inspiratory flow, timing, orientation, adherence and technique errors. The RS01X smart dry powder inhaler was reported to agree strongly with an external standard for inspiratory volume and peak inspiratory flow, with intraclass correlation coefficients above 0.95. This is an important technical validation, but not evidence that sensor feedback consistently improves outcomes over the long term.

Manufacturing is another proposed application. The review describes quality-by-design and digital-twin approaches for jet milling, spray drying, blending and capsule filling. In cited scale-up work, mass median aerodynamic diameter shifted by around 0.3–0.5 micrometres and fine-particle fraction fell by around 5–10 percentage points when dry powder inhaler manufacture moved from laboratory to commercial scale. The finding illustrates why predictive tools could be useful before full-scale production. It does not show that any specific AI platform resolves scale-up variation.

2. AERD: a nasal sample as a possible diagnostic signal

The AERD study takes AI closer to a specific diagnostic decision. AERD combines asthma, severe eosinophilic nasal polyps and reactions to aspirin or related cyclo-oxygenase-1-inhibiting non-steroidal anti-inflammatory drugs. The authors cite an estimated prevalence of around 7% among all people with asthma and 15% among those with severe asthma. Despite this, the condition may be under-recognised.

The reference diagnostic approach—oral non-steroidal anti-inflammatory drug challenge—can be effective, but it is resource-intensive and involves provoking the reaction of interest. The investigators therefore asked whether messenger RNA from an inferior-turbinate nasal brushing could provide an informative, lower-risk molecular profile.

The study used 71 participants with AERD and 57 without. The National Jewish Health training cohort compared AERD with healthy controls. The Scripps Health validation cohort compared AERD with aspirin-tolerant asthma. This independent validation is a strength, but the cohorts were not identical: the validation comparison was narrower and clinically harder.

After adjustment for sex, race, age, body mass index and corticosteroid use, 56 genes were differentially expressed between AERD and healthy controls. When AERD was compared with aspirin-tolerant asthma, 21 genes remained significant after stringent correction. The reported genes included markers associated with eosinophils, mast cells, type 2 inflammation and epithelial barrier function. Charcot-Leyden crystal protein showed approximately 3.16-fold higher expression in AERD than in healthy controls.

The 34-gene classifier achieved 93% accuracy in training, with an area under the receiver-operating characteristic curve of 0.96. Applied without retraining to the independent cohort, it achieved 83% accuracy and an area under the curve of 0.88. This retained performance is notable because validation involved aspirin-tolerant asthma rather than healthy controls. It is also appropriately uncertain: the validation cohort was smaller, and the ideal comparator of aspirin-tolerant asthma with chronic rhinosinusitis with nasal polyps was not fully represented.

The model’s biology was not treated as a substitute for validation, but it adds interpretive context. Interleukin 1 receptor-like 1, the interleukin-33 receptor, and Charcot-Leyden crystal protein were among the relevant signals. The authors describe these as consistent with established inflammatory pathways in AERD. They also report unexpected lower expression of adaptive immune-related transcripts and FCER1A in nasal epithelial brushing samples. Their proposed explanation—that this may reflect the sampled tissue compartment—remains a hypothesis requiring further study.

The cohort findings reinforce clinical complexity. AERD participants more commonly reported nasal polyps, nasal surgery and alcohol-triggered respiratory symptoms than those with aspirin-tolerant asthma. The study reports alcohol-triggered respiratory symptoms in 66% of AERD participants versus 17% of aspirin-tolerant asthma participants. These features may inform clinical suspicion, but they do not establish a diagnostic rule.

3. AIRE-KIDS: prediction linked to an operational decision

AIRE-KIDS focuses on what happens after a paediatric asthma ED visit. The stated objective is not prediction in isolation, but more targeted access to specialised asthma services when capacity is limited. The authors note that up to 25% of children with an asthma ED visit may return within one year, while not every child can be enrolled in specialised care.

The study’s design used temporally distinct cohorts: 2,716 children in the pre-COVID-19 training dataset and 1,237 children in the post-COVID-19 validation dataset. This is more demanding than a random split because it tests performance after changes in care patterns and patient populations. The primary outcome included repeat asthma ED attendance or admission, plus receipt of specialised asthma care; a sensitivity analysis evaluated a stricter outcome limited to acute-care events.

The models used 68 features derived from structured records, environmental data and the Ontario Marginalization Index. The comparison between boosted-tree methods and large language models is especially informative because language models have attracted interest even when data are already structured. Patient features were converted to text for the language-model arm, and open-source models were used locally because institutional privacy rules prohibited external data transfer.

LightGBM consistently performed best. For repeat ED use under the primary outcome, the final six-feature model used prior asthma ED attendance, Canadian Triage Acuity Scale score, medical complexity, food allergy, prior non-asthma respiratory ED attendance and age. It achieved validation area under the curve 0.712 and F1 0.510 at a 0.250 threshold. The local rules-based best-practice alert achieved F1 0.334.

For future admission, the final five-feature model achieved validation area under the curve 0.65 and F1 0.375, compared with F1 0.313 for the existing alert. In sensitivity analysis using only repeat acute-care events, the ED model had lower performance, with area under the curve 0.644 and F1 0.407. The authors found broadly similar influential features.

The emphasis on F1 is central. In a limited-capacity programme, a useful model must balance identifying children likely to need support with avoiding an unmanageably large referral group. AIRE-KIDS therefore offers a concrete example of performance measurement matched to an implementation context. It does not yet show that acting on predictions improves care or reduces subsequent acute-care use.

Why the studies should or should not be compared directly

These publications share a respiratory focus but assess different evidence questions. The inhaler paper is a narrative review with endpoints such as particle-size distribution, emitted dose and manufacturing consistency. The AERD study evaluates diagnostic classification against challenge-confirmed disease status. AIRE-KIDS evaluates future acute-care risk and specialist-care receipt in a defined paediatric service setting.

Their metrics must therefore remain separate. A strong R² for aerosol prediction does not equate to diagnostic accuracy. An area under the curve for AERD classification cannot be interpreted against an F1 score selected for a capacity-limited referral pathway. The studies also differ in their data sources: device and formulation parameters, nasal transcriptomics, and structured electronic health data linked to environmental and neighbourhood measures.

There are, however, meaningful parallels. All three sources recognise that model performance is conditional on the data and setting. The AERD investigators tested an independent institution. AIRE-KIDS used a later temporal cohort. The inhaler review identifies independent validation, standardisation and explainability as recurring needs. Each source also avoids presenting AI as an autonomous replacement for clinical expertise.

What this means for clinical practice

For respiratory specialists, the immediate value of these papers lies in their disciplined framing of what AI might do. In pulmonary delivery, it may make product development and process optimisation more efficient, and smart sensors may identify inhalation patterns or errors. Translation depends on reliable data capture and whether feedback can be incorporated into care without increasing inequity or workload.

In AERD, the nasal classifier offers a plausible future adjunct when clinical history is ambiguous or challenge is unsuitable. The authors propose that a targeted assay could eventually provide an interpretable probability score. The current study does not establish replacement of supervised oral challenge, specialist assessment or clinical judgement.

In paediatric asthma, the AIRE-KIDS model may eventually help ED teams direct limited specialist resources. Its small final feature set is potentially attractive for workflow integration. Yet the authors appropriately propose clinician acceptability work and prospective silent validation before any live decision-support use.

What remains uncertain

The pulmonary-delivery review highlights variable digital engagement, narrow training datasets, sensor and connectivity problems, cost, cybersecurity and possible inequities for populations less represented in connected data. Many described systems remain at proof-of-concept or early-validation stages, and sustained clinical benefit from digital inhalers remains limited in the cited evidence.

The AERD study requires larger prospective multicentre validation, broader population calibration and standardised sampling and laboratory workflows. It did not fully account for treatment effects on gene expression. Further work is needed to determine how a targeted panel would perform in the most challenging overlapping presentations and how its result should influence clinical assessment.

AIRE-KIDS remains single-centre and retrospective. Relevant factors such as controller adherence and household smoke exposure were not reliably available in structured records. Local outcome capture may be incomplete for care obtained elsewhere. Most importantly, prospective research is needed to determine whether use of the model changes referral decisions or subsequent clinical outcomes.

Conclusion

Across inhaler design, AERD classification and paediatric asthma triage, these studies show a field moving towards more tailored respiratory care while retaining important limits. The pulmonary-delivery review identifies potential roles for AI in prediction and monitoring, but also substantial practical barriers. The AERD study provides independently validated evidence that nasal epithelial gene-expression patterns may help identify a difficult subtype. AIRE-KIDS shows that a focused boosted-tree model can outperform tested language models for structured paediatric ED data. The most compelling common message is not technological replacement, but careful augmentation: AI tools may earn a role when they are validated beyond their development setting, understandable within clinical workflows and demonstrably useful for patients and services.

Educational content only. It does not replace clinical assessment, current guidelines, or patient-specific professional advice.