Machine learning for identifying COPD subtypes from spirometry
Spirometry gives clinicians a fast view of lung function: how much air can be forced out in one second, how much can be exhaled overall, and whether the airways narrow after bronchodilator. Yet the same broad diagnosis of chronic obstructive pulmonary disease can conceal different combinations of airflow limitation, breathlessness and exacerbation risk. Using machine learning to identify subtypes of chronic obstructive pulmonary disease from spirometry data offers a route to uncover repeatable physiological patterns within routine pulmonary-function tests.
A useful result must go beyond sorting patients into severity labels. It needs to connect measurements with clinical outcomes, remain reliable across hospitals, and be understandable to the people receiving care. For an Australian audience, that also means working with varied primary and specialist services, community representation, and robust privacy and ethics safeguards.
Spirometry provides the clinical starting point
For COPD, persistent airflow obstruction is demonstrated by a post-bronchodilator FEV1/FVC below 0.70, interpreted alongside symptoms and exposure history. The ratio compares forced expiratory volume in one second (FEV1) with forced vital capacity (FVC); FEV1 also grades the severity of limitation. Spirometry does not directly measure emphysema, airway inflammation or the frequency of chest infections, so a model needs additional context to define a clinically meaningful subtype.
| Measure | What it contributes | What it cannot settle |
|---|---|---|
| FEV1/FVC | Identifies airflow obstruction when persistently reduced | Does not establish its cause or measure symptom burden |
| FEV1 | Grades the severity of airflow limitation | Depends on effort, technique and reference equations |
| FVC | Supports quality review and estimates exhaled volume | A low value has many causes and does not confirm restriction |
| PRISm pattern | Highlights preserved airflow ratio with reduced FEV1 | Is not itself COPD and requires clinical interpretation |
| Flow–volume loop | Shows patterns associated with obstruction or poor effort | Cannot distinguish all causes of obstruction |
Clinical assessment fills some gaps by adding symptoms, smoking exposure, exacerbation history and comorbidities. A model based on lung measurements alone can therefore offer a candidate physiological grouping, not a complete clinical phenotype or an automatic treatment recommendation.
Machine learning finds candidate disease groups
Supervised learning can predict outcomes such as future exacerbation from spirometry and other clinical variables; unsupervised learning searches for groups without assigning them in advance. The latter is especially relevant to subtype discovery, using clustering and related methods to detect combinations that may not be apparent to clinicians.
Algorithms such as gradient boosting, random forests and mixed-effects models can capture nonlinear relationships, but sophistication does not guarantee biological validity. Candidate groups should be stable, produce well-calibrated probabilities and be associated with meaningful outcomes such as treatment response, hospital attendance or mortality. Without those checks, a cluster may reflect coding habits rather than disease biology.
Data quality determines whether patterns generalise
Dataset construction often matters more than model choice. Researchers need consistent diagnostic definitions, device calibration records, bronchodilator timing, quality grades for acceptable manoeuvres, and variables recorded at the time of assessment. Age, smoking history, symptoms and comorbidities can help distinguish causes of reduced airflow, while careful splitting by patient and time prevents information leakage from future visits.
De-identified clinical and pulmonary-function records can be drawn from Queensland hospitals, primary-care spirometry services and relevant research cohorts. Australian electronic records are not uniform, so missing values, measurement units and documentation must be audited before training. Secure linkage may support longitudinal analysis, but consent, data-access agreements and ethics approvals should be established before data assembly.
Validation must cross service boundaries
A model trained at one Brisbane respiratory service may perform differently at another hospital, a regional clinic or a metropolitan general practice elsewhere in Australia. The country’s geography and separate state health systems create variation in equipment, populations and referral pathways. External validation across sites is therefore essential, with performance reported by device, service type and relevant population groups rather than as a single headline figure.
Prospective evaluation should test whether using the model changes decisions and patient outcomes, not merely whether it predicts an existing label. That stage needs predefined thresholds, a comparison with usual care, and assessment of unnecessary tests, missed cases and clinician workload.
Explainability and equity shape responsible use
Explainability should answer a practical question: which measurements drove a result? Summaries of feature contributions can show whether a model relies on FEV1, FVC and bronchodilator response or on a proxy such as age. Clinicians and patients need a clear explanation of uncertainty because no spirometry result should be treated as a diagnosis on its own.
Equity also requires attention to who is represented and who can benefit. Aboriginal and Torres Strait Islander communities experience substantial respiratory burden, yet spirometry evidence can be limited or shaped by longstanding gaps in access, housing and preventive care. Governance should include community representatives, culturally safe participation, and checks for unequal false-negative and false-positive rates. Personal and health information must comply with the Australian Privacy Act 1988 and applicable Queensland requirements.
Plain-language information on lung testing, symptoms and follow-up is as important as the algorithm. Resources such as breathing and lung health can support shared decisions and help patients understand why a result may lead to further assessment rather than an immediate label.
Linking patterns to clinical decisions
If validation identifies a stable group, the clinical payoff could be better prioritisation of further assessment, such as lung-volume testing, imaging or review of exacerbation risk. Spirometry alone should not dictate treatment. Decisions about inhalers, smoking cessation, vaccination and pulmonary rehabilitation still require clinical assessment, symptoms and patient preferences.
Implementation should fit Australian primary care and specialist workflows, including the Medicare Benefits Schedule pathway for diagnostic testing and referral. Decision-support tools need clear triggers, override options and audit trails. Evaluation should measure whether patients receive appropriate follow-up, not just whether an algorithm achieves a high discrimination score.
Partner for better respiratory outcomes
Progress depends on collaboration across disciplines and sectors. Brisbane Diamantina Health Partners connects research institutes, universities and health services in a Queensland health translation network, creating a natural setting to bring respiratory clinicians, data scientists, biostatisticians and consumers together around a shared study protocol. National and industry partnerships can add computing expertise, recruitment reach and implementation experience.
Clinical services, research teams and technology partners can contribute data, expertise or implementation sites, while community representatives and people living with COPD help shape outcomes, governance and communication. With NHMRC support, competitive research funding and appropriate industry engagement, this collaborative pathway can turn stronger lung-function data into earlier, more equitable care.
