AI in PRImary Care spirometry for diagnosis of Lung Disease ((APRIL)
Research type
Research Study
Full title
Real-world evaluation of an Artificial-Intelligence support software (ArtiQ.Spiro) in Primary Care spirometry pathways for the detection of Lung Disease.
IRAS ID
324175
Contact name
William Man
Contact email
Sponsor organisation
Royal Brompton & Harefield Hospitals, Guy’s and St Thomas’ NHS Foundation Trust
Clinicaltrials.gov Identifier
Duration of Study in the UK
1 years, 0 months, 1 days
Research summary
Respiratory disease is the third biggest cause of death in the United Kingdom (UK) with 20% of all deaths attributed to respiratory disease and lung cancer. Respiratory disease is also one of the commonest causes of emergency hospital admissions (>700,000 per year) and winter bed pressures faced by the National Health Service (NHS). Prior to COVID-19, lung conditions were already costing the NHS around £9.9 billion each year. There are strong links between lung disease, social deprivation and health inequalities with environmental drivers of lung disease such as smoking, outdoor pollution, housing and occupational hazards more prevalent in socio-economically deprived communities. The NHS Long Term Plan, which outlines the actions to improve the nation’s health outcomes, identified respiratory disease as a national clinical priority.
Spirometry is an essential diagnostic procedure recommended for the diagnosis and monitoring of chronic obstructive pulmonary disease (COPD) and asthma, the commonest long-term respiratory conditions. However, spirometry provision in primary care is suboptimal. Only 13.4% of spirometry performed in primary care meet international criteria, more than 40% fail at least one quality criteria, there are low levels of confidence in identifying technical errors or interpreting spirometry and there are poor levels of agreement in spirometry interpretation between primary care and specialist respiratory physicians. Important consequences include under-diagnosis, mis-diagnosis and unnecessary referral to secondary care.
ArtiQ.Spiro is a decision support software that combines two sub-components – one focussing on quality assessment (ArtiQ.QC), and one on spirometry interpretation (ArtiQ.PFT). It is intended to be used as an adjunct to spirometry to assist with the grading of spirometry quality and the interpretation of spirometry by providing the probability of six disease / or no disease categories. An example is shown in Appendix A.
Quality Assessment
International guidelines comprise objective criteria for ensuring good quality spirometry but also partly rely on the subjective opinion of technicians. Even in cohorts of trained and accredited staff, there is inter-rater variability when assessing spirometry quality. ArtiQ.QC leverages deep learning methods to perform the subjective elements of spirometry quality assessment, as well as implementing the criteria. The AI component of the software is trained to mimic the subjective, visual inspection of data usually performed by technicians. Spirometry measurements from the National Health and Nutritional Examination Survey (NHANES 2011-12) were used as training data. Each measurement was assessed by multiple experienced technicians to determine whether the data quality was acceptable; in total the NHANES dataset contained 36,873 measurements, of which 54% were acceptable. Over 29,000 measurements were used to train a convolutional neural network, with the remainder used for validation/testing.Spirometry Interpretation
ArtiQ.PFT is the only CE-marked clinical decision support software incorporating AI in the field of respiratory physiology diagnostics. It uses a random forest machine learning model to estimate the probability of disease from the eight commonest categories (7 diseases + normal lung function) detectable with lung function testing. 1430 cases, where the final diagnosis was known, were used to train the random forest model. For use in primary care settings, ArtiQ.PFT has been adapted to identify six categories – Asthma, COPD, interstitial lung disease (ILD), normal lung function, other obstructive disease, and other unidentifiable disease – using the same training set as for ArtiQ.PFT (1430 cases of spirometry-only data). The rationale for reducing the number of categories to identify is based on the reduced lung function dataset available when working in a primary care environment compared to a hospital-based pulmonary function testing laboratory.A possible limitation to the ArtiQ.Spiro is that the training set for interpretation only included patients undergoing full lung tests in hospital lung function laboratories where the lung function skillset of the practitioner is likely to be higher than seen in primary care. Furthermore, the patient population in the training set was almost exclusively Caucasian.
The NHS Long Term Plan (LTP) has identified quality assured spirometry as a clinical priority through investment into training more community staff to perform and interpret spirometry more accurately. However there are significant workforce issues and training of available staff takes time. There is likely to be variability and inequality in workforce and training across the NHS. Disagreement over spirometry interpretation can still occur between trained staff.The findings of this proposed feasibility study will inform the design of a multicentre randomised controlled trial to assess the real-world implementation and impact of an Artificial Intelligence-powered quality control and interpretative software (ArtiQ.Spiro) on the diagnostic accuracy, care processes, patient and health economic outcomes.
REC name
North West - Greater Manchester South Research Ethics Committee
REC reference
23/NW/0017
Date of REC Opinion
6 Mar 2023
REC opinion
Further Information Favourable Opinion