What is artificial intelligence and how might it be applicable to respiratory medicine?
Artificial intelligence (AI) describes the ability of a computer software programme to undertake tasks that would typically require human (or organic) intelligence. The application of artificial intelligence software to enhance or complement current clinical pathways in respiratory medicine is a rapidly changing landscape.
There are a broad range of AI techniques including machine learning where computer algorithms learn and adapt automatically from patterns identified in sample data to accomplish specific tasks. Deep learning is a type of machine learning that combines layers of algorithms or ‘neurons’ into artificial neural networks mimicking the human brain to perform higher level tasks. A convolutional neural network describes the application of neural networks to image-based data, typically a CT scan in the case of respiratory medicine.
Broadly speaking, machine learning can be thought of as supervised or unsupervised. Supervised learning takes pre-labelled input data to ‘learn’ patterns which are then applied to new unlabelled data. An unsupervised learning approach takes unlabelled data to find novel patterns or clusters that have not previously been identified which can be mapped to an outcome.
These different software approaches have been tailored to provide solutions to several common fields of respiratory medicine.
How is AI currently being used in the diagnosis and treatment of respiratory disease?
Artificial intelligence is being applied across respiratory medicine including pulmonary physiology, pathology, biomarker discovery and thoracic radiology. AI algorithms are being developed to predict the malignancy risk of pulmonary nodules, triage chest radiographs, identify pulmonary emboli and predict and monitor disease outcomes in interstitial lung disease. However, there are several hurdles which need to be overcome before these technologies are established in routine clinical use and at present AI primarily exists as a research tool.
Lung cancer screening is one field whereby the application of AI could not only enhance reporting but could be crucial to large scale rollout. The availability of specialist thoracic radiologists to report significant numbers of thoracic CT scans is a recognised hurdle to screening rollout in the UK. The Royal College of Radiologists are currently predicting a 40% shortfall in radiology workforce by 2028.[1] The use of AI to streamline interpretation and reporting of screening CTs could be part of the solution. Algorithms have been developed which can out-perform the current Brock malignancy risk model for pulmonary nodules [2]. These algorithms may avoid unnecessary follow up of benign nodules reducing workload and enhancing patient experience. AI has also been postulated as a method of performing non-invasive tissue typing in lung cancer, a so called “virtual biopsy”. A team from Imperial College London utilised metabolomic and imaging data in a deep learning model to generate an algorithm that could make a tissue diagnosis based on imaging alone[3].
Artificial intelligence may also have a role in acute respiratory pathways such as the diagnosis of venous thromboembolism. A study at the Royal United Hospitals NHS Foundation Trust is examining the role of an AI solution to pulmonary embolus detection on CTPA (NCT06093217). The prospective study will examine the impact of the AIDOC algorithm in enhancing the detection of acute pulmonary emboli. The study will also assess automated risk stratification with AI assessment of right ventricular strain.
What are potential applications for AI outside of thoracic imaging in respiratory medicine?
The interpretation of pulmonary physiology requires expertise to ensure accurate diagnosis and appropriate treatment. The NHS has aimed to increase availability of spirometry to patients in the community and test interpretation could be a novel application of AI. A current randomised control trial is evaluating primary care clinician interpretation of spirometry with or without an AI decision support software[4]. The researchers are examining whether the ArtiQ.Spiro AI software can improve quality assessment and diagnostic accuracy of spirometry interpretation.
AI may also be coming after the humble stethoscope. Studies are ongoing investigating whether computerised lung sound analysis can benefit from new machine learning approaches. Artificial neural networks have been trained to pick up sound signatures for conditions such as COPD, asthma and interstitial lung disease[5]. The application of digital auscultation in primary care or acute setting may be a cheap and easily accessible point of care application of AI.
Strain on pathology resources from increased complexity and intensity of lung cancer workload may also be assisted by AI. Deep learning approaches have demonstrated similar ability to expert pathologists to distinguish between adenocarcinoma and squamous cell carcinomas and have even predicted the presence of common gene mutations[6].
What obstacles exist preventing the role out of AI in respiratory medicine in the UK?
As previously highlighted, there are several current and potential barriers to the role out of AI in respiratory medicine. These barriers include ethical constraints, technological limitations, regulatory frameworks, workforce and patient safety concerns and social barriers[7].
Ethical frameworks were established prior to the advent of AI and have been slow to adapt to a rapidly changing landscape. Concerns regarding privacy, consent and conflicts of interest must be met with clear regulatory frameworks to ensure transparency in the role out of AI. There are legitimate patient safety concerns where any task is handed over to an automated process and prospective clinical trials are required to demonstrate not just the efficacy of a software but to highlight their place within current and future clinical pathways.
The confidence of clinicians to understand the application and interpretation of AI solutions is critical to its establishment in mainstream healthcare. AI solutions need to have explainable methodology to avoid the “black box” stigma already established in the field. Given the substantial potential of AI to revolutionise so many areas of healthcare, the financial incentive to bring products to market risks compromising data quality, transparency and equity. Despite these hurdles the NHS remains poised to benefit from the widespread adoption of well-placed AI solutions.
How will you be using AI within your PhD?
I am undertaking a Wellcome Trust GW4-PhD Fellowship with the University of Exeter in collaboration with North Bristol NHS Trust, Royal Devon University Healthcare NHS Foundation Trust and Royal United Hospitals Bath NHS Foundation Trust. I am interested in the role of quantitative CT (qCT) in the prognostication and disease monitoring of fibrotic interstitial lung disease. Quantitative CT is a catch-all term for the use of AI to analyse CTs to provide objective and reproducible radiomic biomarkers. QCT can use techniques ranging from simple assessment of CT lung attenuation values to complex unsupervised deep learning algorithms which identify patterns in high resolution CT images to provide diagnostic or prognostic information.
These biomarkers have shown great promise in several studies to date but this hasn’t yet translated into clinical use. I am aiming to prospectively examine the added benefits of using qCT in addition to conventional methods such as respiratory physiology, visual CT interpretation and patient reported outcome measures to investigate the driving mechanisms of progressive pulmonary fibrosis (PREDICT-ILD NCT05609201).
How can I find out more about the applications of AI in respiratory medicine?
Resources available in the UK include the NHS AI lab (https://transform.england.nhs.uk/ai-lab/) and the Alan Turing Institute (https://www.turing.ac.uk/).
References
- Radiologists RCo. Clinical Radiology Workforce Census 2023; 2023.
- Baldwin DR, Gustafson J, Pickup L, Arteta C, Novotny P, Declerck J, Kadir T, Figueiras C, Sterba A, Exell A, Potesil V, Holland P, Spence H, Clubley A, O'Dowd E, Clark M, Ashford-Turner V, Callister ME, Gleeson FV. External validation of a convolutional neural network artificial intelligence tool to predict malignancy in pulmonary nodules. Thorax 2020: 75(4): 306-312.
- Boubnovski Martell M, Linton-Reid K, Hindocha S, Chen M, Moreno P, Alvarez-Benito M, Salvatierra A, Lee R, Posma JM, Calzado MA, Aboagye EO. Deep representation learning of tissue metabolome and computed tomography annotates NSCLC classification and prognosis. NPJ Precis Oncol 2024: 8(1): 28.
- Doe G, El-Emir E, Edwards GD, Topalovic M, Evans RA, Russell R, Sylvester KP, Van Orshoven K, Sunjaya AP, Scott DA, Prevost AT, Harvey J, Taylor SJ, Hopkinson NS, Kon SS, Jarrold I, Spain N, Banya W, Man WD. Comparing performance of primary care clinicians in the interpretation of SPIROmetry with or without Artificial Intelligence Decision support software (SPIRO-AID): a protocol for a randomised controlled trial. BMJ Open 2024: 14(6): e086736.
- Gonem S, Janssens W, Das N, Topalovic M. Applications of artificial intelligence and machine learning in respiratory medicine. Thorax 2020: 75(8): 695-701.
- Coudray N, Ocampo PS, Sakellaropoulos T, Narula N, Snuderl M, Fenyo D, Moreira AL, Razavian N, Tsirigos A. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med 2018: 24(10): 1559-1567.
- Ahmed MI, Spooner B, Isherwood J, Lane M, Orrock E, Dennison A. A Systematic Review of the Barriers to the Implementation of Artificial Intelligence in Healthcare. Cureus 2023: 15(10): e46454.