Application of artificial intelligence in respiratory medicine: Has the time arrived?

Abstract

Artificial intelligence (AI) has seen increasing use across all sectors, as enabled by increasing volumes of data, greater computing power and novel algorithm architectures. In health care, there has been an exponential increase in research involving AI, as reflected by a surge in publications and academic funding. AI could play a role in automating lung cancer screening1 and is being trialled by the UK's National Health Service for triaging chest X-rays for radiologist review.2 Potential applications within respiratory medicine are legion, particularly in the analysis of data for diagnosis, but there are associated challenges with its implementation. Additionally, its role will be as an adjunct, rather than a replacement, for the experience and human touch that skilled physicians provide. Machine learning (ML) encompasses a group of AI methods by which computers can identify patterns and relationships between data (such as radiological images or blood tests) and outcomes of interest. Traditional statistical methods characterize such patterns with mathematical equations. In ML, the computer analyses large volumes of data to learn complex, non-linear relationships that enable greater accuracy and cannot easily be expressed as an equation. ML also enables the analysis of types of data that were previously not amenable to computational analysis, such as imaging and auditory data. The diagnosis of respiratory conditions relies on central tenets of medicine: the history, the examination, bedside tests and imaging. Skilled respiratory physicians can detect subtle changes in breath sounds on auscultation, interpret variations in pulmonary function test (PFT) scores and analyse images from X-rays, computed tomography (CT) scans and bronchoscopy. AI may play a role in supporting physicians in these areas. ML analysis of breath sounds obtained from electronic stethoscopes is objective, not prone to inter-clinician variability and is not restricted to the human auditory frequency range. Genetic algorithms and neural networks have shown good specificity and sensitivity for detecting wheezes and crackles.3 However, we should avoid to become overly reliant on such analysis, or we risk becoming de-skilled. ML may enhance the analysis of PFT scores. Complex, multidimensional patterns of PFT variation may identify disease subtypes, personalizing diagnosis and treatment. Standardization of interpretation could also be improved. One AI model identified the correct diagnostic category more often than pulmonologists,4 although the true clinicians' performance was underestimated as they received limited clinical information. Using ML in bronchoscopy, to analyse images and diagnose potential cancers, has been explored. One study achieved 86% diagnostic accuracy.5 However, treatment decisions are better informed by the definitive histology results, which are typically reported within days. Real-time highlighting of potentially missed lesions could be a more useful approach, and analogous efforts have shown promise in colonoscopy. Convolutional neural networks (CNNs) are specialized ML methods, excelling at imaging analysis, which may support assessment of respiratory disease on chest X-ray or CT scan. The former is the most commonly performed radiological investigation. Thorough interpretation of such volume is time-consuming, which can lead to fatigue-based diagnostic error and requires a sufficient number of radiologists. Early ML studies focused on identifying specific lesions while more recently CNNs have been trained to detect multiple pathologies.6 While reasonable accuracies have been obtained, such algorithms will not enable full automation unless they are trained to recognize all possible pathologies, or else rarer conditions may be missed. When such algorithms are implemented clinically, it is important to communicate to users the specificity of their focus. Such algorithms could still play a useful role in prioritizing scans for review, screening for potentially missed pathologies and enabling objective measurements of radiographical features. One study demonstrated that an AI algorithm for prioritizing chest X-rays reduced the average time for scans with a critical finding to be reviewed by a radiologist from 11.2 to 2.7 days.7 AI may enable an objective, standardized measurement of signs, such as consolidation, where there is significant variability in reporting but developing standardized guidelines is challenging. ML may support the use of CT scans for diagnosis and screening. Nodules are prevalent but only a small percentage is cancerous. One algorithm analysed parenchymal features to identify cancerous nodules with an area under the curve (AUC) of 0.965.8 A novel ‘end-to-end’ deep learning model, which learns the features of lesions and compares them to earlier scans, identified cancer with an AUC of 0.944.1 There is potential for ML to enhance methods for treating respiratory disease, but the research field is less substantiated. One application is to support lung movement tracking for the accurate delivery of radiotherapy to lung cancers. The system latency of existing techniques limits their accuracy while real-time retraining of neural networks has been shown to improve precision of delivery.9 There are important considerations to ensure the safe and effective implementation of AI into health care. Data collection and usage must follow informed consent. Although AI is complex, we should proactively communicate to patients exactly about how their data are being used. Data security and privacy are key features. Cases like the Theranos scandal damage public trust and waste investment potential, which could jeopardize the development of potentially life-saving technology. Development of AI algorithms requires large volumes of well-structured data. Clinician input is key, given their unique insight, so such collaboration should be encouraged. We must guard against over-reliance on algorithms and de-skilling of our workforce. There are challenges of apportioning responsibility to algorithms, meaning it may ultimately lie with the clinician. Good performance metrics do not guarantee a positive clinical impact, and this must be validated through prospective trials in clinical settings. Clinical workflows must also be considered. Algorithms must generalize to different populations and work with variable levels of data quality. One challenge is limited interpretability, particularly of deep learning algorithms. AI has many potential applications in respiratory medicine, particularly with assisting diagnosis from examination, bedside tests and imaging. While there are studies supporting the effectiveness of such analysis, it remains to be validated in clinical settings and the nature of integration into clinical workflows is yet to be seen. If implemented successfully, AI can overcome inefficiencies and better equip physicians with enhanced diagnostic tools, freeing them to utilize their unique skills and empathy in the provision of high-quality care. C.A.L. is an employee of Cera Care. M.M. is an investor and employee of Cera Care. Cera Care is a domiciliary care provider conducting research into how AI can be used to improve the care delivered to elderly people living at home. No funding was received for this work.