
Artificial intelligence (AI) is capable of handling large volumes of patient data to identify patterns and trends relevant to clinical practice. Given the sheer volume of data that is analyzed by this technology, some have argued that AI tools are “more experienced” than individual physicians at pattern recognition and may help physicians in all aspects of patient care, including diagnosis, clinical decision-making, and prognostication. Additionally, AI models are adaptable and can “learn” over time as more data become available.

The current paradigm for diagnosing interstitial lung disease (ILD) requires synthesizing radiological, pathological, and clinical data by clinical experts in a multidisciplinary review. This is typically available at academic centers, leaving community-based pulmonologists at a disadvantage. The median time to ILD diagnosis is up to 2.1 years.1 Patients whose disease has progressed in that time often experience worse outcomes. Several institutions have reported on their experiences using convolutional neural networks (CNN), a type of deep learning model adept at pattern recognition, to help address these diagnostic delays. Mei and colleagues utilized a joint CNN combining chest CT scans and clinical data to classify five ILD subtypes and compare the diagnostic accuracy of this model with human experts.2 The model outperformed human experts in diagnosing usual interstitial pneumonia with 82.4% sensitivity and 68.1% specificity. It performed as well as human experts in diagnosing nonspecific interstitial pneumonia, chronic hypersensitivity pneumonitis, and sarcoidosis.

AI may also help predict which patients with ILD are at risk for disease progression and who might respond to ILD therapies. Deep learning methods quantifying fibrosis and honeycombing on chest CT scans have been trained on longitudinal data to predict disease progression and prognosis.3 AI tools combining imaging, clinical, and lung function data show promise in predicting clinical outcomes such as decline in FVC. Predictive tools such as these could be invaluable in informing decisions about starting therapies and improving patient outcomes.
Despite the promising potential for AI to improve ILD diagnosis and management, these tools are still early in their clinical implementation and have significant limitations.4 First, most tools are developed at single centers and lack external validation. Secondly, ILDs are a rare group of diseases with more than 200 subtypes, so the ability to obtain good, structured clinical data on which to train AI tools is limited and mostly available at large tertiary care centers. Furthermore, the models generated through AI methods are frequently “black boxed,” meaning that the data that inform the model are known, but the decisions made in generating the model are not.5 Finally, human biases that are reflected in clinical data can be translated into AI bias, perpetuating errors that already plague clinical practice.
AI has a promising future in improving care for patients with ILD. It has the potential to shorten the time to diagnosis, predict disease progression, and hopefully improve patient outcomes. However, given the limitations described above, AI tools will need to undergo significant refinement before widespread adoption in clinical practice.
References
1. Hoyer N, Prior T S, Bendstrup E, Wilcke T, Shaker S B. Risk factors for diagnostic delay in idiopathic pulmonary fibrosis. Respir Res. 2019;20(1):103.
2. Mei X, Liu Z, Singh A, et al. Interstitial lung disease diagnosis and prognosis using an AI system integrating longitudinal data. Nat Commun. 2023;14(1):2272.
3. Dack E, Christe A, Fontanellaz M, et al. Artificial intelligence and interstitial lung disease: diagnosis and prognosis. Invest Radiol. 2023;58(8):602-609.
4. Gonem S. Artificial intelligence in respiratory medicine. In: Artificial Intelligence in Clinical Practice. Academic Press; 127-135.
5. Baselli G, Codari M, Sardanelli F. Opening the black box of machine learning in radiology: can the proximity of annotated cases be a way? Eur Radiol Exp. 2020;4(1):30.