AI and Pulmonary Fibrosis: New Paths to IPF Prognosis

Idiopathic pulmonary fibrosis (IPF) is one of the most complex and unpredictable respiratory diseases in modern pulmonology. Despite advances in treatment, the prognosis remains highly variable: some patients experience a rapid decline, whilst others show a slower and more stable progression.

In this context, a recent study proposes an innovative approach based on artificial intelligence (AI) to improve the ability to predict the course of the disease, using a simple test already available in clinical practice: chest computed tomography (CT).

The problem: an unpredictable disease

IPF is characterised by:

1. progressive loss of lung function;
2. worsening respiratory symptoms;
3. a median survival of approximately 3–5 years from diagnosis.

However, the clinical course is extremely heterogeneous. The tools currently used to estimate prognosis – such as the GAP index (Gender, Age, Physiology) or functional parameters (FVC, DLCO) – have significant limitations and do not always capture the complexity of the disease.

A new approach: artificial intelligence applied to CT

A recent study by South Korean authors published in the American Journal of Respiratory and Critical Care Medicinedescribes a deep learning model called ORACLE-IPF (Outcome Risk Assessment via Computational Learning for Estimating mortality).

This is a fully automated system that uses a single chest CT scan performed at diagnosis to estimate the risk of mortality up to 5 years.

The model is based on:

• three-dimensional convolutional neural networks;
• ‘attention’ mechanisms;
• analysis of the distribution and extent of pulmonary fibrosis.

The aim is not only to recognise the disease, but to predict its clinical progression.

How the study was conducted

The study involved:

• over a thousand patients for model development;
• an independent cohort of 667 patients for external validation;

The model was designed to:

• automatically analyse the CT scan;
• identify relevant fibrotic patterns;
• estimate the risk of mortality at 1, 2, 3, 4 and 5 years;

A particularly interesting feature is the system’s ability to generate attention maps, which highlight the areas of the lung considered most relevant for prognosis.

The results: a more accurate prognosis

The study’s results are particularly significant:

• AUC between 0.75 and 0.81 for overall mortality;
• AUC up to 0.87 for respiratory mortality.

The model demonstrated:

• superior performance compared to traditional methods;
• greater accuracy than the GAP index;
• the ability to add value even when combined with other clinical parameters.

Furthermore, ORACLE-IPF showed high reproducibility, both in repeat scans on the same day and in longitudinal follow-up.

Why this study is important

This work introduces some key concepts for the future of respiratory medicine:

1. CT as a prognostic biomarker

CT is no longer merely a diagnostic tool, but can become a quantitative indicator of mortality risk.

2. Automation and objectivity

The model eliminates the inter-observer variability typical of traditional radiological assessment.

3. Personalised medicine

The ability to estimate individual risk paves the way for more targeted patient management.

Clinical implications

The potential applications of this approach are numerous:

• early identification of high-risk patients;
• more appropriate selection for lung transplantation;
• planning of follow-up and treatment intensity;
• support in prognostic communication with the patient.

Furthermore, the model could be used as a biomarker in clinical trials, improving patient selection and the efficiency of studies.

Limitations of the study

Like any innovative study, this one also has certain limitations:

• retrospective design;
• predominantly Asian population;
• need for validation in different clinical settings.

Furthermore, integration into clinical practice will require further prospective studies.

This study represents an important step towards an increasingly data-driven approach to respiratory medicine. The use of artificial intelligence allows complex prognostic information to be extracted from existing tests, without additional costs or new procedures.

If confirmed in future studies, these tools could significantly change the way we assess and manage patients with pulmonary fibrosis.

Idiopathic pulmonary fibrosis remains a difficult disease to predict and manage. The introduction of artificial intelligence-based models, such as ORACLE-IPF, represents one of the most promising innovations of recent years.

The ability to estimate the risk of mortality from a single CT scan could mark the beginning of a new era in pulmonology, in which diagnosis, prognosis and treatment decisions are increasingly integrated.