From images to outcomes: artificial intelligence-based standalone risk prediction in idiopathic pulmonary fibrosis

Idiopathic pulmonary fibrosis (IPF) is a fatal, progressive fibrosing interstitial lung disease associated with irreversible decline in lung function and poor survival, with an estimated median survival time of 3-5 years after diagnosis.1 Despite its prognosis, the clinical course of IPF is highly heterogeneous, with some individuals experiencing rapid progression while others remain relatively stable for a prolonged period. This marked variability makes accurate prognostic assessment a persistent challenge in clinical practice.2 Yet, reliable prognostication is central to IPF management, informing key decisions, including timing of lung transplantation referral, intensity of follow-up, and clinical trial enrollment. At present, prognostic evaluation in IPF relies primarily on clinical characteristics, pulmonary function tests, and composite indices such as the GAP score.3 High-resolution computed tomography (HRCT), integral to diagnosis and longitudinal disease assessment, also contains rich prognostic information and offers considerable potential for risk stratification.4-6 However, radiologic evaluation in routine practice remains largely qualitative and subjective, with substantial interobserver variability. In recent years, deep learning-based analysis of CT images has emerged as a promising strategy to address these limitations. Nevertheless, most prior studies have focused on image segmentation, feature extraction, or pattern classification, with prognostic associations evaluated secondarily, rather than developing models explicitly trained to predict mortality.4-6 Moreover, proprietary restrictions on access to models and code have often constrained reproducibility and independent validation, and meaningful clinical translation of artificial intelligence (AI)-based prognostic models for fibrotic lung diseases has yet to occur. In this issue of the Journal, Lee et al. report the development and external validation of ORACLE-IPF,7 a fully automated, attention-based deep learning model designed to predict 1- to 5-year all-cause and respiratory mortality in IPF from a single baseline chest CT scan. Across internal and external validation cohorts comprising 1690 patients, ORACLE-IPF demonstrated robust discrimination for long-term mortality prediction, with high reproducibility across repeat imaging, longitudinal scans, and varying CT reconstruction parameters. Notably, the model consistently outperformed deep learning-based CT segmentation metrics and showed comparable or superior performance relative to the GAP index. When integrated with established prognostic tools, ORACLE-IPF provided incremental prognostic value, whereas adding conventional clinical measures to the AI model yielded minimal additional benefit. Collectively, these findings suggest that ORACLE-IPF captures prognostic information embedded within CT images that is not fully represented by traditional clinical or physiologic assessments. Several aspects of this study warrant emphasis. First, ORACLE-IPF was explicitly developed as an outcome-driven prognostic model, trained directly on mortality endpoints rather than on surrogate imaging features presumed to correlate with outcomes.4,8,9 This design choice enhances its clinical relevance. Second, the incorporation of fibrosis-guided attention mechanisms offers a degree of interpretability, indicating that the model preferentially attends to radiologic features long associated with adverse outcomes, such as honeycombing.10 Third, the investigators demonstrated strong model reproducibility and robustness against varying CT acquisition parameters, a prerequisite for real-world deployment that remains insufficiently addressed in much of the AI literature. Importantly, the investigators have made the model and code publicly available, supporting future independent validation and prospective evaluation. The potential clinical implications are considerable. The work represents an important shift away from current tools that reduce a complex, spatially heterogeneous disease to low-dimensional summaries of physiology or extent, toward models trained directly on clinically meaningful endpoints. By demonstrating that a single baseline CT scan, analyzed using attention-based deep learning, can generate reproducible and clinically meaningful mortality predictions, this work underscores the prognostic value of routine imaging. As AI matures, such approaches have the potential to bridge the longstanding gap between radiologic appearance and individual patient-centered outcomes in IPF. A fully automated, CT-based prognostic tool requiring no manual annotation or additional clinical inputs could be readily integrated into routine workflows, facilitating early risk stratification, informing transplant referral and monitoring strategies, and improving patient selection for clinical trials, including in settings where pulmonary function testing is unavailable or unreliable. Nevertheless, several limitations warrant careful consideration. The study’s retrospective design introduces the potential for selection bias, despite the use of large cohorts and external validation. Both cohorts consisted exclusively of Asian patients, and generalizability to other ethnic populations, healthcare systems, and CT acquisition protocols remains to be established. In addition, the model was evaluated solely against mortality outcomes; other clinically relevant endpoints, such as acute exacerbations, functional decline, and lung transplantation, were not examined. Although the authors explored cause-specific outcomes, it remains uncertain to what extent the model’s predictions reflect fibrosis-related risk versus broader comorbidity or systemic risk signals (eg, cardiovascular disease), a distinction that will require prospective evaluation to clarify. Finally, direct comparisons with expert radiologist assessments were not performed, leaving open the question of how AI-based predictions might complement or augment human interpretation in practice. Looking ahead, large prospective multicenter studies across diverse populations are needed to establish the robustness and generalizability of ORACLE-IPF. To support this, standardized imaging protocols and shared standards for model validation will be essential. Furthermore, rigorous evaluation is required to determine whether such an AI-based prognostic tool, either alone or integrated with longitudinal clinical data, changes clinical decision-making or improves patient outcomes, thereby paving the way for its regulatory adoption and routine clinical use. In summary, the study by Lee et al. represents a significant step forward in the application of artificial intelligence to the longstanding challenge of IPF prognostication. By demonstrating that a single baseline CT scan, analyzed through attention-based deep learning, can provide reproducible and clinically meaningful mortality predictions, this work highlights the untapped prognostic potential of routine imaging. As AI continues to mature, such approaches may help bridge the longstanding gap between radiologic appearance and patient-centered outcomes in IPF. Supplementary material is available at American Journal of Respiratory and Critical Care Medicine online. Please see the ICMJE disclosure forms, which have been provided as supplementary material.