Idiopathic Pulmonary Fibrosis Mortality Risk Prediction Based on Artificial Intelligence: The CTPF Model

Abstract

Background: Idiopathic pulmonary fibrosis (IPF) needs a precise prediction method for its prognosis. This study took advantage of artificial intelligence (AI) deep learning to develop a new mortality risk prediction model for IPF patients. Methods: We established an artificial intelligence honeycomb segmentation system that segmented the honeycomb tissue area automatically from 102 manually labeled (by radiologists) cases of IPF patients' CT images. The percentage of honeycomb in the lung was calculated as the CT fibrosis score (CTS). The severity of the patients was evaluated by pulmonary function and physiological feature (PF) parameters (including FVC%pred, DLco%pred, SpO2%, age, and gender). Another 206 IPF cases were randomly divided into a training set (n = 165) and a verification set (n = 41) to calculate the fibrosis percentage in each case by the AI system mentioned previously. Then, using a competing risk (Fine-Gray) proportional hazards model, a risk score model was created according to the training set's patient data and used the validation data set to validate this model. Result: The final risk prediction model (CTPF) was established, and it included the CT stages and the PF (pulmonary function and physiological features) grades. The CT stages were defined into three stages: stage I (CTS≤5), stage II (5 < CTS<25), and stage III (≥25). The PF grades were classified into mild (a, 0-3 points), moderate (b, 4-6 points), and severe (c, 7-10 points). The AUC index and Briers scores at 1, 2, and 3 years in the training set were as follows: 74.3 [63.2,85.4], 8.6 [2.4,14.8]; 78 [70.2,85.9], 16.0 [10.1,22.0]; and 72.8 [58.3,87.3], 18.2 [11.9,24.6]. The results of the validation sets were similar and suggested that high-risk patients had significantly higher mortality rates. Conclusion: This CTPF model with AI technology can predict mortality risk in IPF precisely.

Keywords: artificial intelligence (AI); deep learning; disease severity grade; idiopathic pulmonary fibrosis (IPF); pulmonary fibrosis stage; semantic segmentation.

FIGURE 1

Case screening process. In total, 232 cases were diagnosed as IPF according to the 2018 IPF diagnosis and treatment guidelines. A total of 26 patients were excluded, two patients were diagnosed as interstitial pneumonia with autoimmune features (IPAF) during follow-up; 24 patients had incomplete CT and lung function data. Finally, 206 cases were included in the retrospective analysis (including 16 cases of lung transplantation): 93 surviving cases, including 11 lung transplants; 81 deaths, out of which 10 died from lung cancer, 67 died from acute exacerbation of IPF, and 4 died after lung transplantation; and 32 patients failed to follow up, including one failed to follow up after lung transplantation.

FIGURE 2

Correlation of AI–CT fibrosis score and lung function parameters. (A) Correlation between CT-score and FVC%pred, Spearman correlation coefficient rs = -0.40, p < 0.01; (B) correlation between CT-score and DLco%pred, Spearman correlation coefficient rs = -0.66, p < 0.01; (C) correlation between CT-score and SpO2%, Spearman correlation coefficient rs = -0.44, p < 0.01; (D) correlation between CT-score and CPI, Spearman correlation coefficient rs = 0.65, p < 0.01; and (E) correlation between CT-score and manual-CT scores by radiologists, Spearman correlation coefficient rs = 0.80, p < 0.01.

FIGURE 3

Analysis of CT stage and PF grading and mortality.(A) shows the relationship between CT staging and mortality risk based on Fine–Gray regression CT staging univariate analysis, which might be mixed with the influence of PF grade. (B) shows the same relationship in the multivariate analysis of CT staging and PF grading. Adjusted PF grading means the effect of PF grading was eliminated. The results showed that CT stage, with both PF grade adjusted and unadjusted, was positively correlated with mortality risk. (C) shows the relationship between PF grade and mortality risk based on Fine–Gray regression PF grade univariate analysis, which might be mixed with the influence of CT staging. (D) shows the same relationship between the multivariate analysis of CT staging and PF classification. CT staging adjusted means the effect of the CT stage was eliminated. The results show that the PF grade, with both CT staging adjusted and unadjusted, is positively correlated with mortality risk.

FIGURE 4

Model prediction nomogram and calibration curve. (A) Mortality nomogram of CTPF as the predictive model. (B–D) Calibration curves after cross-validation using CT staging, PF staging, CTPF comprehensive staging, and GAP staging to predict patients’ cumulative mortality risk at 1, 2, and 3 years. The CTPF prediction model has the best AUC value, Brier score, and stability.

FIGURE 5

Examples of patient’s original lung CT image, honeycomb lung region segmentation, and staging. (A-1,2,3) are the original CT images, the segmented lung region, and honeycomb lung region identified by the deep learning model of patient Zhang. The corresponding stage of this patient is II c. Similarly, (B-1,2,3) are the corresponding images of patient Xu, whose stage is Ia. (C) shows their comprehensive CTPF staging; patient Zhang’s comprehensive stage is IIc; the comprehensive stage of patient Xu is Ia.