Article Preview
TopIntroduction
Chronic obstructive pulmonary disease (COPD) is a respiratory disease characterised by airflow limitation (Vestbo et al., 2013). It poses a serious threat to human health with a high incidence. In fact, the disease is one of the most significant public health problems affecting global economic and social development (Alkhathlan et al., 2020; Corlateanu et al., 2020; Halpin et al., 2021). COPD is the third leading cause of death worldwide and the fifth disease causing a substantial social burden, with approximately three million deaths each year (Lozano et al., 2012) and more than 300 million people who suffer it (Adeloye et al., 2015). With an aging population, the number of people affected is expected to increase (Lopez-Campos et al., 2016).
Pulmonary function tests (PFTs) are essential tools for diagnosing COPD and assessing the function of the respiratory system (Crapo, 1994). However, interpreting complex, multidimensional, nonlinear, and heterogeneous data in PFT reports can be challenging. Experts rely on international guidelines to identify disease patterns (obstructive, restrictive, mixed, and normal) and grade their severity (Pellegrino et al., 2005; Vogelmeier et al., 2017). The final clinical diagnosis requires combining PFT results with patient information, symptoms, and other tests, such as imaging, blood analysis, and biopsy (Galie et al., 2016; Martinez et al., 2017).
Although PFT reports contain rich clinical information, reliance on human interpretation has limitations. Owing to international diagnostic rules, physicians can only make preliminary judgments based on some of the test values. Furthermore, clinicians may lack experience, leading to misdiagnosis and delayed COPD treatment (Tinkelman et al., 2006).
To improve the accuracy of COPD diagnosis, it is essential to mine hidden clinical information from PFT reports. Utilising the entire report can reduce unnecessary examinations and wasted medical resources. However, few studies have utilised PFT reports to aid in the diagnosis of COPD, likely because of difficulties in acquiring the necessary data. The data contained in PFT reports are physically separated from other hospital data, creating a data island. These data are stored in standalone spirometers; they cannot be integrated and analysed with other data like electronic medical records (EMRs).
Building on the research findings of the authors’ laboratory, this study represents the first effort to extract 230,000 PFT reports from stand-alone spirometers at Xiangya Hospital of Central South University, thereby breaking down the data barriers that hindered previous studies. Given the focus of this study on the auxiliary diagnosis of COPD, the authors examined the bronchodilation reports in the PFT reports, totalling 16,012 reports.
The PFT is globally standardised, making it an ideal candidate for developing artificial intelligence (AI) algorithms for assisted diagnosis (Jordan & Mitchelle, 2015; Kononenko, 2001). AI can identify subtle and decisive features that are difficult for humans to detect. Then, the technology can incorporate the features into powerful differential diagnostic algorithms.
This study developed a complete solution for the auxiliary diagnosis of COPD using PDF-format PFT reports. The proposed algorithm comprises six parts: (1) pre-processing PDF report data; (2) matching report ID; (3) handling missing data; (4) data selection; (5) feature selection; and (6) model training. Specifically, the proposed algorithm initially uses text detection and recognition algorithms (Liao et al., 2020; Shi et al., 2017) to process the PDF format and obtain structured data. It then matches the missing patient identifications (IDs) in the PFT reports and handles the missing values in the examination indicators. Subsequently, data inclusion and exclusion are performed to obtain a clean dataset.