Chronic obstructive pulmonary disease (COPD) is characterized by airflow limitation associated with persistent respiratory symptoms. Most patients with COPD experience exacerbation of symptoms and are at high risk of developing comorbidities such as cardiovascular disease [1].

Long-term exposure to tobacco smoke, vapor, gas, dust, and fumes is an established major risk factor for COPD [2]. However, only a small percentage of smokers develop airflow limitation, while nonsmokers can develop COPD [3]. These inconsistencies indicate that risk factors other than long-term smoking are associated with COPD [4].

The prevalence of COPD has been reported to be 12% to 13% among smokers [5]. However, only 9.4% of patients with airflow limitation have a previous diagnosis of COPD, and European data indicate that up to 80% of COPD cases are undiagnosed [6], suggesting delays in the diagnosis of COPD. The ARCTIC observational cohort study showed that late COPD diagnosis was associated with a higher exacerbation rate and increased comorbidities and costs compared with early diagnosis [7].

To address the issue of undiagnosed COPD, significant risk factors for airflow limitation other than smoking should be identified and evaluated in routine clinical practice. In a cohort study of 9040 individuals from the Japanese general population, concomitant Chlamydia pneumoniae and Mycoplasma pneumoniae seropositivity was found to be an independent risk factor for airflow limitation [8]. Additionally, Sato et al employed an annual health examination with pulmonary function tests measuring airflow limitation to identify undiagnosed patients with COPD among the Japanese population and found that iron deficiency might be associated with COPD development [9]. However, the follow-up duration of these cohorts was short (<3 years), limiting their ability to identify risk factors for COPD in the general population.

A large questionnaire-based surveillance demonstrated some improvement in diagnostic rates for COPD; however, approximately 60% of eligible participants failed to respond to the questionnaire [10]. While these results suggest that identifying robust and relevant risk factors is likely to improve early diagnosis, the slow progression and heterogeneity of the disease have hindered the identification of such risk factors for COPD development.

The recently reported “Subtype and Stage Inference” machine learning computational model identified subtypes of patients with COPD [11]. Compared with traditional approaches, the advantages of machine learning include the ability to process complex nonlinear relationships between predictors and to provide novel outputs. Therefore, the aim of this study was to apply machine learning methods to predict possible risk factors for the development of airflow limitation, an essential feature of COPD diagnosis, using a Japanese medical check-up database comprising data from a number of healthy subjects to support the early diagnosis of COPD by general practitioners and occupational health physicians.

This study applied a machine learning method, a powerful tool to analyze large quantities of complex data, to predict risk factors for COPD. This is the first study to investigate more than 300,000 records from working-age adults in Japan utilizing an annual medical check-up database. This system allows healthy employees to track their health conditions over time by clinical measurements and questionnaires. We found that the most significant predictor of COPD diagnosis was the absolute value of FEV₁/FVC, indicating that low FEV₁ in early adulthood is an important factor in the development of COPD. Childhood asthma is associated with impaired lung function, lower lung function in adulthood, and higher risk of COPD even for nonsmoker participants, as previously reported by Martinez et al [21]. In our speculation, some part of the nonsmoker COPD population might have had a history of childhood asthma, increasing susceptibility to passive smoke exposure or airway pollution and resulting in the early diagnosis of COPD in nonsmokers compared with exsmokers in the study. Smoking status had the second highest impact on disease diagnosis. Among individuals with a smoking history, the peak age of COPD diagnosis was older in exsmokers than in current smokers. This finding suggests that smoking cessation delays the diagnosis of COPD, consistent with a previous study in which smoking cessation was reported to affect the natural history of COPD [22].

Erythrocyte indices (MCV and MCHC) might also be available as potential predictors of COPD diagnosis in addition to lung function measurements. These data are supported by a previous report in which continuous smoking had a significant effect on hematological parameters compared with nonsmoking, and it may be associated with an increased risk of COPD [23]. The increased levels of MCV and MCHC in individuals with COPD support a previous finding that impaired lung function has a strong association with ischemic heart disease [24]. Conversely, the presence of an allergic disease appeared to have a preventive effect on airflow limitation, which is in contrast with observations from the Tasmanian Longitudinal Health Study in which the presence of allergic diseases was an early predictor of lung trajectories toward COPD [25]. However, the Hokkaido cohort study showed that subjects with multiple asthma-like features had slower lung function decline [26]. From the findings of our observational study in Japan, we can speculate that early diagnosis and intervention for allergic diseases may have less impact on lung function and that regular and frequent medical intervention could lead to an overall increase in life expectancy among patients who can readily access appropriate treatment by respiratory specialists.

Furthermore, individuals with decreased levels of serum albumin and total protein, as well as lower hemoglobin A_1c and BMI may be at risk of developing cachexia, a common condition among patients with COPD [27]. With respect to other identified risk factors, a retrospective cross-sectional study showed an association between EOS and airflow limitation in patients with COPD [28]. Given that increased alanine aminotransferase levels have been observed in patients with obstructive sleep apnea [29], individuals at risk of developing COPD might be exposed to intermittent hypoxia, indicating that a reduced sleeping time, as determined in the study questionnaire, might also represent a risk factor for COPD. Even minor changes in hematological parameters might be attributable to hypoxic conditions, leading to sleep disruption. Additionally, frequently missing breakfast might accelerate malnutrition in the COPD group. Furthermore, significantly higher prevalence rates of chronic neck and lower back pain in patients with COPD compared with healthy individuals were observed in a population-based study, although the findings were not confirmed by logistic regression analysis [30], and the link between COPD and back pain remains unknown. The observation of increased WBC counts in patients with COPD compared with healthy controls [31] suggests that systemic inflammation may be involved in the pathogenesis of COPD [32].

Our results also indicate that smoking cessation should be prioritized for the prevention of COPD and that smokers with sleep disturbances, back pain, and/or low BMI and malnutrition may be at increased risk of developing COPD and should be considered as candidates for lifestyle intervention therapy. Furthermore, the five key questions included in our questionnaire should be validated in future investigations and potentially implemented in daily practice as part of an annual medical check-up to prevent COPD.

The positive predictive value of the XGBoost model was comparable to that of a self-scored persistent airflow obstruction screening questionnaire in the Japanese population previously reported by Samukawa et al [33]. However, our models showed more accuracy because the sensitivity and specificity of our models achieved higher figures, and the AUC reached over 0.9 compared with that of the questionnaire, which ranged from 0.595 to 0.612. The AUCs of the XGBoost and logistic regression models were similar, while the most important factor related to COPD diagnosis was FEV₁ in both models. However, some variables differed in importance in each model. Kuhn et al reported that machine learning approaches can incorporate high-order nonlinear interactions among predictors that cannot be addressed by traditional modeling approaches (eg, logistic regression models) [34]. However, machine learning methods cannot elucidate whether a causal relationship exists between the identified variable and the disease. Thus, the association between risk factors detected using a machine learning model and COPD requires validation in future prospective studies.

A strength of this study was the use of longitudinal lung function test data from healthy individuals from April 1998 to March 2019. In general, medical checkup data are not linked to medical records, meaning that profiles of lung function tests over time could not be investigated. However, it was possible to evaluate longitudinal lung function tests because the database included data from individuals from a point in time when they were healthy until they had developed COPD. Additionally, data from healthy individuals were included, allowing lung function test results from when they were diagnosed with COPD to be investigated. Finally, both clinical measurements and questionnaire variables were included in the database, thereby increasing the potential to identify several different risk factors for COPD.

The limitations of this study include the definition of COPD diagnosis by airflow limitation with pulmonary function tests. Instead of post-BD spirometry data as suggested by the ATS/European respiratory guidelines, we employed pre-BD spirometry data for the diagnosis of COPD since no post-BD spirometry was performed in the annual medical check-up. The precise diagnosis of COPD cannot always be demonstrated by airflow limitation alone; however, we believe that the diagnostic approach was reasonable from a clinical perspective as airflow limitation has been reported to be a poor prognostic factor in the general population [35]. Low lung function values (FEV₁/FVC <0.7) might be observed at a single time point in some individuals for no discernable reason. Therefore, we considered COPD as lung function of FEV₁/FVC <0.7 on two consecutive occasions. In terms of differentiation between asthma and COPD, we cannot exclude the possibility of misclassification of asthma as COPD in some patients since reversibility tests were not performed in the annual medical check-up, but participants with a medical history of asthma were excluded. Additionally, a database from a single organization was analyzed in this study; thus, the results might include bias based on the type of industry or the organizational structure of the company, limiting the generalizability of the findings. To obtain more generalizable findings, studies using other databases are necessary. Finally, some unknown confounders may have remained; therefore, we plan to perform model validation by analyzing other databases. Well-controlled prospective studies should be conducted to confirm the predictive factors for COPD diagnosis.

In conclusion, our machine learning method applied to longitudinal medical check-up data, including general questionnaires and laboratory parameters, identified hematological, nutritional, and inflammatory parameters as potential risk factors for COPD. These parameters, along with lung function and smoking status, may be useful in identifying at-risk individuals and may lead to an earlier diagnosis.