Development and validation of the EHS-COPD model to predict sex-specific risk of chronic obstructive pulmonary disease (COPD) in older Chinese adults: Hong Kong's Elderly Health Service Cohort.

BACKGROUND: No screening program is recommended for chronic obstructive pulmonary disease (COPD) in adults based on current clinical practice guidelines. Risk prediction models for COPD developed in Western settings may not be directly applicable to older Chinese adults. To evaluate the performance of an existing risk prediction model for COPD developed in a Western setting in Chinese adults and investigate whether a new risk prediction model performs better in predicting 5-year risk of COPD (EHS-COPD).
METHODS: This study is based on 135,822 participants aged 65+ years from Hong Kong's Elderly Health Service (EHS) cohort. We assessed the performance of an existing risk prediction model in the entire cohort, and in a random sub-sample of 91,133 participants, we recalibrated the existing model and derived a new model using extended Cox proportional hazards regression. Candidate risk predictors from the literature and the EHS cohort were considered for inclusion. Risk prediction performance, discrimination, and calibration of the newly derived models were assessed in the remaining 44,689 participants.
RESULTS: The existing risk prediction model overestimated the 5-year risk of COPD in older Chinese adults (65+ years); after recalibration, it still overestimated the 5-year risk of COPD for both men and women. The new EHS-COPD risk prediction model, including time-varying factors (i.e., age and smoking status) and time-invariant factors (i.e., education level, public assistance, alcohol use, body mass index, physical activity, existing hypertension, recent falls, cognitive function, and self-rated health status), had an improved performance. For men, EHS-COPD explained 19.5% of COPD risk, the D statistic was 23.1, and Harrell's C statistic was 0.93. The corresponding values for women were 8.5%, 21.1, and 0.93.
CONCLUSIONS: The existing COPD risk prediction model overpredicted COPD risk in older Chinese and could not be recalibrated to predict well. A revised prediction model using time-invariant and time-varying factors provides a better tool for identifying older Chinese adults at high risk of developing COPD. 2022 Annals of Translational Medicine. All rights reserved.

Introduction

Chronic obstructive pulmonary disease (COPD) and its complications are major causes of morbidity, mortality, and hospitalizations worldwide (1). In China, more than 86 million of ~633.0 million adults aged 40+ years were estimated to have COPD in 2014–2015, with the total cost of COPD ranging from US $1,964 to $3,449 per patient, representing 33–40% of the average household income (2-4). In Hong Kong, ~0.5% of the non-institutionalized population aged 15+ years had COPD in 2014–2015 (~43.5% were older adults aged 65 years or above), which resulted in ~2.7% of all registered deaths in 2017 (5). COPD is associated with substantial morbidity, increased healthcare use, disability, and death (1,6). Up to ~70% of COPD cases are undiagnosed and the prevalence of COPD is increasing globally (7-9), raising the possibility that earlier identification of those at high risk of COPD for suitable interventions could be beneficial.

Screening for COPD is not currently recommended in adults (10), because there is no direct evidence showing benefits of screening asymptomatic adults for COPD using either questionnaires or office-based screening via pulmonary function testing. With recent clinical trials demonstrating that smoking cessation (11-13) and pharmaceutical interventions (12,14-16) in people with mild or moderate COPD can ameliorate the decline of lung function and delay the risk of COPD exacerbations, identifying people at high risk of developing COPD using risk prediction models for targeted interventions might help reduce the burden of COPD.

To date, several risk prediction models have been developed for COPD, which either incorporate well-established risk factors (e.g., age and smoking) (17,18) or are restricted to asthma patients (19) from specific cohorts, limiting their application to the general population. In China, risk prediction models for COPD, including environmental, population, and genetic attributes (20) or peak expiratory flow measurement (21), are recommended for screening people at high risk. However, the inclusion of genetics or hard-to-measure factors limits their use in primary care. Existing risk prediction models do not always generalize to new settings because of health-related differences between populations. For example, the most widely used risk prediction models developed in Western settings, e.g., the Framingham cardiovascular disease score (22) and the Framingham diabetes score (23), have not always been applicable in a Chinese setting even after recalibration (24).

In this study, we first assessed the performance of an existing risk prediction model developed by Kotz et al. (17) in a Western setting, which uses four risk factors (i.e., age, smoking, prior asthma, and deprivation), in a Chinese setting. Secondly, we investigated whether a new model to predict COPD risk within the next 5 years in older Chinese adults (65+ years) performed better, based on a large population-based cohort, i.e., Hong Kong’s Elderly Health Services (EHS) Cohort. We followed the Transparent reporting of a multivariable model for individual prognosis or diagnosis (TRIPOD) guidance for development and reporting of a new model (available at https://dx.doi.org/10.21037/atm-21-3270) (25).

Methods

Study design and data source

The EHS cohort is a large, contemporary, prospective cohort study of older Chinese adults aged 65+ years in Hong Kong enrolled at 18 territory-wide Elderly Health Centres. Baseline and follow-up data were provided by the Department of Health, as detailed elsewhere (26). Specifically, the study aimed to promote understanding of aging in a global context and to take advantage of Hong Kong as a sentinel for Chinese populations currently experiencing very rapid economic development. The initial cohort enrolled ~66,820 participants from July 1998 to December 2001 at the Elderly Health Centres. All older adults in Hong Kong were encouraged to enroll at a nominal annual fee (22,26). At the centres, nurses and doctors provided physical check-ups and health assessments using clinical examinations and structured interviews, including demographics, lifestyle, socio-economic position, health status, self-rated health, physical functioning, social contact, depressive symptoms, and cognitive functioning. Self-reports were supplemented by clinical diagnoses based on medical history, if available. The participants were followed up at subsequent visits. Vital status was ascertained via record linkage to the Hong Kong Deaths Registry. Up to December 31, 2012, ~136,309 participants were enrolled in the EHS cohort, which was about 13.6% of the Hong Kong population aged 65+ years in 2012. This study excluded participants with existing COPD at baseline or those with >50% missing values for all predictors. For each participant, the entry date was the date of the baseline health assessment. Participants were censored at the earliest of date of death, attrition, or study end date (December 31, 2015). We randomly allocated two-thirds of the participants to the derivation cohort and the remainder to the validation cohort.

Outcomes

The outcome was the 5-year predicted risk of COPD. Self-reported new COPD cases were identified during follow-up visits. The time to event was measured from entry to date of self-report of a new COPD case.

Candidate risk predictors

We first assessed the performance of the model developed by Kotz et al. (17) using similar risk predictors [i.e., age, a measure of deprivation (Carstairs index), asthma, and smoking history]. We also selected additional potential risk predictors, available for the EHS cohort, from the relevant literature ().

Table 1

Sex-specific baseline characteristics of participants aged 65 years or older in derivation and validation cohorts at study entry

Characteristics	Derivation cohort				Validation cohort
	Men (n=32,222)	Women (n=58,911)	Overall (n=91,133)		Men (n=15,690)	Women (n=28,999)	Overall (n=44,689)
Age (years), mean (SD)	72.18 (5.20)	71.68 (5.37)	71.86 (5.31)		72.11 (5.12)	71.69 (5.40)	71.84 (5.31)
Type of housing
Public or aided	14,221 (44.1)	26,488 (45.0)	40,709 (44.7)		6,935 (44.2)	13,016 (44.9)	19,951 (44.6)
Private (rented)	1,336 (4.1)	2,347 (4.0)	3,683 (4.0)		693 (4.4)	1,147 (4.0)	1,840 (4.1)
Private (self-owned)	15,517 (48.2)	27,154 (46.1)	42,671 (46.8)		7,484 (47.7)	13,360 (46.1)	20,844 (46.6)
Others	1,146 (3.6)	2,920 (5.0)	4,066 (4.5)		578 (3.7)	1,476 (5.1)	2,054 (4.6)
No. of missing values	2 (0.0)	2 (0.0)	4 (0.0)		NA	NA	NA
Public assistance
Yes	4,301 (13.3)	8,187 (13.9)	12,488 (13.7)		2,005 (12.8)	4,053 (14.0)	6,058 (13.6)
No	27,915 (86.6)	50,717 (86.1)	78,632 (86.3)		13,685 (87.2)	24,942 (86.0)	38,627 (86.4)
No. of missing values	6 (0.0)	7 (0.0)	13 (0.0)		0 (0.0)	4 (0.0)	4 (0.0)
Education
Illiterate	2,089 (6.5)	19,994 (33.9)	22,083 (24.2)		990 (6.3)	9,790 (33.8)	10,780 (24.1)
Literate but no formal education	2,948 (9.1)	9,659 (16.4)	12,607 (13.8)		1,428 (9.1)	4,845 (16.7)	6,273 (14.0)
Primary	14,638 (45.4)	19,280 (32.7)	33,918 (37.2)		7,152 (45.6)	9,353 (32.3)	16,505 (36.9)
Secondary or above	12,539 (38.9)	9,962 (16.9)	22,501 (24.7)		6,118 (39.0)	5,006 (17.3)	11,124 (24.9)
No. of missing values	8 (0.0)	16 (0.0)	24 (0.0)		2 (0.0)	5 (0.0)	7 (0.0)
Body mass index (kg/m2)
<18.5	1,593 (4.9)	2,707 (4.6)	4,300 (4.7)		827 (5.3)	1,332 (4.6)	2,159 (4.8)
18.5–<23.0	11,021 (34.2)	19,030 (32.3)	30,051 (33.0)		5,403 (34.4)	9,512 (32.8)	14,915 (33.4)
23.0–<25.0	8,188 (25.4)	13,040 (22.1)	21,228 (23.3)		3,794 (24.2)	6,494 (22.4)	10,288 (23.0)
25.0+	11,415 (35.4)	24,132 (41.0)	35,547 (39.0)		5,665 (36.1)	11,657 (40.2)	17,322 (38.8)
No. of missing values	5 (0.0)	2 (0.0)	7 (0.0)		1 (0.0)	4 (0.0)	5 (0.0)
Smoking status
Never smoker	12,864 (39.9)	52,572 (89.2)	65,436 (71.8)		6,351 (40.5)	26,017 (89.7)	32,368 (72.4)
Ex-smoker	13,710 (42.5)	3,813 (6.5)	17,523 (19.2)		6,576 (41.9)	1,744 (6.0)	8,320 (18.6)
Current smoker	5,569 (17.3)	1,715 (2.9)	7,284 (8.0)		2,737 (17.4)	813 (2.8)	3,550 (7.9)
No. of missing values	79 (0.2)	811 (1.4)	890 (1.0)		26 (0.2)	425 (1.5)	451 (1.0)
Alcohol use
Never drinker	21,645 (67.2)	53,081 (90.1)	74,726 (82.0)		10,589 (67.5)	26,174 (90.3)	36,763 (82.3)
Ex-drinker	3,156 (9.8)	1,842 (3.1)	4,998 (5.5)		1,521 (9.7)	905 (3.1)	2,426 (5.4)
Social drinker	4,440 (13.8)	3,324 (5.6)	7,764 (8.5)		2,109 (13.4)	1,610 (5.6)	3,719 (8.3)
Moderate drinker	1,686 (5.2)	410 (0.7)	2,096 (2.3)		854 (5.4)	202 (0.7)	1,056 (2.4)
Excessive drinker	1,288 (4.0)	251 (0.4)	1,539 (1.7)		615 (3.9)	104 (0.4)	719 (1.6)
No. of missing values	7 (0.0)	3 (0.0)	10 (0.0)		2 (0.0)	4 (0.0)	6 (0.0)
Physical activity (hours per week)
None	4,215 (13.1)	7,104 (12.1)	11,319 (12.4)		2,000 (12.7)	3,494 (12.0)	5,494 (12.3)
<1.5	2,516 (7.8)	4,943 (8.4)	7,459 (8.2)		1,258 (8.0)	2,442 (8.4)	3,700 (8.3)
1.5–<3.0	5,017 (15.6)	10,185 (17.3)	15,202 (16.7)		2,515 (16.0)	4,880 (16.8)	7,395 (16.5)
3.0–<4.5	3,983 (12.4)	7,157 (12.1)	11,140 (12.2)		1,844 (11.8)	3,482 (12.0)	5,326 (11.9)
4.5+	16,485 (51.2)	29,513 (50.1)	45,998 (50.5)		8,073 (51.5)	14,699 (50.7)	22,772 (51.0)
No. of missing values	6 (0.0)	9 (0.0)	15 (0.0)		0 (0.0)	2 (0.0)	2 (0.0)
Family history of cardiovascular diseases
No	15,333 (47.6)	28,691 (48.7)	44,024 (48.3)		7,445 (47.5)	14,060 (48.5)	21,505 (48.1)
Yes	2,715 (8.4)	5,655 (9.6)	8,370 (9.2)		1,417 (9.0)	2,858 (9.9)	4,275 (9.6)
No. of missing values	14,174 (44.0)	24,565 (41.7)	38,739 (42.5)		6,828 (43.5)	12,081 (41.7)	18,909 (42.3)
Hypertension (yes)	702 (2.2)	1,322 (2.2)	2,024 (2.2)		337 (2.1)	586 (2.0)	923 (2.1)
Asthma (yes)	24 (0.1)	85 (0.1)	109 (0.1)		15 (0.1)	32 (0.1)	47 (0.1)
Recent falls
Yes	274 (0.9)	922 (1.6)	1,196 (1.3)		134 (0.9)	474 (1.6)	608 (1.4)
No	31,939 (99.1)	57,978 (98.4)	89,917 (98.7)		15,556 (99.1)	28,517 (98.3)	44,073 (98.6)
No. of missing values	9 (0.0)	11 (0.0)	20 (0.0)		NA	8 (0.0)	8 (0.0)
Hospital admission (yes)
Yes	4,554 (14.1)	7,075 (12.0)	11,629 (12.8)		2,189 (14.0)	3,442 (11.9)	5,631 (12.6)
No	27,666 (85.9)	51,835 (88.0)	79,501 (87.2)		13,501 (86.0)	25,557 (88.1)	39,058 (87.4)
No. of missing values	2 (0.0)	1 (0.0)	3 (0.0)		NA	NA	NA
Physical functioning
Normal	30,509 (94.7)	54,021 (91.7)	84,530 (92.8)		14,858 (94.7)	26,571 (91.6)	41,429 (92.7)
Poor	1,588 (4.9)	4,672 (7.9)	6,260 (6.9)		770 (4.9)	2,316 (8.0)	3,086 (6.9)
No. of missing values	125 (0.4)	218 (0.4)	343 (0.4)		62 (0.4)	112 (0.4)	174 (0.4)
Cognitive function
Normal	31,268 (97.0)	54,246 (92.1)	85,514 (93.8)		15,272 (97.3)	26,708 (92.1)	41,980 (93.9)
Poor	889 (2.8)	4,557 (7.7)	5,446 (6.0)		390 (2.5)	2,245 (7.7)	2,635 (5.9)
No. of missing values	65 (0.2)	108 (0.2)	173 (0.2)		28 (0.2)	46 (0.2)	74 (0.2)
Depressive symptoms
No/fewer	30,446 (94.5)	53,643 (91.1)	84,089 (92.3)		14,850 (94.6)	26,430 (91.1)	41,280 (92.4)
More	1,714 (5.3)	5,143 (8.7)	6,857 (7.5)		824 (5.3)	2,512 (8.7)	3,336 (7.5)
No. of missing values	62 (0.2)	25 (0.2)	187 (0.2)		16 (0.1)	57 (0.2)	73 (0.2)
Self-rated health status
Better	19,794 (61.4)	33,412 (56.7)	53,206 (58.4)		9,701 (61.8)	16,383 (56.5)	26,084 (58.4)
Normal	2,262 (7.0)	4,108 (7.0)	6,370 (7.0)		1,043 (6.6)	1,959 (6.8)	3,002 (6.7)
Poor	10,126 (31.4)	21,302 (36.2)	31,428 (34.5)		4,932 (31.4)	10,621 (36.6)	15,553 (34.8)
No. of missing values	40 (0.1)	89 (0.2)	129 (0.1)		14 (0.1)	36 (0.1)	50 (0.1)

Values are numbers (percentages) of participants unless stated otherwise. NA, not available; SD, standard deviation.

Missing data

Potential risk predictors with more than 50% missing values were excluded. Multiple imputation was used to impute risk predictors with ≤50% missing values assuming any missing values of the risk predictors were missing at random (49,50). We conducted ten imputations because this has relatively high efficiency, accounting for the large datasets and computing power (50). Rubin’s rules were used to combine coefficients accounting for missing data (50). Imputed values were used for derivation and testing of all the risk prediction models.

Validation and recalibration of the Kotz COPD risk prediction model (17)

We first assessed the performance of the original and recalibrated Kotz risk prediction models (17) using similar risk predictors. We recalibrated both models to survival in the EHS cohort because life expectancy is longer in Hong Kong than in the UK (51). We also replaced a composite measure of deprivation, the Carstairs index (52), with five levels based on seven measures (i.e., income deprivation, employment deprivation, education, skills and training deprivation, health deprivation, disability, crime, barriers to housing and services and living environment deprivation) with an index of deprivation based on the type of housing, education level, and public assistance (i.e., Comprehensive Social Security Assistance) as these are more relevant and contextually specific indices in Hong Kong, especially for older adults (53). Furthermore, due to the relatively low prevalence of asthma (~0.13%, n=175) at study entry, we excluded this risk predictor from the recalibrated Kotz model.

Statistical analysis for development of a new COPD risk prediction model

We secondarily investigated whether a new model to predict COPD risk within the next five years in older Chinese adults (65+ years) performed better than did the Kotz model. We used a sex-specific extended Cox proportional hazards regression to estimate the coefficients for each potential risk predictor in the derivation cohort. Risk predictors were selected using backward elimination as recommended for constructing clinical prediction models (54). The proportional hazards assumption was tested for each potential risk predictor, with additional adjustment for other risk predictors, using Schoenfeld residuals before model development (55). Non-linear relations for continuous risk predictors were detected, and the appropriate relation via fractional polynomial terms was identified based on Martingale residuals (56). Significant interactions between age and the risk predictors were included in the final models. The risk of COPD was calculated using the following equation:

where S0(t) is the baseline survival at time t, which was estimated using the observed survival in the derivation cohort; X and X(t) are respectively time-invariant and time-varying risk predictors. We considered three periods (i.e., 0–4.9, 5.0–9.9, 10+ years) for possible time-varying effects of age and smoking.

Model validation

We assessed the predictive performance of the original Kotz model in the whole EHS cohort and of the recalibrated Kotz model and the new model (EHS-COPD) in the validation cohort. Measures assessing predictive performance [i.e., integrated Brier score (IBS) and explained variance (R2)], discrimination (i.e., D statistic and Harrel’s C statistic), and calibration [i.e., calibration-in-the-large (or calibration intercept) and calibration slope] were computed (57). Of these, IBS was calculated by integrating the Brier score for all the entire follow-up periods, which quantifies the mean squared error of the difference between the predicted and the observed survival probability. The observed survival probability derived from the Kaplan-Meier estimator was considered as the benchmark value (), in which an IBS of zero indicates a perfect model (58). Based on the IBS, we further computed the explained variance using . We calculated the D statistic (which quantifies the prognostic separation between the COPD and non-COPD cases) and Harrell’s C statistic (which quantifies the probability of correct ordering in terms of shorter time to event for the participant with higher predicted risk for a randomly selected pair of participants) to evaluate the discriminative ability, in which a higher value indicates better model discrimination. Calibration was illustrated by comparing the mean 5-year predicted risk with the observed risk. The calibration slope was estimated to evaluate the agreement between predicted and observed risk. A value of 1 for the calibration slope and 0 for the calibration in the large parameter suggest perfect calibration, whereas a value for the calibration slope diverging from 1 indicates a poorer agreement. The performance measures were estimated based on each imputed validation cohort and combined using inverse variance weighting (59).

Risk stratification

Since there is no clear threshold for classifying the risk of developing COPD; we calculated sensitivity, specificity, and positive and negative predicted values of the EHS-COPD model at the top 5%, 10%, and 50% estimated risk of COPD over 5 years to stratify the participants into four groups: high, moderate, mild, and low, as in a previous similar study (60).

All statistical analyses were conducted using R software (version 3.6.3, https://cran.r-project.org/) (61).

Ethical review

The EHS Cohort was conducted with the Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (HKU/HA HKW IRB) ethical approval. This study is an analysis of routinely collected data; informed (non-written) consent was obtained by the participants implicitly agreeing to their data being used for research by using the service. This study conformed to the Declaration of Helsinki (as revised in 2013).

Data availability

Currently, the data are not publicly available; we would welcome collaborations and research proposals.

Results

Overall study population

We identified a total of 136,309 participants aged 65 years or older in the EHS cohort. We excluded 487 (0.36%) participants who had COPD at baseline. We included 135,822 participants (47,912 men and 87,910 women) when assessing the performance of the original and recalibrated Kotz models and also to derive the sex-specific EHS-COPD models. Two-thirds of the participants (32,222 men and 58,911 women) were randomly allocated to the derivation cohort, with the remaining one-third (15,690 men and 28,999 women) allocated to the validation cohort, as shown in .

Figure 1

Flowchart of participants selection for the chronic obstructive pulmonary disease (COPD) risk prediction model based on the Hong Kong’s Elderly Health Services (EHS) Cohort.

Baseline characteristics

shows sex-specific baseline characteristics of the participants in both the derivation and validation cohorts. In the derivation cohort, the mean age was 72.2 years for men and 71.7 years for women. The distribution of type of housing, public assistance, physical activity, family history of cardiovascular diseases, existing hypertension, hospital admission, and recent falls were comparable between men and women. Women were more likely than men to be illiterate or obese, based on Asian BMI cut-offs (62) (BMI 25.0+ kg/m2), and to have poor physical functioning, poor cognitive functioning, depressive symptoms, or poor self-rated health. Furthermore, 59.8% of men and 9.4% of women were ex- or current smokers, while 32.8% of men and 9.9% of women were ever drinkers.

Incidence rates of COPD

shows the number of incidence cases, person-years of follow-up, and incidence rates by sex, age, and observation period. Overall, 3,823 COPD cases were observed, with 2,610 COPD cases in 0.92 million person-years of follow-up in the derivation cohort and 1,213 COPD cases in 0.45 million person-years of follow-up in the validation cohort. The incidence rate of COPD was 2.27 times [95% confidence interval (CI): 2.13–2.42] higher in men than in women, and increased rapidly during the first 2 years of follow-up and decreased afterwards (Figure S1). The original Kotz model had about a 2 times higher crude incidence rate of COPD [5.53 per 1,000 person-years (95% CI: 5.46–5.60)] than that in the EHS cohort (2.78 per 1,000 person-years, 95% CI: 2.69–2.87). The median follow-up time for the derivation cohort was 10.0 years [interquartile range (IQR), 6.3–15.0 years], with 75,543 (82.9%) participants having 5+ years of follow-up and 44,198 (48.5%) participants having 10+ years of follow-up. In the validation cohort, the median follow-up time was 10.0 years (IQR, 6.4–15.0 years), with 37,096 (83.0%) participants having 5+ years of follow-up and 21,890 (49.0%) participants having 10+ years of follow-up.

Table 2

Number of incidence cases, person-years of follow-up, and incidence rate per 1,000 person-years of observation in the derivation and validation cohorts

Characteristics	Derivation cohort				Validation cohort
	Incidence cases	Person years	Incidence rate per 1,000 person-years (95% CI)		Incidence cases	Person years	Incidence rate per 1,000 person-years (95% CI)
Men	1,352	302,230	4.47 (4.24–4.72)		657	148,465	4.43 (4.09–4.78)
Women	1,258	618,752	2.03 (1.92–2.15)		556	305,396	1.82 (1.67–1.98)
Age (years)
65–69	995	405,158	2.46 (2.31–2.61)		462	200,740	2.30 (2.10–2.52)
70–74	874	295,446	2.96 (2.77–3.16)		398	144,575	2.75 (2.49–3.04)
75–79	489	149,370	3.27 (2.99–3.58)		224	72,337	3.10 (2.70–3.53)
80–84	180	53,417	3.37 (2.90–3.90)		100	27,444	3.64 (2.96–4.43)
85+	72	17,591	4.09 (3.20–5.15)		29	8,753	3.31 (2.22–4.76)
Time period (years)
0.0–4.9	2,005	429,709	4.67 (4.46–4.87)		912	210,796	4.33 (4.05–4.62)
5.0–9.9	453	673,190	0.67 (0.61–0.74)		208	331,560	0.63 (0.54–0.72)
10.0+	152	637,778	0.24 (0.20–0.28)		93	315,885	0.29 (0.24–0.36)
Overall	2,610	920,982	2.83 (2.73–2.94)		1,213	453,861	2.67 (2.52–2.83)

CI, confidence interval.

Potential risk predictors

Table S1 shows the coefficients and adjusted hazard ratios for the recalibrated Kotz model for men and women in the derivation cohort. shows the coefficients and adjusted hazard ratios for the EHS-COPD model for men and women in the derivation cohort. The EHS-COPD model included time-varying variables (i.e., age and smoking status) and time-invariant variables (i.e., education level, public assistance, alcohol use, BMI, physical activity, existing hypertension, recent falls in the past 6 months, cognitive function, and self-rated health status). The other potential risk predictors did not meet the inclusion criteria in the final model because of extremely low or high prevalence.

Table 3

Sex-specific adjusted hazard ratios (95% confidence interval) with time-varying effects of age and smoking status for COPD in the derivation cohort

Characteristics	Men			Women
	Coefficients	Adjusted HR (95% CI)*		Coefficients	Adjusted HR (95% CI)*
Age
Time period: 0.0–4.9 years	0.0112	1.01 (1.01–1.02)		0.0081	1.01 (1.00–1.01)
Time period: 5.0–9.9 years	−0.0077	0.99 (0.98–1.00)		−0.0113	0.99 (0.98–1.00)
Time period: 10.0+ years	−0.0319	0.97 (0.95–0.98)		−0.0191	0.98 (0.96–1.00)
Smoking status
Time period: 0.0–4.9 years
Never smoker		Reference			Reference
Ex-smoker	0.4868	1.63 (1.55–1.71)		0.7780	2.18 (2.05–2.31)
Current smoker	1.0053	2.73 (2.59–2.89)		1.0978	3.00 (2.79–3.22)
Time period: 5.0–9.9 years
Never smoker		Reference			Reference
Ex-smoker	0.5285	1.70 (1.54–1.87)		1.0375	2.82 (2.48–3.21)
Current smoker	1.072	2.92 (2.63–3.24)		1.6642	5.28 (4.61–6.04)
Time period: 10.0+ years
Never smoker		Reference			Reference
Ex-smoker	0.5000	1.65 (1.41–1.93)		1.0660	2.90 (2.28–3.71)
Current smoker	1.1787	3.25 (2.76–3.83)		1.4618	4.31 (3.28–5.67)
Public assistance (yes)	0.2302	1.26 (1.20–1.32)		0.2236	1.25 (1.19–1.31)
Alcohol use
Never drinker		Reference
Ex-drinker	0.3786	1.46 (1.39–1.53)		0.0176	1.02 (0.94–1.11)
Social drinker	0.3271	1.39 (1.33–1.45)		0.1105	1.12 (1.05–1.19)
Moderate drinker	0.2116	1.24 (1.16–1.32)		−0.2036	0.82 (0.67–0.99)
Excessive drinker	−0.0786	0.92 (0.84–1.01)		−1.7951	0.17 (0.09–0.31)
Body mass index (kg/m2)
<18.5	0.4515	1.57 (1.47–1.68)		0.4666	1.59 (1.49–1.71)
18.5–<23.0		Reference			Reference
23.0–<25.0	−0.0541	0.95 (0.91–0.99)		−0.1573	0.85 (0.81–0.90)
25.0+	−0.0962	0.91 (0.87–0.95)		0.0110	1.01 (0.97–1.05)
Physical activity (hours per week)
None		Reference			Reference
0.1–<1.5	−0.2346	0.79 (0.74–0.85)		−0.2693	0.76 (0.71–0.83)
1.5–<3.0	−0.0561	0.95 (0.90–1.00)		0.0793	1.08 (1.02–1.15)
3.0–<4.5	−0.8430	0.43 (0.40–0.47)		−0.9286	0.40 (0.36–0.43)
4.5+	−0.2890	0.75 (0.71–0.78)		−0.0592	0.94 (0.89–0.99)
Hypertension (yes)	0.5051	1.66 (1.52–1.81)		0.8540	2.35 (2.17–2.54)
Recent falls (yes)	−0.9162	0.40 (0.30–0.53)		0.2171	1.24 (1.10–1.40)
Cognitive function (poor)	−0.3933	0.67 (0.59–0.77)		−0.2269	0.80 (0.74–0.86)
Self-rated health
Better	0.1113	1.12 (1.04–1.20)		0.0315	1.03 (0.96–1.11)
Normal		Reference			Reference
Poor	0.1925	1.21 (1.17–1.26)		0.1741	1.19 (1.15–1.23)

*, well-established predictors in EHS-COPD are selected with P value <0.05. EHS-COPD, chronic obstructive pulmonary disease (COPD) risk prediction model based on Hong Kong’s Elderly Health Services (EHS) Cohort; CI, confidence interval.

Figures S2,S3 show the smoothed hazard ratios for the time-varying effects of age and smoking status including and excluding education level, public assistance, alcohol use, BMI, physical activity, existing hypertension, recent falls, cognitive function, and self-rated health status in men and women, respectively. For each of these potential risk predictors, the hazard changed over time from study entry.

shows the performance of the original Kotz model, the recalibrated Kotz model, and the EHS-COPD model in predicting the 5-year risk of COPD in men and women. Overall the values of D and C statistics were greater in women than in men except for the C statistic for the recalibrated Kotz model, suggesting that both the original and the recalibrated Kotz model performed relatively better in women than in men. However, the performance of the EHS-COPD model for predicting 5-year risk of COPD was better than those of the original and the recalibrated Kotz models in both sexes, as indicated by the smaller IBS, larger R2, higher D statistic, and the higher Harrell’s C statistic.

Table 4

Performance of sex-specific original Kotz model, recalibrated Kotz model, and the EHS-COPD model for predicting the 5-year risk of COPD in men and women aged 65 years or older in the validation cohort

Statistics	Men	Women
Original Kotz model*
Integrated Brier Score
Benchmark model	0.0367	0.0177
Original Kotz model	0.0364	0.0176
R2 (%)	0.82	0.56
D statistic	1.614 (1.591 to 1.636)	2.220 (2.193 to 2.248)
Harrell’s C	0.603 (0.598 to 0.607)	0.684 (0.677 to 0.692)
Calibration slope	0.391 (0.287 to 0.495)	N.A.
Calibration-in-the-large	0.003 (−0.004 to 0.011)	N.A.
Recalibrated Kotz model**
Integrated Brier Score
Benchmark model	0.0370	0.0164
Recalibrated Kotz model	0.0364	0.0163
R2 (%)	1.62	0.61
D statistic	2.458 (2.420 to 2.497)	2.472 (2.429 to 2.515)
Harrell’s C	0.666 (0.659 to 0.673)	0.657 (0.649 to 0.665)
Calibration slope	0.737 (0.692 to 0.782)	1.437 (1.312 to 1.562)
Calibration-in-the-large	−0.005 (−0.007 to −0.002)	−0.011 (−0.013 to −0.008)
EHS-COPD***
Integrated Brier Score
Benchmark model	0.0370	0.0164
EHS-COPD	0.0298	0.0150
R2 (%)	19.46	8.54
D statistic	23.147 (23.097 to 23.187)	21.420 (21.374 to 21.465)
Harrell’s C	0.930 (0.928 to 0.932)	0.928 (0.925 to 0.930)
Calibration slope	2.126 (1.952 to 2.301)	3.928 (3.663 to 4.193)
Calibration-in-the-large	−0.018 (−0.024 to −0.012)	−0.018 (−0.021 to −0.015)

*, the baseline survival at 5 years was 0.9700 for men and 0.9842 for women in the original Kotz model; **, the baseline survival at 5 years was 0.9703 for men and 0.9835 for women in the recalibrated Kotz model; ***, the baseline survival at 5 years was 0.9720 for men and 0.9845 for women in the EHS-COPD model. N.A., not available. The performance of the original Kotz model was assessed using the whole EHS cohort, whilst those of the recalibrated Kotz and EHS-COPD models were assessed using the validation cohort. Due to the limited predicted risk using the original Kotz model, only 2 risks were calculated, inducing 2 subgroups so that the calibration slope and the calibration-in-the-large could not calculated. EHS-COPD, chronic obstructive pulmonary disease (COPD) risk prediction model based on Hong Kong’s Elderly Health Services (EHS) Cohort; CI, confidence interval.

Compared with the Kaplan-Meier estimator (), the predictive error of IBS for the recalibrated Kotz model was 0.0364 for men. It explained 1.62% of the variation in COPD risk in the validation cohort. The D statistic was 2.458 and Harrell’s C statistic was 0.666. The corresponding values in women were 0.0163, 0.61%, 2.472, and 0.657. The recalibrated Kotz model performed better than the original Kotz model for men, but performed worse for women. In contrast, the newly developed EHS-COPD model outperformed the original and recalibrated Kotz models in both sexes. Similar results were also observed for calibration slope and calibration in the large ().

Furthermore, based on the EHS-COPD risk prediction model, the 5-year predicted risk of COPD was 0.114 for men and 0.042 for women, similar to the observed 5-year risk (i.e., 0.152 for men and 0.073 for women). Figure S4 depicts the observed risk and mean predicted risk of developing COPD at 5 years by 20th predicted risk in the validation cohort for men and women, respectively. The EHS-COPD model tended to underestimate the 5-year risk of developing COPD. Overall, the EHS-COPD model had better performance in men than in women.

shows sensitivity, specificity, positive and negative predictive values for the 5-year risk of COPD in the validation cohort. Table S2 shows the characteristics of participants in the validation cohort classified into the following four groups according to EHS-COPD model:

Table 5

Sensitivity, specificity, and positive and negative predictive values for death at different thresholds of predicted risk of chronic obstructive pulmonary disease over 5 years in the validation cohort among both men and women

Threshold	Risk threshold (%)	True-positive count	False-positive count	False-negative count	True-negative count	Sensitivity (%)	Specificity (%)	Positive predictive value (%)	Negative predictive value (%)
Top 1%	0.1434	112	336	800	43,440	12.31	99.23	25.07	98.19
Top 2%	0.1057	213	676	699	43,100	23.34	98.46	23.93	98.40
Top 3%	0.0853	308	1,006	604	42,770	33.81	97.70	23.45	98.61
Top 4%	0.0722	391	1,334	521	42,442	42.83	96.95	22.65	98.79
Top 5%	0.0626	446	1,662	466	42,114	48.88	96.21	21.15	98.91
Top 6%	0.0544	497	1,987	415	41,789	54.49	95.46	20.02	99.02
Top 7%	0.0473	541	2,264	371	41,512	59.38	94.83	19.30	99.11
Top 8%	0.0418	564	2,532	348	41,244	61.83	94.22	18.21	99.16
Top 9%	0.0380	590	2,793	322	40,983	64.75	93.62	17.46	99.22
Top 10%	0.0348	616	3,058	296	40,718	67.56	93.02	16.79	99.28
Top 15%	0.0254	782	4,341	130	39,435	85.83	90.08	15.28	99.67
Top 20%	0.0199	871	5,437	41	38,339	95.54	87.58	13.81	99.89
Top 30%	0.0112	907	6,161	5	37,615	99.45	85.93	12.83	99.99
Top 40%	0.0074	912	6,646	0	37,130	100	84.82	12.07	100
Top 50%	0.0057	912	6,671	0	37,105	100	84.76	12.03	100

❖ The high-risk group included 2,237 participants (i.e., 5.0% of 44,688), in the top 5% for risk of COPD in the next 5 years;

❖ The moderate-risk group included 2236 participants (i.e., 5.0% of 44,688), in the next 10% for risk of COPD in the next 5 years;

❖ The mild-risk group included 18,028 participants (i.e., 40.3% of 44,688), in next 50% for risk of COPD in the next 5 years;

❖ The low-risk group included the remaining 22,187 participants (i.e., 49.6% of 44,688).

In the high-risk group, the COPD incidence rate was 65.0 per 1,000 person-years, and the mean age was 74.6 years. In addition, 88.1% were men, 92.2% were either ex- or current smokers, 27.1% had public assistance, 14.2% were illiterate, 40.9% were ever drinkers, 15.7% had low BMI, 21.6% had no physical activity per week, 2.1% had existing hypertension, 0.8% had had 2 or more falls in the last 6 months, 4.0% had poor cognitive function, and 43.1% had poor self-rated health in comparison with the previous year.

Clinical examples

Table S3 shows four hypothetical participants as examples to illustrate how to use the EHS-COPD model.

Discussion

Principal findings

We assessed the performance of the original and the recalibrated Kotz models for predicting the risk of developing COPD within the next 5 years in older Chinese adults (65+ years) in Hong Kong. Neither model predicted the risk of developing COPD well. A newly developed EHS-COPD risk prediction model had a better performance in an independent validation cohort. It explained 19.5% and 8.5% of COPD risk for men and women, respectively. Moreover, the EHS-COPD model yielded an excellent overall performance, which outperformed the original and recalibrated Kotz models, especially for men.

Comparison with other studies

Incidence rate

The crude incidence rate of COPD, i.e., 2.78 per 1,000 person-years (95% CI: 2.69–2.87), in the EHS cohort (Figure S1) was lower than that in some previous studies (17,33) but comparable to that in others (63).

Risk predictors

The original Kotz model is largely driven by age (17), but in the EHS cohort, age did not predict COPD in men or women (Table S1). Ex- or current smoking did not predict COPD as strongly in the EHS cohort as in the original Kotz model, possibly due to the small number of cigarettes smoked per day in Hong Kong and the existence of other context-specific causes of COPD (), such as incense burning (64,65). As previously, low BMI, low socioeconomic position, and ever drinking all predicted COPD (8,32,37-44,66,67). Several markers of ill-health, such as poor cognitive function, recent falls, and self-rated health status, also predicted COPD. However, asthma did not predict COPD, possibly because of the relatively low prevalence of asthma among older adults in this setting (68).

Other risk prediction models

A recent review (27) identified 4 risk prediction models for COPD. However, one of them was designed for use in out-patients, and 3 of them were developed in Western settings.

Strengths and limitations

The COPD risk prediction model developed here has the advantages of incorporating time-varying risks of age and smoking status, a large sample with long follow-up, a population-based cohort, and minimal recall and response bias. Nevertheless, limitations exist. First, self-reported COPD cases were not validated against medical records or spirometry, but measurement errors should be random. Second, participants in the EHS cohort were volunteers, so they may not represent the entire population, limiting its applicability to those who are home-bound. Third, for feasibility, the model only included easy-to-assess risk predictors, but did not include other potentially significant COPD risk predictors that are unlikely to be available in a clinical setting, such as indoor/outdoor air pollution (31,69), occupational exposures (e.g., dust, chemical agents, and fumes) (70), genetic predisposition (71,72), measures of lung function (72), asthma and airway hyper-reactivity (73,74), and history of severe childhood respiratory infections (75,76). The model explained less variation for women (i.e., ~8.5%) than it did for men, possibly because women had low smoking prevalence. Lastly, recalibration of the EHS-COPD model might be needed for use in all Chinese adults.

Conclusions

An existing COPD prediction model, the Kotz model, had low discrimination and calibration in older Chinese adults (65+ years) in Hong Kong. However, a revised model, EHS-COPD, had excellent performance compared to the original and recalibrated Kotz models.

The article’s supplementary files as

Box 1 Potential risk predictors for COPD considered for inclusion in the COPD risk prediction model

Risk predictors in the relevant literature

Age (continuous variable) (8,17-19,27-35)

Smoking status (never, ex-, or current smoker) (8,17-20,27-32,34-37)

BMI (<18.5 kg/m2, 18.5–<23.0 kg/m2, 23.0–<25.0 kg/m2, 25.0+ kg/m2) (20,27,34,37-40)

Existing asthma at study entry (17,28)

Physical activity (none, 0.1–<1.5 h/week, 1.5–<3.0 h/week, 3.0–<4.5 h/week, 4.5+ h/week) (37,41-44)

Recent falls (2 or more falls in the past 6 months) (45,46)

Cognitive function (poor functioning defined as mini-mental state examination score ≤18 [illiterate]/20 [1–2 years schooling]/22 [2+ years schooling] or abbreviated mental test score ≤7 or simplified memory test without correctly answering all three questions) (47)

Existing hypertension at study entry (48)

Alcohol use (never, ex-, social, moderate, or excessive drinker) (37)

Potential risk predictors available in the EHS cohort

Education level (illiterate, literate but no formal education, primary, secondary or above)

Family history of cardiovascular disease in either first- or second-degree relatives

Hospital admission in the past 12 months

Physical functioning [poor functioning defined as ADL >8, or IADL ≥6]

Depressive symptoms (geriatric depression scale ≥8)

Type of housing (public or aided housing, private rented housing, private self-owned housing, others)

Public assistance (in receipt of Comprehensive Social Security Assistance)

Self-rated health (health condition this year compared with last year)

COPD, chronic obstructive pulmonary disease; BMI, body mass index; EHS, Elderly Health Services; ADL, activities of daily living score; IADL, instrumental activities of daily living score.