Stratification of telehealthcare for patients with chronic obstructive pulmonary disease using a predictive algorithm as decision support: A pilot study

Abstract

Introduction

The number of patients needing care who suffer from chronic obstructive pulmonary disease (COPD) is expected to increase in the future. The consequences thereof will increase the socio-economic burden for both patients and society. Telehealthcare technologies have shown potential in reducing hospitalisation-related costs and in improving health-related quality of life (HRQOL) for some COPD patients, but not all. The aim of this study was to investigate the potential of predictive algorithms for helping the general practitioner to stratify telehealthcare for COPD patients in a way that maximises HRQOL and minimises COPD-related costs.

Methods

Data from 553 COPD patients based in the North Denmark Region were analysed and used as predictors for four multiple linear regression models. The models were trained and evaluated for their abilities to predict individual patient’s future health- and cost-related developments, with and without telehealthcare.

Results

The average root-mean-square error (RMSE) of the health and cost models was 5.265 HRQOL scores and US dollars (US$)5430.49, respectively. The accuracy regarding the polarity of the predicted changes ranged from 61–65% for the health models and 74–75% for the cost models. While differences in the magnitude of predictions with and without telehealthcare were statistically significant (p < 0.01), the polarity of predictions was similar across models in 82.05% of all cases.

Discussion

Our results indicate that it may be possible to predict the magnitude and polarity of a COPD patient’s future health- and cost-related developments with and without telehealthcare. Predictive algorithms may provide a useful decision support tool in stratifying telehealthcare for COPD patients.

Keywords

Chronic obstructive pulmonary disease TeleCare North trial Short Form-36 Health Survey telemedicine linear regression model

Introduction

Chronic obstructive pulmonary disease (COPD) is an umbrella term used to describe chronic lung diseases causing airflow limitations in the lungs.¹ Approximately 64 million people suffer from COPD worldwide, and the disease is a major cause of morbidity and mortality.^2–4 Despite a decreasing trend in mortality,⁵ COPD caused 3.1 million deaths in 2012, making it the third most common cause of death that year.^2,6 In Denmark, COPD affects 430,000 people and is the direct cause of 3,500 deaths and 23,000 hospitalisations annually.^7–9 The large number of deaths and hospitalisations puts substantial socio-economic pressure on both patients and healthcare systems.^7,10,11 In fact, currently, 10% (approximately US$440 m) of the total annual Danish healthcare budget for citizens over 40 years is related to the care and treatment of COPD.⁸ The prevalence of and socio-economic pressure exerted by COPD is expected to increase as a result of mortality reduction related to improved treatment and a generally increasing incidence of disease.^7,10,11

Home support technologies have been investigated as a way to address the expected increase in costs generated by COPD¹⁰ and the increasing need for home care services.¹² According to a review by Paré et al.,¹⁰ nine of 10 recent studies have reported a statistically significant decrease in the number of hospitalisations when applying home telehealthcare compared with conventional care. Several studies have also reported a decrease in total COPD-related costs.¹⁰ Despite these positive effects, not all COPD patients seem to benefit from telehealthcare, which is why an in-depth screening of patients should be conducted before offering such a service.⁶ To our knowledge, no existing studies have addressed the issue of stratifying telehealthcare for COPD patients.

The aim of this present study was to investigate whether it is possible to use predictive algorithms to help stratify telehealthcare for COPD patients in a way that maximises health-related quality of life (HRQOL) and minimises the hospitalisation costs of the individual patient.

Methods

Data

The data for this study were obtained from 1,225 COPD patients included in the TeleCare North trial, which was a large-scale, cluster-randomised telehealthcare randomised controlled trial (RCT) conducted from 2012–2015 as a cross-sectional collaboration between Aalborg University and general practitioners (GPs), as well as the North Denmark Region and its 11 municipalities.^13–15

The patients were divided into an intervention group that received the intervention and a control group that did not (see Table 1).^15,16 Prior to the study, each patient answered a health-related questionnaire at home and had several physiological parameters, which served as baseline values, measured by their affiliated GPs. During the study, the intervention group performed weekly home measurements of oxygen saturation, blood pressure, pulse, and weight and answered questions related to COPD exacerbation symptoms. The patient data was monitored asynchronously by a nurse. At the end of the study (12-month follow-up), all patients were asked to complete the health-related questionnaire again.

Table 1.

Characteristics of the patients.

Variable	Intervention group	Control group
Number	263	290
Age (years)	69.09 ± 9.02	70.05 ± 8.52
BMI	26.46 ± 4.87	26.55 ± 7.08
Gender
Male (n, col %)	130 (49.45%)	132 (45.52%)
Female (n, col %)	133 (50.55%)	158 (54.48%)
HRQOL score	43.08 ± 7.74	43.36 ± 7.76
Hospitalisation cost 2013 (US$)^a	3,800.55 (7226.48)^b	3,187.26 (5788.43)^b
FEV₁ (%)	49.27 ± 17.42	48.18 ± 21.91
FVC (%)	72.56 ± 18.20	73.03 ± 23.08
FEV₁/ FVC (%)	56.05 ± 17.19	53.22 ± 14.21
GOLD classification
Stage 1 (n)	19 (7.22%)	12 (4.14%)
Stage 2 (n)	148 (56.27%)	154 (53.10%)
Stage 3 (n)	85 (32.32%)	116 (40.00%)
Stage 4 (n)	11 (4.18%)	8 (2.76%)

BMI: body mass index; HRQOL: health-related quality of life; US$: US dollars; FEV1: forced expiratory volume in one second; FVC: forced vital capacity; GOLD: Global Initiative for Chronic Obstructive Lung Disease.

Only patients with costs in 2013; ^bmedian and interquartile range (IQR).

The data was extracted from the TeleCare North trial database and the Danish National Patient Register (DNPR) based on patient social security number. Each patient had a set of 39 predictors, of which 11 were physiological measurements, 24 were extracted from the questionnaire, and four were extracted from the DNPR.

The physiological parameters consisted of systolic, diastolic and mean arterial blood pressure, pulse, body mass index (BMI), spirometry measures (percentage of expected forced expiratory volume in one second (FEV₁ (%)), FEV₁ measured in litres (FEV₁L (l)), percentage of expected forced vital capacity (FVC (%)), FVC measured in litres (FVCL (l)), FEV₁/FVC), and oxygen saturation.

The questionnaire was divided into two sections: (a) the standardised European Quality of Life-5 Dimensions questionnaire (EQ-5D),¹⁷ which consists of medical- and COPD-related questions, such as questions concerning the presence of comorbidities (e.g. diabetes mellitus type 2, cardiovascular disease, mental illness, cancer and disease affecting muscles, bones or joints), years since COPD diagnosis and smoker/non-smoker status; and (b) the Short Form-36 Health Survey (SF-36) questionnaire,^18,19 which is a well-documented and validated patient-reported survey consisting of 36 questions that are computed into eight subscales; vitality, physical functioning, bodily pain, general health perceptions, physical role functioning, emotional role functioning, social role functioning, and mental health. A mental and a physical component summary (MCS and PCS) score are computed from the eight subscales. Henceforth, the notion of HRQOL equals the average of MCS and PCS of each patient on a scale from 0–100. The lower the score, the more disability is present. Data from the DNPR provided information on the number and total cost of hospitalisations during 2013 and 2014.

Preprocessing

Baseline BMI data were not available for all patients because height information was not reported at baseline for 18.01% of the patients. Heights reported at the 12-month follow-up were imported, and any height values still missing were subsequently imputed using single imputation. Approximately 34% of the spirometry parameters were missing. These were computed using pre-existing parameters.²⁰ In cases in which the GP did not report any oxygen saturation measurements prior to the study, the mean of the first 10 home measurements for the individual intervention patient was calculated and used as a baseline substitute. Oxygen saturation was removed as a predictor for the control patients due to a large proportion (80.22%) of missing values.

Patients who did not answer the 12-month follow-up questionnaire (intervention group: n = 126, control group: n = 164) and patients who had more than four missing predictors (intervention group: n = 49, control group: n = 183) were excluded. Patients with excessively large hospitalisation costs (intervention group: n = 9, control group: n = 10) were also excluded due to high-cost treatments not related to COPD. Patients who performed less than 10 home measurements (intervention group: n = 131, control group: n = 0) were excluded due to an inadequate use of the intervention. A total of 553 patients (intervention group: n = 263, control group: n = 290) were included in this study (see Table 1). Any missing values from the remaining patients were imputed using single imputation, as the proportion of missing values amounted to only 0.97%.

Regression models

To predict COPD patient HRQOL and hospitalisation cost changes with and without telehealthcare, four multiple linear regression models were developed (see Figure 1). The predictors were used to train the models to predict health and cost changes with telehealthcare for the patients in the intervention group, whereas the same were used to train the models to predict these changes without telehealthcare for the control group. Only patients with hospitalisation costs throughout the baseline year (2013) and the intervention period (2014) were included in the development and evaluation of the cost models.

Figure 1.

Illustration of the development of the four multiple linear regression models and an example of how their outputs may be implemented in a decision support tool. COPD: chronic obstructive pulmonary disease; QOL: quality of life.

Each model was trained to fit the data to a criterion variable using the 12 predictors with the largest correlation with the criterion variable. All possible combinations of the 12 predictors (n = 4,095) for each of the four models were trained on 70%, validated on 15%, and tested on the remaining 15% of the data using five-fold cross-validation. Based on the validation process, the best-performing models were chosen, and the performance metrics from the test process were reported. The performances of the models were evaluated based on the spread of the differences between predicted and criterion values (residuals). For each of the four types of predictions, the model resulting in the lowest residual spread was chosen.

The agreement of the models (H_I vs H_C and C_I vs C_C) regarding the polarity of the predicted changes was assessed by applying each model to the counterpart’s test data (e.g. using H_I to predict changes based on the test data of H_C). Thereby, the predictions of the two health and the two cost models were applied to the same sets of patients, which allowed for direct comparison. The degree of agreement was calculated, and statistical analysis was used to assess differences in predictions.

Statistical analysis

The data sets were exported into SPSS (IBM SPSS Statistics). The Shapiro-Wilk test revealed non-normal distributions in the predicted values of each of the four models. Therefore, the medians of the predictions of the health models (H_I and H_C) and the cost models (C_I and C_C) were compared using the paired-samples sign test.

Results

The four best performing multiple linear regression models were based on seven to nine predictor variables (see Table 2), including both demographic, DNPR, SF-36 and spirometry information. Furthermore, all the models included a constant that was calculated during the development of each model. This is called the intercept and describes where the fitted regression line crosses the y-axis. The two health models consisted primarily of SF-36 and spirometry predictors, whereas the two cost models consisted primarily of DNPR predictors combined with SF-36, spirometry, and demographic predictors.

Table 2.

The predictors of the four multiple linear regression models.

H_I	C_I	H_C	C_C
Intercept	Intercept	Intercept	Intercept
Mental health	Hospitalisation cost in 2013	Social role functioning	Hospitalisation cost in 2013
Bodily pain	Municipal share of hospitalisation cost in 2013	Mental health	Municipal share of hospitalisation cost in 2013
Emotional role functioning	Number of hospitalisation days in 2013	Bodily pain	Number of hospitalisation days in 2013
Vitality	Presence of disease in muscles, bones or joints	Physical role functioning	FVC
FEV₁	Social role functioning	Vitality	Number of people in household
Baseline MCS	Gender	Age	Civil status
		General health perceptions
		Educational level
		FEV₁/FVC-ratio

FEV₁: forced expiratory volume in one second; FVC: forced vital capacity; MCS: mental component summary.

Based on the five-fold cross-validation of each model, model RMSE and accuracy are reported using the average and standard deviation (SD) of the corresponding five underlying measures. On average, the predicted health changes diverged by 5.27 ± 0.17 HRQOL scores from the value of the criterion variables (see Table 3). The predicted cost changes diverged, on average, by US$5430.49 ± 393.53 from the value of the criterion variables.

Table 3.

Overview of the performance of the four models. All metrics are provided as the mean and standard deviation (SD) of a five-fold cross-validation. The accuracies presented represent the models ability to correctly predict an increase or decrease.

Parameter	H_I	C_I	H_C	C_C
Adjusted R²	0.13 ± 0.03	0.50 ± 0.08	0.09 ± 0.02	0.512 ± 0.075
RMSE	5.41 ± 0.12^a	5055.33 ± 388.86^b	5.12 ± 0.21	5805.65 ± 398.19^b
Mean of predictions	–0.13 ± 0.42^a	–681.52 ± 2268.87^b	–1.32 ± 0.32^a	1946.86 ± 1557.55^b
SD of predictions	2.63 ± 0.47^a	7733.08 ± 2600.23^b	2.19 ± 0.33^a	8188.78 ± 3578.47^b
Mean of criterion variables	–0.05 ± 0.46^a	–1619.27 ± 2186.23^b	–1.29 ± 0.56^a	381.33 ± 2345.65^b
SD of criterion variables	6.02 ± 0.71^a	9597.14 ± 2745.65^b	4.83 ± 0.68^a	11,112.68 ± 5136.62^b
Mean of residuals	0.07 ± 0.64^a	–1309.30 ± 1942.04^b	0.03 ± 0.43^a	–1106.21 ± 995.11^b
SD of residuals	5.51 ± 0.66^a	6337.37 ± 1534.51^b	4.66 ± 0.35^a	6906.89 ± 1260.82^b
True increases	10.80 ± 12.50	5.80 ± 2.28	6.40 ± 9.92	8.00 ± 2.35
False increases	5.40 ± 6.11	2.80 ± 2.17	4.80 ± 7.46	3.40 ± 0.55
True decreases	13.20 ± 16.72	7.60 ± 1.95	22.80 ± 29.76	5.20 ± 2.05
False decreases	9.60 ± 11.99	1.80 ± 0.84	10.00 ± 11.25	1.00 ± 1.00
Accuracy	0.61 ± 0.07	0.74 ± 0.11	0.65 ± 0.03	0.75 ± 0.05

RMSE: root-mean-square error; SD: standard deviation; R2: coefficient of determination.

HRQOL scores^{; b}US dollars.

The accuracies represented the models ability to predict the correct polarity of the change from baseline to the end of the study (i.e. an increase or decrease in HRQOL or cost). The accuracies of H_I, C_I, H_C, and C_C were 61.25 ± 0.07%, 74.44 ± 0.11%, 64.62 ± 0.03%, and 75.03 ± 0.05%, respectively (see Table 3). The average overall probability that a predicted increase actually was an increase (positive predictive value (PPV)) was 65.02 ± 20.00%, and the average overall probability that a predicted decrease was actually a decrease (negative predictive value (NPV)) was 72.36 ± 10.12%. In general, the cost models performed better than the health models.

By applying the intervention models (H_I and C_I) to the data from the corresponding control models (H_C and C_C, respectively), a comparable measure of model effects was obtained. Statistically significant differences in medians were found when comparing predictions of H_I and H_C based on the same data (p < 0.01). C_I and C_C revealed statistically significant differences in medians when applying C_C to the C_I test data (p < 0.01), but the reverse was not true (p = 0.171). The average median absolute differences of the predictions of H_I and H_C on the counterpart’s test data was 1.12 (interquartile range (IQR): 1.51) and 1.37 (IQR: 1.54) HRQOL scores, respectively. For C_I and C_C, these metrics were US$1605.40 (IQR: US$1901.85) and US$1874.49 (IQR: US$1990.03), respectively. In 76.87% of all cases, the H_I and H_C predict HRQOL changes with the same polarity, while C_I and C_C predict cost changes with the same polarity in 87.23% of all cases.

Discussion

The aim of this study was to investigate whether the prediction of future health- and cost-related changes for a specific COPD patient, with and without telehealthcare, was possible. Polarity accuracies ranging from 61.25 ± 0.07% to 75.03 ± 0.05% indicate that it is possible to predict the direction in which a COPD patient’s future HRQOL and hospitalisation cost will develop. However, RMSEs of 5.27 ± 0.17 HRQOL scores and US$5430.49 ± 393.53 for the health and cost models, respectively, suggest that it is difficult to predict the exact magnitudes of the changes. When applying both health and cost models to the same patients, the models predict changes in the same direction in 76.87% and 87.23% of the cases, respectively. The health models predict changes with statistically significant different median values when applied to the test data of both H_I and H_C, while the cost models only predict changes with statistically significant different median values when applied to the test data of C_C.

The spread of predictions was 2.63 ± 0.47 and 2.19 ± 0.33 HRQOL scores for H_I and H_C, respectively, which was markedly lower than the spread of the criterion values (see Table 3). For C_I and C_C, the spreads were US$7733.08 ± 2268.87 and US$8188.78 ± 3578.47, respectively, which were also lower than the spread of the criterion values. The lower spread in the predicted values compared with the observed values indicated a higher density, suggesting that some degree of predictive potential is present. This is in line with all the models having accuracies and both positive and negative predictive values well above 50.00%.

As expected, the health models contained predictors concerning MCS, PCS, health status and limitations, whereas the cost models contained predictors concerning total hospitalisation days and cost in 2013. The predictors used within the two health and two cost models differ by some degree (see Table 2), which is interesting as the cluster-randomised study design¹⁴ implies that the physiological and demographic properties of the groups are homogenous. An interesting difference between the models with and without telehealthcare was the overall type of their respective predictors. H_I and C_I consisted mostly of predictors directly related to either mental or physical health, whereas H_C and C_C, to a larger degree, consisted of socially related predictors, such as social function, educational level, number of people in the household and civil status. This indicates that social factors affect HRQOL and hospitalisation cost to a larger degree for patients who did not receive the telehealthcare intervention.

FEV₁ has also been found to predict changes in HRQOL,²¹ which is in line with the inclusion of FEV₁ in H_I and possibly the FEV₁/FVC-ratio in H_C. A systematic review regarding the risk factors of hospitalisation and readmission found that previous hospitalisations, low FEV₁, comorbidity, impaired health status and gender were risk factors for hospitalisation. Living alone, length of hospitalisation and physical activity were risk factors for readmission.^22,23 This may explain the inclusion of hospitalisation cost and number of hospitalisation days in 2013, presence of disease in muscles, bones or joints, gender, FVC, number of people in the household and civil status as predictors in the two cost models.

The models were developed based on predictors from 553 COPD patients. The set of predictors for each model was chosen based on the correlation with the corresponding criterion value, and the feature combination resulting in the lowest residual spread was chosen. Other methods, including both forward and backward elimination of predictors, did not improve model performance despite the use of different inclusion and exclusion criteria.

A reassessment of the exclusion criteria would most likely result in a larger patient population, e.g. imputing missing information for patients with more than four missing predictors and following an intention-to-treat approach would result in a population of 916 patients; an increase of 65.64%. The use of other types of models, e.g. non-linear regression models, may also improve model performance and allow a clear distinction between the changes with and without telehealthcare.

When applied to the same data set, the two health and two cost models predicted changes with the same polarity in 76.87% and 87.23% of all cases, respectively. In combination with the statistically significant differences between the medians of both of the two pairs of models, these results indicate that the HRQOL and hospitalisation cost developments are predictable and that there are differences in developments with and without telehealthcare. However, the large prediction variability does complicate the distinction of changes with and without telehealthcare.

There are several limitations to our study. The high loss-to-follow-up might have been prevented by revising the exclusion criteria. The linear regression approach used in this study provides a tool to predict the direction in which a patient will develop; however, more complex approaches (e.g. non-linear regression models) might provide more accurate predictions of the magnitude of the changes.

In conclusion, our findings indicate that it is possible to predict whether the future HRQOL score and cost for a COPD patient will increase or decrease within the following year, with or without telehealthcare. Furthermore, our results suggest that an optimisation of the models and predictor selection may permit clear distinctions between developments with and without telehealthcare, and thereby provide useful decision support to GPs in stratifying telehealthcare for COPD patients. Therefore, we suggest that more research in this area should be conducted.

Footnotes

Acknowledgements

The authors wish to thank the COPD patients participating in the TeleCare North trial, as well as the clinical and industrial partners contributing to the completion of the trial. For further details and funding information, see .

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

References

World Health Organization. Chronic obstructive pulmonary disease (COPD), www.who.int/respiratory/copd/en/ (2015, accessed 5 August 2015).

World Health Organization. The top 10 causes of death, www.who.int/mediacentre/factsheets/fs310/en/ (2014, accessed 3 August 2015).

Shavelle

Paculdo

Kush

. Life expectancy and years of life lost in chronic obstructive pulmonary disease: Findings from the NHANES III follow-up study. Int J Chron Obstruct Pulmon Dis 2009; 4: 137–148.

Rabe

Hurd

Anzueto

. Global strategy for the diagnosis, management, and prevention of chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2007; 176: 532–555.

López-Campos

Ruiz-Ramos

Soriano

. Mortality trends in chronic obstructive pulmonary disease in Europe, 1994–2010: A joinpoint regression analysis. Lancet Respir Med 2014; 2: 54–62.

Dinesen

Huniche

Toft

. Attitudes of COPD patients towards tele-rehabilitation: A cross-sector case study. Int J Environ Res Public Health 2013; 10: 6184–6198.

Laursen

Ravn

Sygdomme i thorax. In: Schroeder

Schulze

Hilsted

(eds). Basisbog i Medicin og Kirurgi, 5th ed. Copenhagen: Munksgaard Danmark, 2012, pp. 401–408.

Danmarks Lungeforening. Hvor hyppig er KOL?, www.lunge.dk/hvor-hyppig-er-kol (2015, accessed 4 August 2015).

Juel K and Døssing M. KOL i Danmark: Sygdommen der hver dag koster 10 danskere livet. Report. Copenhagen: The Danish National Institute of Public Health, April 2003.

10.

Paré

Poba-Nzaou

Sicotte

. Comparing the costs of home telemonitoring and usual care of chronic obstructive pulmonary disease patients: A randomized controlled trial. European Research in Telemedicine 2013; 2: 35–47.

11.

Fernández-Granero

Sánchez-Morillo

Léon-Jiménez

. Automatic prediction of chronic obstructive pulmonary disease exacerbations through home telemonitoring of symptoms. Bio-Med Mater Eng 2014; 24: 3825–3832.

12.

Polisena

Tran

Cimon

. Home telehealth for chronic obstructive pulmonary disease: A systematic review and meta-analysis. J Telemed Telecare 2010; 16: 120–127.

13.

TeleCare Nord KOL. Business case TeleCare Nord, www.rn.dk/∼/media/Rn_dk/Sundhed/Til%20sundhedsfaglige%20og%20samarbejdspartnere/TelecareNord/KOL/Om%20projektet/TeleCare%20Nord%20Business%20Case.ashx (2012, accessed 5 August 2015).

14.

TeleCare Nord KOL. Forskningsdesign i TeleCare Nord, www.rn.dk/Sundhed/Til-sundhedsfaglige-og-samarbejdspartnere/TeleCare-Nord/Evaluering-og-forskning/∼/media/Rn_dk/Sundhed/Til%20sundhedsfaglige%20og%20samarbejdspartnere/TelecareNord/KOL/Evaluering%20og%20forskning/Evaluerings-%20og%20forskningsdesign.ashx (2012, accessed 5 August 2015).

15.

Udsen

Lilholt

Hejlesen

. Effectiveness and cost-effectiveness of telehealthcare for chronic obstructive pulmonary disease: Study protocol for a cluster randomized controlled trial. Trials 2014; 15: 178–178.

16.

Lilholt

Heiden

Hejlesen

User satisfaction and experience with a telehealth system for the Danish TeleCare North Trial: A think-aloud study. In: Lovis

Séroussi

Hasman

(eds). eHealth – For Continuity of Care 2014; vol. 205, European Federation for Medical Informatics and IOS Press, pp. 900–904.

17.

EuroQol Research Foundation. EQ-5D, www.euroqol.org/about-eq-5d.html (2015, accessed 20 October 2015).

18.

Bjørner

Damsgaard

Watt

. Dansk manual til SF-36: Et spørgeskema om helbredsstatus, 1st ed. Denmark, Copenhagen: Lægemidddelindustriforeningen, 1997.

19.

Ware JE. SF-36® Health Survey Update, www.sf-36.org/tools/SF36.shtml (2015, accessed 21 October 2015).

20.

Brorson S, Gorlén T, Heebøll-Nielsen NC, et al. KOL i almen praksis – diagnostik, behandling, opfølgning, rehabilitering. Report. Copenhagen: Dansk Selskab for Almen Medicin, 2008.

21.

Hesselink

Van Der Windt

DAWM

Penninx

BWJH

. What predicts change in pulmonary function and quality of life in asthma or COPD? J Asthma 2006; 43: 513–519.

22.

Bahadori

Fitzgerald

. Risk factors of hospitalization and readmission of patients with COPD exacerbation – systematic review. Int J Chron Obstruct Pulmon 2007; 3: 241–251.

23.

Niewoehner

Lokhnygina

Rice

. Risk indexes for exacerbations and hospitalizations due to COPD. Chest 2007; 131: 20–28.