Chronic obstructive pulmonary disease (COPD) is a leading chronic condition in many countries [1]. Patients can experience exacerbation of COPD (eCOPD) that often requires assessment in an emergency department (ED) and hospitalization. Exacerbations play a major role in the burden of COPD, its evolution, and its cost [2, 3]. Some exacerbations can be severe, leading to death or the need for intubation, whereas others are more moderate, requiring little more than an adjustment of the patient’s current medical therapy. Currently, ED physicians must rely largely on their experience and the patient's personal criteria for gauging how an eCOPD will evolve. A clinical prediction rule that could help predict eCOPD evolution would allow ED physicians to make better informed decisions about treatment.

A couple of severity scores have been developed for patients with stable COPD [4, 5], but these do not apply to patients with eCOPD. Several studies have attempted to develop severity scores for eCOPD, but none has been adopted by clinicians, largely because they do not include important clinical variables or they have other methodological problems [6–10]. Several key issues affect the development of severity scores for patients with eCOPD being evaluated in an ED. To date, most such studies have included only patients who are admitted to the hospital, which excludes a significant percentage of patients with eCOPD [11, 12]. Another limitation is the choice of reliable predictors that can easily be gathered in the ED. For a clinical prediction rule to be adopted, its development must follow strict methodological norms and must be based on easily available parameters [13, 14].

The goal of this study was to develop a clinical prediction rule for short-term death following eCOPD. We defined 'short term 'as any time during the hospital admission or within 1 week after discharge from the ED to home. Data for developing this rule were collected from a large prospective cohort of patients with eCOPD attending a number of different EDs.

This prospective cohort study covered 16 hospitals belonging to the Spanish National Health Service (Hospital Costa del Sol, Hospital Valme, Hospital de Motril, Corporació Sanitaria Parc Taulí, Hospital del Mar, Hospital Universitario de La Princesa, Hospital Universitario Gregorio Marañón, Hospital Universitario La Paz, Hospital de Móstoles, Hospital Marqués de Valdecilla, Hospital Santa Marina, Hospital San Eloy, Hospital Galdakao-Usansolo, Hospital Txagorritxu, Complejo Hospitalario Donostia, and Hospital Cruces). The Institutional Review Boards of the participating hospitals approved this project. Patients with eCOPD attending the EDs of any of these hospitals were informed of the goals of the study, and invited to voluntarily participate and sign an informed consent form. All information was kept confidential. Recruitment started in June 2008 and ended in September 2010. A description of the study protocol was published previously [15].

Patients were eligible for the study if they presented to the ED with symptoms consistent with eCOPD. COPD was confirmed if the patient had a forced expiratory volume in 1 second/forced vital capacity (FEV₁/FVC) quotient of less than 70%. Exacerbation was defined as an event in the natural course of the disease characterized by a change in the patient’s baseline dyspnea, cough, and/or sputum that was beyond normal day to day variations and may have warranted a change in regular medication in a patient with underlying COPD [16]. For cases of COPD newly diagnosed in the ED to be included in the study, they had to be confirmed by spirometry within 60 days after the index episode at a time when the patient was stable [17]. Patients were excluded from the study if, at the time they were seeing at the ED, they had eCOPD complicated by a comorbidity such as pneumonia, pneumothorax, pulmonary embolism, lung cancer, or left cardiac insufficiency. Other exclusion criteria included a diagnosis of asthma, extensive bronchiectasis, sequelae of tuberculosis, pleural thickening, or restrictive disease. Patients who did not wish to participate were also excluded.

A total of 3,276 episodes were assesses for the study. Of these, 198 (6%) were excluded because COPD was complicated by other major pathologies at the time of ED admission (cardiovascular conditions: 59 (29.8%); pneumonia: 55 (27.8%); cancer: 21 (10.6%); other respiratory problems, 13 (6.6%); and other conditions, 50 (25.2%)). A further 56 episodes were excluded because the COPD diagnosis was not confirmed by spirometry within 60 days of the index episode. Another 145 episodes were lost because of incomplete data without the possibility of retrieving the information needed for the study. Finally, 390 episodes were excluded because the patient had more than one ED visit. No differences were observed in mortality (4.4% versus. 2.7%; P = 0.32) or other parameters, such as arterial blood gas as pH upon ED arrival (P = 0.52) between episodes excluded because patients had multiple eCOPD episodes and episodes included from patients who had only a single eCOPD episode.

The final population included 2,487 patients. Of these, 1,537 (61.8%) were admitted to the hospital and 950 (38.2%) were discharged from the ED to home.

All variables with P values below 0.20 were entered into a multivariate logistic regression model. Five parameters shaped the final model for predicting death during hospital admission or within 1 week of discharge from the ED to home: age; previous history of LTHOT or NIMV; altered consciousness measured by GCS; use of accessory inspiratory muscles or paradoxical breathing upon ED arrival; and baseline dyspnea measured by the MRC scale (model AUC = 0.85, 95% CI 0.77 to 0.93). In separate models, we evaluated the effect on the final model of the role of arterial blood gases by adding the values of PO₂, PCO₂, and pH measured at both ED arrival and at the time that a decision was made to hospitalize the patient or discharge them from the ED to home. None of these was statistically significant.

We assigned a score to each category of the variables selected in the model based on their parameter estimates. Adding these individual scores yielded a final risk score that we called the continuous death in eCOPD (DeCOPD) risk score (AUC = 0.85, 95% CI 0.77 to 0.93) (Table 3). A categorical DeCOPD score was created by dividing the continuous DeCOPD score into four risk categories (mild, moderate, severe, and very severe) based on the risk of experiencing the outcome (Figure 1). This yielded an AUC of 0.84 (95% CI 0.76 to 0.91). Hosmer-Lemeshow tests were all P > 0.62. The continuous DeCOPD score (AUC = 0.84; 95% CI 0.77 to 0.92) and the categorical DeCOPD risk classes (AUC = 0.85; 95% CI 0.78 to 0.92) were validated in a separate sample (validation sample). Introducing the treating hospital variable into any model did not change the model, nor was the individual hospital significantly related to mortality (P = 0.064 unadjusted; P = 0.96 adjusted by the DeCOPD continuous score).

Outcome prediction model/parameter	Estimate	OR	95% CI of the AUC	Pvalue	Score
Short-term mortalitya
Age, years
75 to 85 versus <75	0.2417	1.273	0.540 to 3.003	0.5808	0
>85 versus <75	1.3062	3.692	1.134 to 12.022	0.0301	3
Previous use of LTHOT or NIMV at home	1.1557	3.176	1.278 to 7.893	0.0128	3
Altered consciousness	1.1852	3.271	1.031 to 10.376	0.0442	3
Use of inspiratory accessory muscle or paradoxical breathing upon ED arrival	1.4740	4.367	1.912 to 9.972	0.0005	4
Baseline dyspnea (MRC scale)
Grade 5 versus grades 1 to 4	2.0456	7.734	2.923 to 20.463	<0.0001	5
Missing versus grades 1 to 4	2.5581	12.912	4.654 to 35.821	<0.0001	5
bDerivation sample: AUCmodel = 0.85 (0.77 to 0.93); AUCscore = 0.85 (0.77 to 0.93) AUCscore _ cat = 0.84 (0.76 to 0.91)
bValidation sample: AUCmodel = 0.88 (0.82 to 0.94); AUCscore = 0.84 (0.77 to 0.92) AUCscore _ cat = 0.85(0.78 to 0.92)
1 month mortality (validation sample)
Age, years
75 to 85 versus <75	0.8199	2.270	1.092 to 4.719	0.0281
>85 versus <75	1.9253	6.857	2.701 to 17.411	<0.0001
Previous use of LTHOT or NIMV at home	1.3737	3.950	1.956 to 7.976	0.0001
Altered consciousness	0.9513	2.589	0.887 to 7.557	0.0817
Use of inspiratory accessory muscle or paradoxical breathing upon ED arrival	1.2508	3.493	1.850 to 6.596	0.0001
Baseline dyspnea (MRC scale)
Grade 5 versus grades 1 to 4	1.1743	3.236	1.581 to 6.625	0.0013
Missing versus grades 1 to 4	1.3324	3.790	1.643 to 8.742	0.0018
bValidation sample: AUCmodel = 0.85 (0.81 to 0.90); AUCscore = 0.83 (0.78 to 0.88); AUCscore _ cat = 0.83 (0.78 to 0.88)

Abbreviations: AUC area under the receiver operating characteristic curve, LTHOT long-term home oxygen therapy, MRC Medical Research Council, NIMV non-invasive mechanical ventilation.

^aShort-term mortality: mortality during hospital admission or, if discharged from the ED, in 1 week.

^bAUCs are the results for the whole model, the continuous score, and the categorical score, respectively.

The 1-month mortality was tested in the same model. It had an AUC of 0.85 (95% CI 0.81 to 0.90); AUCs were 0.83 (95% CI 0.78 to 0.88) for the continuous DeCOPD score and 0.83 (95% CI 0.78 to 0.88) for the categorical DeCOPD score.

In addition to short-term mortality, other outcomes that reflect adverse eCOPD evolution (for example, need for IMV, admission to ICU or IRCU) were compared between the four risk classes. Differences were observed across most of the categories for such outcomes (Figure 1). Specifically, differences between the four risk classes were observed for short-term mortality, in-hospital mortality, and 1 month mortality. When the four classes were dichotomized in two (score ≥12) the accuracy results for short-term mortality were: sensitivity 41%, specificity 97%, positive predictive value (PPV) 25%, and negative predictive value (NPV) 99%. When the cut-off point was 7 or above, the accuracy results for ICU admission, IMV use, or IRCU admission were: sensitivity 53.9%, specificity 79.05%, PPV 34%, and NPV 89.54%. The relationship of the DeCOPD categorical score in two categories (score ≥7) with other outcomes (for example, hospital admission, length of stay, readmissions or new ED visits) is presented (see Additional file 1).

This study describes the development of a clinical prediction rule for short-term mortality in patients with exacerbations of COPD, derived from a large, multicenter prospective cohort. This rule predicts death using clinical data generally available in the ED, and also often available in the primary care setting. Development of this rule followed proper procedures for derivation and validation [11, 12], and the rule provides excellent predictive validity.

Given how common exacerbations are among patients with COPD, and the substantial effects of these exacerbations on health, health-related quality of life, and healthcare costs, predicting the severity of exacerbations could significantly improve both care and allow more targeted allocation of healthcare resources. To date, however, no validated clinical prediction rules are available to stratify all patients experiencing an eCOPD upon their arrival in an ED, or to provide valid and reliable clinical rules or scores for predicting short-term outcomes [6, 9].

To the best of our knowledge, no prior prospective cohort studies including both patients admitted to the hospital and those discharged to home from the ED have been employed to develop and validate a clinical prediction rule for death using variables commonly available in the ED [6, 9]. Although previous authors have proposed predictive models for death or ICU admission, and some have created severity scores for patients with eCOPD, these did not evaluate such a complete range of variables as in our study, including arterial blood gases and other relevant data from the ED, and this limits their results [7–9, 25, 26].

The five factors in our final model—age, previous history of LTHOT or NIMV, use of inspiratory accessory muscles or paradoxical breathing upon ED arrival, altered mental status, and baseline dyspnea—have been observed separately as predictors of poor outcomes in previous studies [6, 8, 9, 11, 26–28]. Age is always a surrogate of other still unknown variables. Previous use of LTHOT or NIMV and the baseline level of dyspnea reflect the basal severity of the patient’s COPD. Altered mental status or the use of inspiratory accessory muscles upon arrival in the ED indicate acute cardiopulmonary compromise of the current presentation. Some of the parameters included in our rule are not modifiable, such as age or previous need of LTHOT or NIMV at home, which reflect the fragility of the patient and their disease. Severe baseline dyspnea can be modified through respiratory rehabilitation programs. Even for patients with severe COPD who cannot undertake respiratory rehabilitation, efforts to avoid new exacerbations or readmissions can be conducted through programs such as telemedicine or continuity of care, as has been reported elsewhere [29]. Finally, use of inspiratory accessory muscles or paradoxical breathing and altered level of consciousness reflect the severity of the current exacerbation. These are not permanent but modifiable and, in conjunction with other factors, should alert the ED physician to the severity of the patient’s eCOPD and guide appropriate treatment and adequate follow-up.

We found that measurement of arterial blood gases upon ED arrival or at the time a decision is made to hospitalize the patient or discharge them to home from the ED did not add any value to our score. We therefore did not include this measurement in our final model due to the high rate of missing data for this parameter, mainly at ED decision time. Nevertheless, future studies should evaluate the final role of these parameters in our score, although room for improvement in its predictive ability appears to be low.

From a practical point of view, our DeCOPD risk classification allows a clinician to identify patients who could be discharged to home (those with mild eCOPD) from those who should be admitted (severe and very severe eCOPD) because the latter are more likely to have an adverse event or to need intensive treatment and closer follow-up. For patients in the moderate category, further clinical information would be needed in the ED to decide if they should be admitted and if they need intensive treatment.

We compared our score with other rules or parameters used to classify patients with COPD, such as the ADO index, HADO score, GOLD FEV1%, and GOLD COPD combined assessment [16, 21, 22, 24]. In each case, these other scores performed poorly compared with our score. However, we must point out that those rules were created for patients with stable COPD, not those experiencing exacerbations. Tabak et al. recently proposed a prediction model derived from a large administrative database of hospitalized patients; their model has AUCs around 0.83, also but requires vital signs upon admission and laboratory results [10].

Our study has several strengths. The prediction rule we propose was created from data prospectively collected from a large (n = 2,487) group of patients experiencing COPD exacerbations recruited from 16 different EDs, each with its own set of guidelines or clinical practices for evaluating and treating patients with eCOPD. This is likely to improve the generalizability of the results. The cohort included both patients admitted to the hospital and patients who were discharged to home. Furthermore, we collected a broad range of clinical variables, many of them from the ED, and tested them in different models. We also strove to properly validate these prediction rules, following best practices for such studies [13, 14].

The lack of data for some key variables is the main limitation of our work. In a study such as this, clinical practice in the ED prevails over research requirements. We did not ask ED physicians to modify their clinical practices in any way. Thus, we had to work with the information available. Occasionally this led to missing data, but this did not appear to affect the results. In the case of baseline dyspnea (measured by the MRC scale), for example, we determined that patients with missing data had similar outcomes to those with class 5 baseline dyspnea, and the AUC of the final model was not affected by leaving these patient in the calculations or removing them.

It must also be noted that the patient population was almost entirely comprised of men (97%). Similar gender distributions have been observed in other studies performed in our country [30], which probably reflects the smoking patterns in Spain in the mid-20th century. Although we do not consider this to be a serious limitation, it could affect the generalizability of the results.

Finally, we did not include any biomarkers. At the time this study was conducted, several biomarkers were being evaluated for eCOPD (C-reactive protein, procalcitonin, copeptin, pro-adrenomedullin, pro-endothelin, and B-type natriuretic peptide). However, these were not routinely employed in the ED or in the hospital [31]. Whether these biomarkers add value to scores based on clinical variables will be the aim of future studies.

We developed a clinical prediction rule for a critical outcome, death, for patients attending an ED with eCOPD. The five variables included in our model are easily available in the ED (and also in primary care settings) and the score is easy to estimate. This clinical prediction rule could be employed in the clinical management of patients with eCOPD to guide their treatment and follow-up by in the ED and also by primary care physicians. Future studies are needed for validation of this prediction rule and to further demonstrate its value in clinical practice.