Data sources
Cause-of-death data
The French death certificates are complying with the WHO international standards. They are exhaustively collected by the Epidemiological Center for the Medical Causes of Death (Inserm -CépiDc) [12]. Since 2000, causes of death are coded according to the 10th revision of the International Classification of Diseases (ICD-10) [11]. This analysis includes all the causes mentioned on the death certificate, 3.4 on average, plus the UCD determined by the ICD-10 rules. The UCD can be one of the causes mentioned on the death certificate or a combination of these causes in a single code (e.g. Diabetes with renal complication).
Hospitalization data
The French acute care hospital database (PMSI-MCO) [13] is designed for hospital payment. It provides medical information for all patients discharged from short-stay hospitals, both public and private. Patient's stays are chained so that the number of hospitalizations within the year before death can be calculated. This study focuses on the last hospitalization before death (both occurring in 2008 or 2009). The patient's gender, age (at admission), and the main diagnosis (ICD10 code) were extracted. The hospitalization database is included in the Social Security database (SNIIRAM) [14].
Linkage
By the time of the analysis, vital status and date of death of the deceased were available only for the beneficiaries of the general health insurance regime. This population accounts for about 70% of French residents (it does not include state employees, students, self-employed, agricultural workers and farmers). Among those hospitalized during the year preceding their death, 96.4% of these beneficiaries could be linked to a single death certificate. The matching was performed through a deterministic methodology allowing at most one difference on one of the following indirect patient identifiers: year and month of birth; year (this variable had to match), month, and day of death; gender; département and commune of residence. Only unique matches were kept in the final set.
Infants deceased before one year of age were excluded because the quality of the vital status assessment for this age-class could not be precisely known. Besides, since the discharge-death time interval was imprecise for 2008, the 2008 records were considered only when death occurred in hospital, or 3 months or more after discharge (exact day of discharge available in 2009, month of discharge only in 2008).
The final database comprises 421,460 subjects deceased in the year following their last discharge.
The linkage of the hospitalization and cause of death data and the study of the resulting dataset were approved by the two French data protection committee and institutional ethical review boards concerned: Institut des Données de Santé (authorization n°16-24/11/2010) and Commission Nationale de l'Informatique et des Libertés (authorization n° 1454315).
Definitions
The French definition of the main diagnosis has changed during the study period, from "condition that takes up the majority of resource use during the hospital stay" before march 2009, to "final diagnosis explaining hospital admission" after. However, this modification of definition had no impact on our results (results not shown).
In order to capture the pathology, which is the relevant information in our purpose, when the main diagnosis of the hospital database was a chapter XXI code (Factors influencing health status and contact with health services), the "main diagnosis" (MD) mentioned in this article was defined as the related diagnosis.
UCD is defined in volume 2 of ICD-10th revision as "(a) the disease or injury which initiated the train of morbid events leading directly to death, or (b) the circumstances of the accident or violence which produced the fatal injury".
Comparison method
Classification
The aim was to compare MD and UCD in order to analyze their independence or consistency.
Consistency was defined as MD and UCD belonging to a same train of events leading to death. If the quality of the information held in both codes seemed sufficient and if MD and UCD could not belong to a same train of events leading to death, they were considered independent.
Four cases were distinguished:
-
Similarity: MD and UCD refer to the same morbid condition, even if precision levels may differ (Eg1: UCD = Intracerebral haemorrahage, unspecified (I61.9) and MD = Intracerebral haemorrahage in cortical hemisphere (I61.1). Eg2: UCD = Pneumonia, unspecified (J18.9) and MD = Bacterial pneumonia, unspecified (J15.9)).
-
Acceptable sequence: the two codes refer to different conditions but belong to a same train of events leading to death. Because UCD is defined as the cause that initiated the process, UCD can precede MD in the causal sequence, even though death occurs chronologically at the end of or after the last hospital stay. Acceptable sequences correspond to cases where MD is a complication of UCD (Eg1: UCD = Alcoholic cirrhosis of liver (K70.3) and MD = Rupture of esophageal varices (I85.0) . Eg2: UCD = Malignant neoplasm of bronchus and lung, unspecified (C34.9) and MD = Secondary malignant neoplasm of brain and cerebral meninges (C79.3)).
-
Independence: both conditions belong to different trains of events leading to death (Eg: UCD = Calculus of bile duct with cholangitis (K80.3) and MD = Primary coxarthrosis, bilateral (M16.0)).
-
Non-informative death certificate: cases that cannot be interpreted in terms of similarity, acceptable sequence or independence because UCD is not informative (Eg: UCD = Cardiac arrest, unspecified (I46.9)).
Similarities and acceptable sequences compose the consistent cases.
Algorithm
In order to classify each death in one of these four cases, an algorithm was designed to compare MD of last stay and UCD, taking all medical conditions mentioned on the death certificate into account (see Additional file 1).
At four stages of the algorithm, the type of relationship between MD and UCD was given by running Iris software (V. 4.0.38) on "test certificates" (see Additional file 2). Iris is a language-independent coding system using international standards [9, 10] (see Additional file 3): the WHO ICD-10 classification, rules and guidelines as well as the knowledge base of the Mortality Medical Data System (MMDS) [15, 16], ACME (Automatic Classification of Medical Entry) software in particular [17, 18].
Artificially introducing MD in a test certificate, at a specific place according to the question asked, permitted us to assess its potential participation in the causal sequence leading to death (5). This method is an update and an extension of the one first proposed by Johansson and Westerling [7].
Statistical analysis methods
In-hospital and out-of-hospital deaths were analyzed separately. The relationships between MD and UCD were studied according to age, gender, discharge-death time interval (in months) and main ICD Eurostat Shortlist chapters of UCD (corresponding ICD chapters): neoplasms (Chapter II), diseases of the nervous system and the sense organs (Chapters VI & VII), circulatory (Chapter IX), respiratory (Chapter X), and digestive (Chapter XI) systems, external causes of morbidity and mortality (Chapter XX), and one class for Others. Because they do not hold any information about the organs originally implied in the death process, imprecise UCDs (ICD-10 code in R99, R96.0, R57.9, R40.2, R09.2, I46.9, I99, I95.9, J96.0, J96.9, P28.5) were excluded from the comparison according to the category of UCD.
Univariate and multivariate log-binomial regression analysis [17] were used to study risk factors for independence vs. consistency, excluding non-informative cases. Relative risks (RRs) of independence and their 95% confidence intervals were estimated, crude and adjusted for age, gender, discharge-death time interval, length of stay, number of stays during the last year of life, and category of UCD. For each variable, the modal class was chosen as reference class.
Age, discharge-death time interval, length of stay, and number of stays during last year of life were included as continuous variables in order to perform trend tests.
Analyses were performed with SAS® version 9.3.