Data Resource Profile: The French National Uniform Hospital Discharge Data Set Database (PMSI)

All major industrialized countries have set up a prospective payment system for piloting health care costs.1 This system involves the production and transmission of structured and standardized billing information, including medical information concerning patients, their health problems and their treatments, commonly referred to as the Uniform Hospital Discharge Data Set (UHDDS). This information is eventually computerized and archived, and has led to the construction of massive databases. The collection of UHDDS records was initiated as a part of the implementation of the French Prospective Payment System (PPS), called the Programme National de Médicalisation des Systèmes d’Information (PMSI). Following a period of testing, experimentation and progressive implementation during the 80s2–7 and the 90s,8–10 UHDDS production by all health care facilities became mandatory in 199611, representing 20–25 million records par year. In 2001, a deterministic anonymous linkage identifier (LinkId) was implemented,12 thus making it possible to link UHDDS records and subsequently study patient re-hospitalization pathways. Since 2004, tariffs have been assigned to diagnosis-related groups (DRGs). The collection of UHDDSs has taken an increasing part in the determination of the income of French hospitals as concerns acute inpatient activity, until finally in 2008 all such hospital reimbursements became entirely based upon the UHDDSs. Furthermore, the UHDDSs are subject to quality and accuracy controls led by the French Social Security administration, potentially resulting in fines for those hospitals not respecting certain coding criteria. This process triggered significant improvement in data quality.13 Following major changes in March 2009, the methodology and terminologies for the production of UHDDSs have now been stable for 7 years. Currently, the UHDDSs describe all inpatient hospital stays in a standardized data set. The latter identifies the hospital and summarizes information about the patient and the stay, notably by describing diagnoses and procedures. It also provides additional information about specific aspects of the stay (for instance, a stay in an intensive care unit). Finally, the data set also contains the DRG the stay has been assigned to. Including the year 2015, the PMSI database amounts to approximately 180 million stays and 40 million LinkIds. A UHDDS is produced by every hospital for every acute inpatient stay, for ambulatory care requiring a hospitalization (e.g. carpal tunnel procedure) and for session treatments (e.g. dialysis session). A part of the data is directly generated from the administrative information system (e.g. date of entry, date of discharge). Medical information is coded from different sources (e.g. laboratory reports, discharge reports) according defined terminologies and coding rules. The Medical Information Department (MID) is in charge of the collection, control, transmission and analysis of the UHDDSs. The UHDDSs are then anonymized, essentially by removing the patient’s date of birth and the dates of entry and discharge, and uploaded to a government agency named the Agence Technique de l’Information sur l’Hospitalisation (ATIH);14 they are then further processed in order to issue payments for hospitals according to their activity. The LinkId is created by hashing a combination of several identifying traits via a non-reversible algorithm. This process is technically as safe as possible, with a collision probability (the occurrence of an identical hash for different patients) of 10−48, and since the identifying information is also combined with several secret keys, dictionary attack is virtually impossible.15,16 LinkIds are created in a two-stage process. At the hospital level, when transmitting the data to the ATIH, a linkage program is used to hash source data supposedly unique to each patient (birth date, sex and social security number) with a secret key. At the ATIH level, the resulting character string is again combined with a secret key and re-hashed. Additionally, a chronological identifier is created at the hospital level: (i) for a given LinkId, a hidden, random reference date is generated; (ii) the difference in days between this reference date and the date of admission is computed; and (iii) the result, labelled as the ‘stay number’, is associated with the record pertaining to this LinkId. Each time the hashing leads to this particular LinkId, the same reference date will be used, resulting in successive stays having increasing stay numbers. The stay number may then be used to chronologically ordinate the records associated with a given LinkId. Each record contains information mainly about: the hospital: national entity identifier; the patient: gender, age, location; the patient’s condition and management: pathologies, procedures; the stay: DRG, length of stay, type of admission and discharge (home, referral), month and year of discharge and additional information about specific wards (e.g. intensive care unit) attended. Pathologies are described by a principal diagnosis, and optionally secondary diagnoses. They are coded using the French version of the ICD 10.17 In specific cases, i.e. essentially session treatments, a ‘related diagnosis’ item is used. Procedures are coded using the Classification Commune des Actes Médicaux (CCAM).18 Laboratory tests and drug prescriptions are not recorded in the UHDDS. Three types of data resource use have successively developed over the years. Table 1 provides example articles for each category. Example articles of PMSI database use Example articles of PMSI database use Most of these studies are focused on a clinical theme, using the PMSI database as a source for identifying cases or analysing the main characteristics of a given population. Over time, some of these studies have incorporated other data sources (e.g. actual medical files, epidemiological registries) to consolidate them. The first domain that addressed the reliability of the PMSI database for disease monitoring as early as 2002 was cancer.19 This topic will remain the most explored over the years, with more and more optimistic conclusions concerning the reliability of the system as time passes and data quality improves.20 In this case, the PMSI database is used with methodological adjustments. Example studies again mainly involve cancer.21,22 However, a noteworthy case of adjusted use of the data occurred in association with the ‘Mediator® scandal’; the latter concerned the increase of valvular heart disease caused by benfluorex. Following several case-control and case report articles,23–27 an increase in the risk of valvular heart disease for diabetic patients treated with benfluorex was assessed for the French population in a study based on the PMSI database.28 This study significantly contributed to the re-assessment of the benefit-risk ratio for this product by the French Medicine drug agency, and ultimately to the suspension of its market authorization in France, shortly followed by a withdrawal recommendation from the European Medicines Agency.29 The UHDDSs are used to determine financial resources, and are frequently and thoroughly verified by both its producer and the paying party, with possible financial and legal consequences. It is thus reasonable to assume that this database is of an acceptable quality, taking into account the natural limits to precision imposed by the methodology and terminologies used to code conditions and procedures. Because it is comprehensive, the PMSI database gives the opportunity to conduct studies on a national scale. The volume of data ensures statistical power and precision. Combined with the study of medical information conveyed by the UHDDS, record linkage gives access to dynamic factors for the study of pathologies and patient behaviour (e.g. re-hospitalization rates and delays, type of hospital attended). The PMSI database is not built according to the thorough criteria required by clinical research. The quality of its data meets the requirement of the billing process, which relies more on completeness of core information and basic consistency (e.g. age or sex consistency with certain procedures). These data should simply be considered as raw data from the point of view of scientific research. Immediate use of the data is hampered by technical and methodological issues that can be classified into four categories. Content data quality: the data may include unuseable records containing errors, surrogate data (e.g. generic location code when patient location is unknown) or missing data. Hospital identification: in an ever-changing environment, every year hospitals close, open, merge or change identifiers. Coding rules: each year the official coding guide is updated, sometimes with important impacts on the way patient disease and patient management are represented through coding. Terminology updates: new codes are regularly added to the French version of the ICD10, along with new coding rules. A public website presents predetermined queries on coarse parameters to the public42. For example, it is possible to obtain the case mix of a given hospital for a given year. However, it is not possible to combine patient parameters (age, sex, location) with disease. This resource is not useful for scientific research. The complete database is available online via a accredited provider,43 provided that regulatory requirements are met concerning data security and dissemination.44 The cost of access currently varies between a few hundred and a few thousand euros, depending on the complexity of the considered project. A more efficient solution is to establish collaboration with the health information management (HIM) department of a French university hospital. Since HIM departments access these data on a daily basis to analyse hospital activity, most of them have developed considerable expertise as concerns the data, the accessibility platform, regulatory requirements and associated limitations. This ensures methodology expertise and safety. The PMSI database was initiated during the establishment of the French Prospective Payment System (PPS). Nation-wide record production has been mandatory since 1996. The database contains a record for each acute inpatient stay and represents about 25 million records per year. The records describe the stays in a standardized data set. This data set summarizes information about the patient and the stay, describing diagnoses and procedures and also providing additional information about specific aspects of the stay (for instance, a stay in an intensive care unit). Finally, the data set contains the diagnosis-related group (DRG) for the stay. Since 2001, an anonymous patient identifier has enabled record linkage. The useage of these data has arisen and evolved over the years, ranging from simple auxiliary useage to studies based on a specific adjusted exploitation. The most elaborate studies are focused on the epidemiology of cancer. This database is available upon registration with and payment to a habilitated provider, or through collaboration with a French university hospital health information management department. Conflict of interest: None declared.