Study design and setting
We conducted a multicentre case-control study using data from 2010 to 2016, obtained from the electronic health records of seven hospitals and their primary healthcare institutions geographically located in two regions of Spain (Madrid and Catalonia). The study methodology has been described elsewhere [9]. The administrative, clinical and laboratory records from the data warehouses (DWs) of Catalan hospitals were linked with the primary healthcare data compiled in SIDIAP (Information System for Research in Primary Care) which contains pseudonymised clinical information from all the primary healthcare centres of the Institut Català de la Salut (ICS) [10]. Bellvitge University Hospital is the only hospital in which a DW integrates both clinical practice and primary healthcare data. In Madrid, the clinical data obtained from different sources and information systems integrated within the hospitals were linked to the Primary Healthcare Electronic Health Record (AP-Madrid) which contains data from all the primary healthcare centres within the Servicio Madrileño de Salud (SERMAS).
Study population
We identified all patients admitted to the hospital with a diagnosis of LA (pH <7.35 and plasmatic lactic acid concentration >5 mM/L within the first 24 and 72 hours after admission, respectively) from 2010 to 2016. The date of admission was used as the index date. The inclusion criteria were as follows: (1) at least 18 years of age, (2) hospital or primary healthcare diagnosis of DM2 before the index date, (3) chronic kidney disease (CKD) stage 3a (mild-moderate), 3b (moderate-severe) or 4 (severe) of the Kidney Disease Improving Global Outcomes (KDIGO) classification [11] during the 2-year period before the index date (excluding the previous 2 weeks), taking into account data from the primary healthcare database, and (4) availability of any information recorded on the primary healthcare database within a 1-year period before the index date. We excluded cases with: (1) diabetic ketoacidosis during the current in-hospital stay; (2) hospital or primary healthcare diagnosis of type 1 diabetes mellitus, human immunodeficiency virus disease or solid organ transplant before the index date; (3) hospital or primary healthcare diagnosis of malignant neoplasm (except skin cancer other than melanoma; including pheochromocytoma) within a 5-year period before the index date [9]. In Catalonia, patients not registered in the hospital referral area were also excluded.
Furthermore, people assigned to the hospital's primary healthcare region were chosen for the control group and matched in a ratio of 10:1 by age, gender, CKD stage and year of admission. They were at least 18 years old, had DM2 diagnosed before the index date and a CKD stage as defined for the cases. Additionally, they had information recorded on the primary healthcare database within a 1-year period before the index date. We excluded controls with: (1) type 1 diabetes mellitus, human immunodeficiency virus disease or a solid organ transplant before the index date; (2) malignant neoplasm (except skin cancer other than melanoma; including pheochromocytoma) within a 5-year period before the index date, and (3) patients not resident in the area of study.
Measurements
The hospital databases provided information on the patients' characteristics, including age and gender, hospital course data (admission date, in-hospital death, admission to critical care unit), laboratory test data (values and dates for lactic acid and haemoglobin concentration, and pH). The primary healthcare databases provided laboratory test data (values and dates for serum creatinine and haemoglobin concentration), and information on the drugs prescribed (anatomical therapeutic chemical [ATC] codes; prescription dates for metformin, other non-insulin antidiabetic drugs [NIADs], insulin, diuretics [high ceiling: furosemide, torasemide; low ceiling: hydrochlorothiazide, chlortalidone, xipamide, indapamide; potassium-sparing diuretics: spironolactone, eplerenone], renin-angiotensin system [RAS] inhibitors, and non-steroidal anti-inflammatory drugs [NSAIDs]; National Drug Code [NDC]; and prescribed posology for metformin). In the case of drug combinations, each drug was classified in its corresponding ATC group with a record of the drug's dosage in the combination. Additionally, diagnosis dates and codes (International Classification of Diseases, 9th and 10th revisions [ICD-9, ICD-10], and International Classification of Primary Care, 2nd revision [ICPC-2]) were obtained from both the hospital and primary healthcare databases. Renal function was estimated using serum creatinine data between 2 years and 2 weeks before the index date. The eGFR was calculated using the CKD-Epidemiology equation and stage according to the KDIGO classification (≥90, 60-89 45-59, 30-44, <30 mL/min/1.73m2) [12]. In the case of eGFR estimates resulting in different CKD stages for an individual patient across the 2-year period, the worst CKD stage closest to the index date was assigned to this patient when this was not followed by a better stage. Detailed information regarding the variables has been described previously [9].
Exposure definition
Exposure to metformin, other NIADs and insulin was defined as current use (prescription within a 30 day-period before the index date) or global use (prescription within a 365 day-period before the index date). We would like to point out that global use was defined in the statistical analysis plan.
The length of the exposure was defined as the time between the start and end prescription dates. A gap in drug prescription of ≤30 consecutive days was not considered as discontinued exposure. The prescribed daily dose of metformin was calculated according to the posology recorded by the prescriber and the strength corresponding to the NDC, and was categorised into <1g, 1-2g, and >2g. Exposure to diuretics, RAS inhibitors and NSAIDs were defined as prescriptions during the 30 day-period before the index date.
Statistical analysis
Baseline characteristics were summarised for cases and controls using standard descriptive statistics and a descriptive comparative analysis was carried out.
Conditional logistic regression was used to control for matches on age, gender, renal stage and year of index date. Crude and adjusted odds ratios (OR) with 95% confidence intervals (CI) were estimated to assess the risk of LA associated with metformin. The following drugs competed with metformin: alpha glucosidase inhibitors, dipeptidyl peptidase 4 (DPP-4) inhibitors, insulins and analogues, other blood glucose lowering drugs, sulfonylureas, thiazolidinediones, glucagon-like peptide 1 (GLP1) receptor agonists, NSAIDs, potassium-sparing diuretics, high ceiling diuretics, low ceiling diuretics and RAS inhibitors.
A stepwise procedure was used to select the baseline covariates to be included in the adjusted models. The variables selected were: alcohol use, liver disease, acute myocardial infarction, arterial peripheric arteriopathy, heart failure, chronic respiratory disease, dementia, seizures, gastroenteritis and dehydration.
Subgroup analyses were performed according to the daily dose of metformin, current and global use, and also the CKD stage.
Additionally, the overall case fatality rate of LA as well as the case fatality rate stratified by CKD stage were calculated from the number of deaths among cases and the total number of cases.
The possibility of detection bias was studied by analysing the frequency of determination of plasmatic lactate levels in patients with metabolic acidosis according to the status of metformin exposure. This analysis was performed with data from two of the participating hospitals in a sample of episodes of urgent hospital admission with pH <7.35 during the first 24 hours. An OR with its 95% CI for each hospital was provided.
All statistical analyses were performed with R statistical package version 3.6.0.