We designed the investigation as an observational, retrospective, multi-center, population-based study and extracted the hospital admission data from the health care records of the patients, none of whom were directly contacted.

We selected subjects with type 2 diabetes who were older than 18 years and hospitalized for COVID-19 (with the diagnosis based on ICD.9/10 codes) between January 2020 and December 2021 in 14 hospitals throughout Italy. We analyzed general data, pre-admission treatment schedules, date of admission or transfer to the intensive care unit (ICU) (i.e., the index date; taken as a marker of increased COVID-19 disease severity), and death (if it occurred) of only those who had been recorded as taking GLP-1ras, SGLT-2is, or DPP-4is as monotherapy at least twice (within a time interval of at least 6 months) before hospitalization.

We excluded those on other hypoglycemic drugs to reduce biases from expected confounding factors as follows: (i) insulin treatment could reflect greater clinical severity in T2DM; (ii) secretagogues carry potential cardiotoxicity or nephrotoxicity; and (iii) metformin might increase the risk for ketoacidosis in the case of respiratory disease, despite its beneficial effects on inflammation. Moreover, secretagogues and metformin have been limited to a secondary role in the latest American Diabetes Association (ADA) Standards of Care [28].

In particular, we need to remind the reader that, in the first months of the pandemic, COVID-19 cases dramatically increased within a few days, thus causing a severe shortage of hospital beds. Therefore, sub-intensive and intensive care wards were also used to cope with the high demand until beds became available again in ordinary wards. In general, patients with severe COVID-19 could not be transferred to the ordinary wards again, and they stayed in sub-intensive or intensive care beds for over a certain number of days, which was suggestive of severe disease. Based on the average duration of hospital stays for non-severely ill patients before transfer to ordinary beds, we arbitrarily chose an over-3-day stay as a marker of disease severity.

We identified treatments using the Anatomical Therapeutic Chemical (ATC) drug classification: DDP-4is as ATC A10BH, GLP1-ras as ATC A10BJ, and SGLT-2is as ATC A10BK.

We classified comorbidities according to the ICD.9/10 codes. They included myocardial infarction, cardiac arrhythmias, cardiac valvulopathies, hypertension, congestive heart failure (CHF), peripheral vascular disease, stroke, chronic obstructive pulmonary disease (COPD), pulmonary circulation disorders, rheumatoid disease, peptic ulcer disease, liver disease, paralysis and other neurological disorders, chronic kidney disease, cancer with/without metastasis, and hypothyroidism. The Charlson comorbidity index (CCI), a prognostic, predictive index of severity and life expectancy in patients with multiple comorbidities, helped us to individually identify the overall impact of comorbidities, treatment, and hospital discharge [29]. We calculated the previous hospitalization rate (× 1000) as the number of in-hospital stays during the 2 years before the index date.

This study complied with good clinical practice standards and followed the ethical guidelines of the 1964 Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the IRB (trial registration: Protocol n. 5, May 16, 2022), and the Ethical and Scientific Committee of the reference center, the Department of Endocrinology, San Raffaele Pisana Clinical Research Institute, Rome, Italy, served as the central reference ethical committee for the 14 affiliated hospitals contributing to the study. All subjects with T2DM who participated in the study signed an informed consent form before being included in the present investigation.

We analyzed the data using SAS 9.1 software and assessed the impact of drug classes on in-hospital mortality using the double propensity score logistic regression: (i) intensive care admission and (ii) no-intensive care admission.

We considered potential confounders, i.e., variables associated with drug choice (age, gender, previous hospitalization, Charlson index), and used them to calculate the propensity score.

We used a random match procedure to generate a 1:1 comparison without diabetes cohort replacement for each drug therapy group by applying the nearest neighbor method. After propensity score matching, we checked the balance achieved across selected variables, if so, ever.

We then used propensity score matching between the three drug classes to assemble a sample in which each patient receiving an SGLT-2i was matched to one on a GLP-1ra, and each patient on a DPP-4i was matched to one on a GLP-1ra, adjusting for covariates, as shown in Table 1. We finally used GLP-1ras as references in the logistic regression.

We chose these matching goals to reflect the relative number of users in each group and matched patients without replacement based on the propensity score within a range of 0.025, i.e., ~ 0.2 times the standard deviation of the propensity score. We obtained the estimated propensity scores from the logistic regression. We took an iterative approach to selecting confounders; a potential confounder was included in the model if this was required to ensure that the variable was balanced across treatment groups, as measured by the standardized mean difference. Imbalances of up to 0.2 were accepted.

We reported patient characteristics as mean ± standard deviation (SD) for continuous variables or number/percentage for categorical variables. We calculated, according to the Poisson regression model, the incidence rate (IR) within the 95% confidence interval (95% CI) for several parameters expressed as the number or percent. We used, as appropriate, analysis of variance (rANOVA) supplemented by the two-tailed paired Student’s t-test with 95% confidence intervals for parametric variables and the Mann–Whitney U test for nonparametric variables. We implemented the \(\chi^2\) test with Yates’s correction or Fisher’s exact test to achieve categorical variable differentiation. Finally, we considered all p values < 0.05 to be statistically significant.

When writing this manuscript, we followed STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines for observational study reporting (see the supplementary material).