*2.3. Measures*

#### 2.3.1. Endpoint Definition

We used a composite endpoint to define AVF failure. EuCliD® has a dedicated module for record AVF failure event. However, reporting in this module may be incomplete. In order reduce the impact of reporting bias, we used a set of proxy variables suggestive of AVF failure. Therefore, we considered as an AVF failure any switch to a different vascular access, the occurrence of procedures aimed at re-establishing AVF patency (e.g., angiography with percutaneous angioplasty, stent placement or surgical AVF revision) and hospitalization due to AVF complications. The exact operative definition of the endpoint variable is described in Supplementary Table S1.

#### 2.3.2. Input Variables

The following classes of variables were considered for model input:


We ascertained diabetes by the occurrence of suggestive ICD10 codes according to the Charlson Comorbidity Index (CCI) definition [31]. Additionally, we extracted age, biological sex, dialysis vintage and number of patient's dialysis access.

### 2.3.3. Features Generation

We computed several metrics (minimum, maximum, average, standard deviation, slope) for continuous variables (e.g., dynamic venous and arterial needle pressure). Each metric was computed considering different time periods (e.g., last 7, 30, 90 days before index date).
