Clinical Evidence of Endoleak Risk Index (ERI)
To validate the ERI in a clinical setting, a multi-centric retrospective study — EnduSim— was conducted across several European centers. Consecutive patients who underwentEVAR with Medtronic Endurant II devices were analysed and subsequently divided into two groups:
- Patients with type Ia Endoleak (EL1A): All patients who developed a type Ia Endoleak after EVAR, regardless of when the complication occurred.
- Control patients: Patients without type Ia Endoleak and with at least three years of follow-up were.
For all patients, simulation of EVAR implantation was performed in the same way as the real procedure.
First Version of ERI
The performance of the first version of ERI was published by Derycke et al. (EurJ Vasc Endovasc Surg, 2024). At that time, machine learning was not yet integrated into the algorithm.
The study showed that conventional sizing parameters — including those recommended in the device’s Instructions for Use (IFU) — did not differ significantly between patients with and without type Ia Endoleak. In contrast, almost all digital twin–derived parameters showed significant differences between the two groups.
Current Version of ERI
The present version of the ERI algorithm integrates machine learning, which the current dataset includes 117 patients in the training set (83 controls and 34 with type Ia endoleak — 13 early and 21 late) and 56 patients in the validation set (36 controls and 20 with type Ia endoleak — 6 early and 14 late).
Preliminary validation results demonstrate strong performance:
- Sensitivity: 80% (16 out of 20 patients with EL1A correctly identified).
- Specificity: 83% (30 out of 36 control patients correctly identified).
These findings confirm that the ERI can reliably distinguish high-risk patients from low-risk cases, offering clinicians a decision-support tool for EVAR planning.
AI behind ERI
ERI algorithm is a machine learning model based on meaningful physical features such as oversizing, proximal neck shape, conicity, apposition, presence of thrombus, and derived parameters from the previous lists. These features have been selected because they characterise the proximal sealing zone and play a role in assessing the risk of type IA Endoleak. ERI combines the most relevant features derived from the simulation with digital twin into a single easy-to-understand risk index.
This machine learning algorithm was trained on retrospective patient data for which the outcomes were known regarding the presence or absence of type IA endoleak. The algorithm was shown examples of patients who developed a type IA endoleak and those who did not. For the simulation-based features described in the previous paragraph, significant differences were identified, as reported in Derycke et al. (EurJ Vasc Endovasc Surg, 2024).
The role of the machine learning training process is to determine the optimal combination of features that best distinguishes patients with a type IA endoleak from those without.
To explain further, we can distinguish between different types of artificial intelligence. In machine learning, the list of potential input parameters—also called features—is defined by a human expert. These features must be meaningful and relevant to the specific problem being addressed, In this case, predicting the risk of a type IA endoleak.
The process of identifying and selecting such features is known as “feature engineering’. Machine learning differs from deep learning, which operates as amore opaque or “black-box” approach: the algorithm automatically identifies features within large datasets.
Because of this, deep learning models typically require much larger amounts of data than traditional machine learning methods. This explains why ERI, being machine learning based, achieved strong performance even with a limited number of patients.