Machine learning validation of the AVAS classification compared to ultrasound mapping in a multicentre study.

Area of Science:

• Vascular Surgery
• Medical Imaging
• Machine Learning

Background:

• The Arteriovenous Access Stage (AVAS) classification aids in assessing vessel suitability for vascular access (VA).
• Previous clinical validation supports the utility of the AVAS classification.
• Machine learning models were employed to evaluate AVAS performance against ultrasound mapping data.

Purpose of the Study:

• To compare the performance of the AVAS classification with detailed ultrasound mapping measurements in predicting vascular access.
• To evaluate the predictive accuracy of AVAS and ultrasound mapping for both predicted VA (pVA) and created VA (cVA).

Main Methods:

• A prospective, multicentre international study (NCT04796558) involving 1151 patients.
• Data collected included demographics, risk factors, vessel parameters, and types of predicted and created VA.
• Random Forest algorithm was used to model pVA and cVA; model performance was compared using Bayesian generalized linear models with ROC AUC as the metric.

Main Results:

• Ultrasound mapping demonstrated higher predictive performance (ROC AUC for pVA: 0.85, cVA: 0.8) compared to AVAS (ROC AUC for pVA: 0.79, cVA: 0.71).
• Incorporating additional parameters with AVAS improved prediction (ROC AUC for pVA: 0.87, cVA: 0.82).
• Combining mapping data with other parameters yielded the highest predictive accuracy (ROC AUC for pVA: 0.88, cVA: 0.85).

Conclusions:

• Multiple ultrasound mapping measurements offer superior accuracy for vascular access prediction compared to the AVAS classification alone.
• The AVAS classification, while less accurate, offers a simpler and faster alternative, potentially making it more suitable for routine clinical practice.