Prognostic Performance Of Artificial Intelligence Models In Predicting 12-Week Healing of Chronic Wounds: A Systematic Review And Meta-Analysis

Abstract

Introduction

Chronic wounds pose a substantial health burden, requiring intensive, long-term management and carrying a high risk of debilitating complications. Accurate prognosis regarding the probability and rate of healing within the critical 12-week timeframe is essential for optimizing specific care strategies and ensuring the effective allocation of scarce medical resources. Artificial Intelligence (AI), particularly through its foundation in Machine Learning (ML), offers significant potential to enhance prognostic accuracy by rigorously processing vast quantities of Electronic Medical Record (EMR) data and advanced wound imagery.

Methods

This systematic review and meta-analysis was conducted in strict adherence to the PRISMA 2020 reporting guidelines. Included studies specifically evaluated AI models designed to predict chronic wound healing outcomes within 12 weeks. The methodological quality of these studies was critically assessed using the specialized Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence (PROBAST+AI). Quantitative synthesis was executed to determine the pooled discrimination performance metric, the Area Under the Curve (AUC), and to measure the independent effects of key predictors using pooled Hazard Ratios (HR).

Results

The analysis incorporated two large-scale primary studies boasting high data volumes, alongside several supporting methodological studies. The resultant pooled AUC for AI models reached 0.805 (95% CI: 0.778–0.832), definitively confirming significant prognostic capability. Specifically, models utilizing Gradient-Boosted Decision Tree (GBDT) algorithms achieved an AUC of 0.853, a performance level that substantially outperformed conventional Logistic Regression models (AUC 0.712). Assessment utilizing PROBAST+AI consistently highlighted systemic methodological quality issues, predominantly stemming from weak internal validation within the Analysis Domain, which consequently elevated the Overall Risk of Bias. The pooled HR analysis, synthesizing data for 10 critical prognostic predictors, confirmed that local wound characteristics are the paramount determinants of prognosis. High-Grade Wound Depth (Stage 3/4) was identified as the single strongest inhibitor of healing (HR 0.65 (95% CI: 0.59–0.71)), whereas Normal/Good Vascularization Status represented the strongest accelerator (HR 1.30 (95% CI: 1.22–1.39)).

Discussion and Conclusion

The prognostic performance demonstrated by AI models is statistically significant and definitively exceeds that of conventional statistical methods. This heightened accuracy is attributed to the inherent capacity of non-linear models to effectively capture complex multi-variable interactions central to wound healing pathophysiology. Notwithstanding the encouraging performance metrics, the documented high risk of overfitting due to analytical bias necessitates strict and rigorous external validation prior to any extensive clinical implementation.