Can Artificial Intelligence Optimize the Early Diagnosis of Invasive Candidiasis? A Systematic Review and Meta-Analysis

TL;DR

This paper reviews AI models for early detection of invasive candidiasis, finding moderate accuracy and highlighting the need for further validation.

Contribution

A systematic review and meta-analysis of AI-based models for early diagnosis of invasive candidiasis in high-risk patients.

Findings

Pooled sensitivity and specificity for candidemia prediction were 81.3% and 81.6%, respectively.

Most models had high negative predictive values but modest positive predictive values due to low event prevalence.

Risk of bias was moderate to high, and evidence certainty was low due to study limitations.

Abstract

The early diagnosis of invasive candidiasis remains challenging in immunocompromised and other high-risk patients, prompting interest in artificial intelligence models for assisting clinical decision-making. We conducted a PROSPERO-registered systematic review and meta-analysis of artificial intelligence-based predictive models for the early identification of invasive Candida infections. We searched multiple databases for studies reporting model performance in hospitalized immuno-compromised patients. Data on study characteristics, model details, validation strategy, and diagnostic accuracy were extracted. A bivariate random-effects meta-analysis was performed for candidemia prediction models with compatible data. Eight studies met inclusion criteria. Models were typically developed using retrospective hospital data with heterogeneous populations and predictors. Five candidemia studies provided threshold-based performance data for meta-analysis. Pooled sensitivity and specificity for candidemia prediction were 81.3% (95% confidence interval (CI) 72.9–87.6%) and 81.6% (95% CI 68.4–90.1%), respectively. Most models achieved high negative predictive values, whereas positive predictive values were modest, reflecting low event prevalence. The risk of bias was generally moderate to high (PROBAST), and the certainty of evidence was low (GRADE) due to study limitations and indirectness. AI models show promise for early candidemia identification with moderate diagnostic accuracy. They may be useful as decision-support tools, but further multicenter prospective validation is needed before routine clinical adoption.

Figures