Early clinical decision support using interpretable artificial intelligence in acute illness (sepsis and infection):...

Sepsis and severe infections remain major causes of morbidity and mortality worldwide, particularly in acute care settings where early recognition is critical. Artificial intelligence (AI)-based clinical decision support systems (CDSS) have emerged as promising tools for improving early diagnosis and prognostic assessment. However, limited interpretability and transparency remain key barriers to clinical adoption. This systematic review and meta-analysis aimed to evaluate the diagnostic and prognostic performance of interpretable AI models in acute illness related to sepsis and infection.

A systematic literature search was conducted in April 2026 across PubMed, Scopus, and the Cochrane Library. Eligible studies included original English-language research published from 2024 onward evaluating interpretable or explainable AI, machine learning (ML), or deep learning (DL) models integrated within CDSS for patients with sepsis or severe infection. Studies using non-interpretable black-box models or lacking patient-level outcomes were excluded. A bivariate random-effects meta-analysis was performed for studies reporting sufficient 2 × 2 contingency data.

A total of 849 records were identified, of which 15 studies met the inclusion criteria. These were categorized into sepsis detection/diagnosis (n = 8), mortality prediction (n = 5), and complication prediction (n = 2). For sepsis detection, pooled sensitivity was 0.685 and specificity was 0.898 (AUC = 0.879). Sensitivity analysis demonstrated stable performance (sensitivity 0.751, specificity 0.873, AUC = 0.890). For mortality prediction, pooled sensitivity was 0.618 and specificity was 0.778 (AUC = 0.753), indicating moderate predictive performance. Quantitative synthesis was not performed for complication prediction because only two clinically heterogeneous studies were available. Fourteen of the 15 included studies were judged to have high overall risk of bias, predominantly because of concerns in the analysis domain.

Interpretable AI-based CDSS models demonstrated good overall discrimination for sepsis detection but only moderate performance for mortality prediction. Explainability methods provided insight into model outputs; however, the included studies did not directly evaluate clinician trust, adoption, behavioral responses, or patient-level implementation outcomes. Confidence in the findings is limited by substantial heterogeneity and predominantly high risk of bias. Future research should prioritize prospective multicenter validation, calibration, temporal modeling, and clinical-utility assessment before routine deployment.