Machine learning and deep learning for predicting acute exacerbation in chronic obstructive pulmonary disease: a systematic review

Introduction Acute exacerbation of chronic obstructive pulmonary disease (AECOPD) imposes a substantial burden on patients and healthcare systems. Although traditional risk assessment tools exist, machine learning (ML) and deep learning (DL) may offer improved prediction. This systematic review aims to evaluate the performance of ML and DL algorithms for predicting AECOPD risk. Methods This review was conducted in accordance with the PRISMA guidelines. PubMed, Embase, Cochrane Library and Web of Science were systematically searched from 1 January 2021 to 6 May 2026. Studies that developed or validated ML and DL models for predicting AECOPD risk were included. Methodological quality, risk of bias and applicability were assessed using the PROBAST + AI, and reporting quality was assessed using the TRIPOD + AI. Because of substantial heterogeneity, we synthesised model performance descriptively. Results Thirteen studies were included, of which only two performed external validation. Overall, reported AUC values ranged widely from 0.661 to 0.9721. By outcome type, AUC values for AECOPD risk prediction ranged from 0.668 to 0.9721, whereas those for AECOPD‑related hospitalisation ranged from 0.661 to 0.885. Similarly, by prediction windows, short‑term prediction achieved AUCs of 0.668‑0.972, compared with 0.661‑0.884 for long‑term prediction. However, PROBAST + AI revealed high methodological quality concerns and risk of bias in all 13 studies, and TRIPOD + AI showed suboptimal reporting transparency with an overall completeness rate of 54.3%. Conclusion Limited evidence indicates that models for any AECOPD may outperform those for hospitalisation, and that models with shorter prediction windows may show higher AUCs, although this association could be confounded by small sample sizes and potential patient-level leakage. These findings provide valuable direction for future model development and validation, although the current evidence base is constrained by methodological quality concerns, risk of bias and suboptimal reporting transparency. Systematic review registration PROSPERO CRD420261391432.