# AI/ML driven prediction of COPD exacerbations and readmissions: a systematic review and meta-analysis

## TL;DR

This study reviews AI/ML models for predicting COPD flare-ups and hospital readmissions, finding moderate accuracy but significant variability and methodological issues.

## Contribution

The first systematic review and meta-analysis of AI/ML models for COPD exacerbation and readmission prediction, evaluating performance and methodological quality.

## Key findings

- AI/ML models showed moderate accuracy (AUC 0.73–0.77) in predicting COPD exacerbations and readmissions.
- Externally validated models had higher accuracy (AUC 0.82) compared to internally validated models (AUC 0.76).
- High heterogeneity and frequent methodological flaws limit the generalizability of current models.

## Abstract

Chronic obstructive pulmonary disease (COPD) exacerbations and hospital readmissions are major drivers of morbidity, mortality, and healthcare costs. Artificial intelligence and machine learning (AI/ML) approaches have been applied to predict these events, but their pooled performance and methodological rigor remain unclear.

Following PRISMA 2020 guidelines, we conducted a systematic review and meta-analysis of peer-reviewed studies developing or validating AI/ML models for predicting acute exacerbations of COPD (AECOPD) or hospital readmissions. Databases (PubMed, IEEE Xplore, Cochrane Library, Semantic Scholar) were searched to 2025. Eligible designs included retrospective and prospective cohorts, randomized trials with embedded prediction, and case–control studies. Study quality was assessed using PROBAST, and evidence certainty with GRADE. Random-effects models pooled area under the ROC curve (AUC); subgroup analyses compared AECOPD vs. readmission outcomes and internal vs. external validation.

Thirteen studies were included, with sample sizes ranging from 110 to 113,786 patients. Most were retrospective cohorts using EHRs or claims data, while two used prospective or trial-based data. Models applied diverse algorithms, including random forests, gradient boosting, neural networks, and ensemble pipelines. The pooled AUC across all studies was 0.77 (95% CI: 0.74–0.80), with very high heterogeneity (I2 = 99.5%). Subgroup analyses showed similar performance for AECOPD prediction (AUC = 0.77; I2 = 98.9%) and readmission prediction (AUC = 0.73; I2 = 19.8%). Externally validated models (n = 4) achieved higher accuracy (AUC = 0.82) than internally validated models (AUC = 0.76), although differences were not statistically significant. Risk of bias was moderate to serious in 69% of studies, mainly due to incomplete reporting and overfitting.

AI/ML models demonstrate moderate-to-high discriminatory accuracy in predicting COPD exacerbations and readmissions, with pooled AUCs of 0.73–0.77. However, high heterogeneity, limited external validation, and frequent methodological concerns restrict generalizability. Standardized reporting frameworks (TRIPOD-AI, PROBAST-AI), rigorous external validations, and prospective implementation studies are needed to translate these promising tools into clinical practice.