Prognostic Models for Disease Progression and Outcomes in Chronic Obstructive Pulmonary Disease: A Systematic Review and Meta-Analysis

<b>Background/Objectives</b>: Prevalence and burden of chronic obstructive pulmonary disease (COPD) are projected to increase in the coming decades. Although prognostic models for disease progression and exacerbation risk have proliferated, especially with the advent of machine learning (ML), their methodological rigor, generalizability, and predictive performance remain inconsistent. This study aimed to systematically review prognostic models for disease progression in adults with COPD, including traditional regression-based methods and ML techniques, evaluating model performance, sources of heterogeneity and methodological issues. <b>Methods</b>: PubMed and Embase were searched for all studies that developed and/or validated prognostic models for mortality (overall and cause-specific), exacerbations, or hospitalizations in adults with COPD over a time window of 1-5 years. Methodological quality was appraised using PROBAST. Model performance was summarized descriptively, and discrimination (c-statistic) was meta-analyzed for externally validated models with sufficient homogeneity. <b>Results</b>: Eighty-seven studies presenting 193 prognostic models across 96 unique cohorts were included. Only 7% of models were based on ML. Thirty-eight percent of records were validations of multidimensional indices. All-cause mortality (n = 85), severe exacerbations (n = 38) and moderate/severe exacerbations (n = 16) were the most frequently studied outcomes. Meta-analysis of exacerbation models was hampered by insufficient homogeneity (median c 0.74). As for mortality, BODE outperformed other indices (pooled c 0.75). Over 40% of studies were flawed by a high risk of bias. <b>Conclusions</b>: Despite a comprehensive literature search and thorough data extraction, we were able to provide a meaningful quantitative synthesis only for externally validated mortality models, as pooling results for other individual outcomes was precluded by substantial heterogeneity. Our findings highlight the predominance of regression approaches, the limited use of ML, the presence of persistent methodological limitations and the need for more robust, validated models capable of handling complex, multimodal patient data.