Integrated Pulmonary Severity Score (IPSS) for COPD: A Psycho-Respiratory Risk Index Supported by Explainable Machine Learning

Background/Objectives: In chronic obstructive pulmonary disease (COPD), forced expiratory volume in one second (FEV1) explains only part of the variability in symptoms and prognosis, while anxiety and depression are common but rarely quantified in composite indices. We aimed to develop and internally validate an Integrated Pulmonary Severity Score (IPSS) that combines respiratory function, symptom burden and affective status. Methods: In a prospective observational study, 390 adults with spirometry-confirmed COPD were consecutively enrolled at two tertiary Romanian centres (October 2022–September 2024). Within 48 h of admission, patients underwent spirometry (FEV1% predicted), dyspnoea grading (mMRC), symptom assessment (CAT), affective evaluation (HADS-Anxiety/Depression) and cognitive screening (MoCA, MMSE). A PulmoScore was built from CAT, mMRC and ventilatory deficit (100 − FEV1%) and extended with a HADS-based psychiatric multiplier to obtain IPSS. Spectral clustering, logistic regression, a multilayer perceptron (MLP) and LIME were used for phenotyping and validation. Results: Spectral clustering identified two phenotypes—psycho-respiratory and predominantly respiratory—with acceptable separation (silhouette coefficient 0.26). The psycho-respiratory group showed lower FEV1%, higher CAT and mMRC scores, more severe anxiety–depression and markedly higher IPSS values. Logistic regression and the MLP achieved an accuracy of 0.89, an AUC of 0.95 and Cohen’s κ ≥ 0.75 for identifying this phenotype when using the same core clinical variables that informed phenotyping and IPSS construction. IPSS values were distributed across four strata (<30, 30–69, 70–119, ≥120 points), reflecting progressively worse respiratory and affective burden. Conclusions: In this cohort, IPSS captured a clinically meaningful psycho-respiratory phenotype and improved integrated severity assessment beyond spirometry alone, with potential utility for risk stratification. It can be computed from routine measures, is compatible with explainable AI workflows and warrants external, longitudinal validation before widespread implementation.