AI-Powered Respiratory HealthMonitoring with Predictive Analytics andPersonalized Insights

Sammanfattning

Chronic obstructive pulmonary disease (COPD) affects millions of people but most casesgo undiagnosed until the disease has already caused serious damage. The standard diagnostic approach uses a single ratio from two spirometry measurements.Recent machine learning models have shown strong potential for improving diagnostic accuracy in respiratorymedicine. The full shape of the expiratory flow-volume curve carries far more informationthan this ratio captures.This thesis investigates whether machine learning applied to standardand novel spirometry features can improve COPD classification and develops a separateearly-warning model trained exclusively on pre-diagnosis visits. Methods:This study analysed spirometry parameters, demographic variables, smoking history, comorbidity profiles, and longitudinal visit records from two population datasets: theUnited States National Health and Nutrition Examination Survey (NHANES) and the SwedishNational Airway Register (SNAR) outpatient registry, comprising over 1.7 million valid patient records collected between 2014 and 2024. The research addressed three classificationproblems: cross-national binary COPD classification, comorbidity-enriched COPD classification, and longitudinal early warning prediction across one, two, and three year horizons.Random Forest and XGBoost models were applied across all three phases. Feature engineering incorporated standard spirometry indices,the expiratory area ratio (AreaFE%) , thebeta angle of the descending expiratory limb, and five longitudinal trend features computedfrom multi-visit patient histories. Model performance was assessed using the area under thereceiver operating characteristic curve(AUC), balanced accuracy, sensitivity and specificity.SHAP analysis was applied to all final models to produce clinically interpretable featureimportance rankings. Results:The Random Forest model trained on NHANES transferred directly to the SNAR Swedishoutpatient population without retraining, achieving an AUC of 0.881, though local retrainingon SNAR raised performance to an AUC of 0.954 with sensitivity of 0.909 and specificity of0.912. The addition of six comorbidity variables improved AUC by only 0.001, confirmingthat spirometry and demographic features already capture the dominant classification signal.The age-excluded early warning model predicted COPD development at one, two, and threeyear horizons, achieving AUCs of 0.72, 0.72, and 0.73 respectively. Conclusions: The findings suggest that machine learning models applied to spirometry datamay enable earlier and more accurate identification of COPD than conventional ratio-baseddiagnostic approaches.This study shows that local population retraining produces the largestperformance gain across all experiments, and that routine clinic visits carry enough signalto flag future COPD patients up to three years before a formal diagnosis is recorded. These1results support the integration of machine learning into portable digital health platforms suchas SpiroLuft. However, prospective validation on larger and more diverse patient populationsacross multiple national healthcare systems remains a critical step before these models canbe responsibly deployed in routine clinical practice.