Deep Learning of Suboptimal Spirometry to Predict Respiratory Outcomes and Mortality

Area of Science:

• Pulmonary Medicine
• Machine Learning in Healthcare
• Respiratory Diagnostics

Background:

• Spirometry is crucial for diagnosing respiratory diseases, typically relying on maximal efforts that meet quality control (QC) criteria.
• The utility of suboptimal spirometry efforts (submaximal or QC-failing) for predicting respiratory outcomes remains unclear.

Purpose of the Study:

• To develop and evaluate a machine learning model capable of predicting respiratory outcomes and mortality using suboptimal spirometry data.
• To assess the value of incorporating all spirometry efforts, including suboptimal ones, into predictive models.

Main Methods:

• A novel Spirogram-based Contrastive Learning Framework (Spiro-CLF) was developed to create lung function representations from all available volume-time curves (maximal and suboptimal).
• The Spiro-CLF model was trained on UK Biobank data and validated on independent UK Biobank and COPDGene cohorts.
• Model performance was evaluated for predicting airflow limitation (FEV1/FVC <0.7), low FEV1 Percent Predicted (FEV1 PP <80%), all-cause mortality, and respiratory phenotypes.

Main Results:

• In the UK Biobank, Spiro-CLF effectively predicted FEV1/FVC <0.7 (AUROC 0.956) and mortality (concordance index 0.647) using suboptimal efforts, and improved asthma prediction by 9-42%.
• In the COPDGene cohort, while suboptimal efforts did not enhance lung function or mortality prediction, Spiro-CLF representations significantly predicted asthma and other respiratory phenotypes (P ≤ 2 × 10^-3).
• A significant proportion (61.6%) of spirometry efforts were suboptimal, highlighting the potential data loss from excluding these attempts.

Conclusions:

• A machine learning model (Spiro-CLF) can successfully predict respiratory phenotypes and mortality using suboptimal spirometry data.
• All spirometry efforts, including suboptimal ones, contain valuable information that can enhance predictive modeling for respiratory diseases.
• Further research is warranted to explore the clinical utility of this approach in specific patient populations and scenarios.