Machine learning-based integration of respiratory oscillometry and spirometry for predicting clinical outcomes in...

Assessing respiratory function in spinal muscular atrophy (SMA) is challenging due to the effort-dependent nature of traditional spirometry. While respiratory oscillometry offers a passive alternative, its clinical utility in predicting clinical status remains underexplored. This study evaluates the efficacy of integrating spirometric and respiratory oscillometric indices through machine learning (ML) to predict ambulatory status and the requirement for bilevel positive airway pressure (BiPAP) in SMA patients. We retrospectively analyzed forty-five patients with a genetically confirmed diagnosis of SMA. To address data imbalance, we utilized the SMOTETomek technique and validated Random Forest models through leave-one-out cross-validation. SHAP (SHapley Additive exPlanations) analysis was used to interpret the contribution of individual physiological markers. In univariate logistic regression, forced vital capacity (FVC), peak expiratory flow (PEF), and forced expiratory flow at 75% of vital capacity (FEF75) were significant predictors for walking support. FEF25, FEF50, FEF75, and FEF 25-75 were associated with the requirement for BiPAP use. None of the respiratory oscillometric parameters reached statistical significance in regression models. However, ML models achieved high predictive accuracy, with SMOTETomek improving BiPAP requirement recall from 0.00 to 1.00. SHAP analysis revealed that respiratory oscillometry gained significant predictive weight that traditional models overlooked. R20Hz emerged as a key predictor for ambulatory status, while R5-20 Hz was identified as a decisive feature for predicting the BiPAP use.

The integration of effort-independent respiratory oscillometry with traditional spirometry via ML has the potential to enhance risk stratification in SMA.

• SMA requires lung function monitoring, but traditional spirometry is challenging due to the need for forceful expiratory maneuvers. • Respiratory oscillometry provides an effort-independent assessment of respiratory mechanics through tidal breathing.

• Machine learning successfully integrates effort-independent respiratory oscillometry with spirometry to significantly enhance clinical predictions in SMA. • It identifies R20Hz and R5-20 Hz as critical physiological markers for predicting ambulatory status and the requirement for respiratory support, respectively.