Background: Respiratory oscillometry (Osc) is pulmonary function (PF) modality that measures respiratory mechanics; however, interpretation of Osc is challenging.

Hypothesis: Machine learning (ML) facilitates Osc interpretation.

Objective: To develop a novel ML architecture to classify patterns of PF and compare its accuracy to expert opinion and a physician verified label.

Methods: Data were taken from 245 subjects with 1924 valid 10 Hz mono-frequency Osc tests. Full PF tests and clinical data were used for the gold standard label. Osc measurements of flow, volume, and pressure were inputted for ML and randomly partitioned into training and validation (70:30) sets based on unique subject count. The MiniROCKET algorithm was applied to generate features from the input data which were resolved by a ridge regression classifier to different patterns of PF. A soft voting scheme was implemented on the output classifier scores to reach a final prediction for each Osc test. Results were averaged over 10 experimental runs. ML performance was compared to 11 experts in the interpretations of 72 randomly selected oscillograms. 

Results: Validation accuracy of ML was 86±4% overall, with highest accuracy in identifying normal (N) and restrictive (R) patterns (91±2%). ML was significantly better than expert interpretations of 72 Osc tests for recognizing all but obstructive (O) patterns, with highest accuracy in identifying N patterns (p<0.001).

Conclusions: ML can resolve mono-frequency Osc recordings to N, R, and mixed O-R patterns with significantly greater accuracy than experts in the field.