Can AI help in detecting respiratory diseases with incomplete lung function data?
Nilakash Das1,2, Armin Halilovic2, Julie Maes2, Kevin Ray2 and Marko Topalovic2
1 Laboratory of Respiratory Diseases and Thoracic Surgery, Department of Chronic Diseases Metabolism and Ageing, KU Leuven, Leuven, Belgium
2 ArtiQ NV, Leuven, Belgium
Accurate differential diagnosis of respiratory diseases requires interpreting the complete set of pulmonary function test (PFT) measurements including spirometry, body-plethysmography, and diffusion capacity. However, plethysmography may not be available or routinely performed in many clinical settings. Here, we explore the diagnostic potential of AI in interpreting PFTs with missing plethysmography.
We compared the diagnostic performance of ArtiQ.PFT, an AI-based diagnostic software leveraging complete PFT data (Topalovic et al. 2019 ERJ), to a machine learning model trained using spirometry and diffusion capacity (model B). The latter model was developed on the same training dataset (N=1400) as ArtiQ.PFT. A previously reported sample of 50 subjects with complete PFT and with a gold standard diagnosis was used for the comparison.
Overall diagnostic accuracy did not differ significantly (p=0.25, N=50) between ArtiQ.PFT (82%) and model B (74%). Sensitivity to COPD (100%, N=11) and asthma (75%, N=8) remained unchanged, as well as to interstitial lung disease (90%, N=10) and neuromuscular diseases (67%, N=3) that require lung volumes for interpretation. Sensitivity to thoracic deformity reduced (60% to 20%, N=5). Finally, we observed that spirometry and diffusion capacity measurements explained a significant (p<0.01) variation in plethysmographic parameters: RV (R2=0.45), TLC (R2=0.82), FRC (R2=0.72) and Raw (R2=0.49).
AI can help detecting respiratory diseases even when PFTs are incomplete. Performance is strong, as plethysmography data are highly correlated with spirometry and diffusion capacity. These results should be replicated with a large sample study.