An exploratory analysis of flow volume loop morphology using convolutional neural networks

Lung function testing produces thousands of data points from a simple respiratory manoeuvre. Human interpretation reduces complex datasets to a few key variables and is operator dependent. Using machine learning to examine the entire dataset offers consistent proficiency and may identify unanticipated associations and novel pathophysiological processes. The flow volume loop (FVL) graphically represents a complete breath cycle lending itself to machine learning. We hypothesised that machine learning of FVL morphology alone could predict abnormal physiological variables. We collected a convenience dataset of 1451 full lung function reports. PDFplumber extracted tabular data as floating point numbers and graphs as x, y datasets. Each FVL was redrawn without axes or annotation limiting interpretation to morphological features only. Physiological variables were labelled as; severe (<50% predicted), Moderate (50-69%), Mild (70-89%) and normal (>90%). Resnet18 was selected from 9 candidate models based on inference speed and accuracy. Individual CNNs were finetuned for each physiological variable using PyTorch. The FEV₁/FVC model demonstrated an accuracy of 82.54%. Precision 78.61%, recall 82.54%, and F1-score 80.08%. Refinement of remaining models offers the potential for sophisticated analysis using captured images, independent of hospital testing or reporter skill. The prediction of gender from FVL morphology was unexpected and is worthy of separate study. <table-wrap><object-id>erj;64/suppl_68/OA2888/TB1</object-id><object-id>T1</object-id><object-id>TB1</object-id><table><colgroup><col></col><col></col><col></col></colgroup><tbody><tr><td></td><td>Train Accuracy</td><td>Validation Accuracy</td></tr><tr><td>Gender</td><td>70%</td><td>70%</td></tr><tr><td>FVC</td><td>81%</td><td>81%</td></tr><tr><td>FEV₁</td><td>72%</td><td>75%</td></tr><tr><td>FEV1/FVC</td><td>82%</td><td>87%</td></tr><tr><td>FEV_25-75</td><td>66%</td><td>68%</td></tr><tr><td>MEF₂₅</td><td>68%</td><td>68%</td></tr><tr><td>TLC</td><td>76%</td><td>80%</td></tr><tr><td>RV</td><td>50%</td><td>45%</td></tr><tr><td>FRC</td><td>61%</td><td>62%</td></tr><tr><td>RV/TLC</td><td>53%</td><td>41%</td></tr><tr><td>DLCO</td><td>49%</td><td>49%</td></tr><tr><td>VA</td><td>62%</td><td>56%</td></tr><tr><td>DLCO/VA</td><td>59%</td><td>46%</td></tr><tr><td>MEP</td><td>55%</td><td>38%</td></tr><tr><td>MIP</td><td>51%</td><td>39%</td></tr></tbody></table></table-wrap>