An Automated Framework for ML-Assisted Lung-Function Assessment From Raw EIT Recordings

Spirometry indices such as FEV1 and the Tiffeneau index (TI) are the clinical gold standard for measuring airflow limitation but provide only global lung measurements without capturing regional ventilation variability. This study aims to develop an automated, machine learning (ML)-assisted pipeline that estimates spirometry-related lung-function indices and discriminates respiratory disease classes directly from electrical impedance tomography (EIT) signals, a radiation-free bedside modality with high temporal resolution.

We processed raw EIT recordings from a Dräger Pulmovista device through an automated pipeline that reconstructs impedance curves, performs quality control, optionally applies adaptive denoising, and segments the forced vital capacity (FVC) maneuver into four phases (A-D) to enable robust feature extraction. We evaluated multiple regression and classification models with hyperparameter optimization across three tasks: FEV1 prediction, TI prediction, and multiclass respiratory disease classification.

AdaBoost achieved the best performance for FEV1 regression (mean squared error [MSE] = 0.1026, $R^{2}$ = 0.9096), while XGBoost performed best for TI regression (MSE = 0.0037, $R^{2}$ = 0.7363). A multilayer perceptron (MLP) achieved the highest disease classification performance (accuracy = 0.8947, balanced accuracy = 0.8857, F1-score = 0.8857, ROC-AUC = 0.9444).

These findings demonstrate that automated maneuver segmentation combined with rigorous quality control and ML-based feature engineering enables reliable estimation of spirometry-derived indices and accurate discrimination of respiratory disease classes from EIT-derived descriptors alone.

This pipeline establishes EIT as a viable radiation-free, bedside alternative for continuous, spirometry-informed pulmonary function monitoring, with potential to improve respiratory disease surveillance in settings where conventional spirometry is impractical.