InvestigatioN of a Smart Probe for Lung lEsion Characterization Using Impedance Technology

The primary endpoint was defined as the proportion of patients in which at least one non-anomalous biophysical measurement was obtained by the BSS in the lesion. This endpoint represents the procedural success rate being defined as the BioSpy System obtaining at least one non-anomalous impedance measurement in the lesion during the procedure, divided by the total number of patients in whom BSS was used.

Electrical impedance measurements were continuously acquired during bronchoscopy using the BioSpy System. The physician annotated time intervals corresponding to sensor contact with suspected lesion tissue and with healthy tissue, using bronchoscopic visualization and imaging guidance.

Impedance data collected within these annotated windows were extracted for analysis. Each window was matched to the corresponding histopathological diagnosis from biopsy samples obtained at the same location.

Machine learning model predictions (lesion vs. healthy) based on the impedance data were compared with the biopsy results to assess diagnostic performance.

Models were evaluated with leave-one-patient-out training: one patient was held for testing while the model trained on the others. This cycle repeated until all patients were tested, and the mean performance across iterations was computed. Mean accurary, mean sensitivity and mean specificity of detecting lesion are reported.

The BioSpy System's ability to differentiate tumoral, inflamed, necrotic and fibrotic tissues was assessed using continuous electrical impedance measurements recorded during bronchoscopy.

Physicians annotated intervals where the sensor contacted suspected lesion or healthy tissue guided by bronchoscopic visualization and imaging. Impedance data from these annotated windows were extracted and paired with histopathological diagnoses from biopsies taken at the same location.

Due to limitations in available labels, the secondary endpoint was evaluated by training a machine-learning model to distinguish cancer from all other tissue types.

Model performance was assessed with a leave-one-patient-out procedure: one patient was used for testing while the others were used for training, repeating the process until all had been tested.

Mean accuracy, mean sensitivity and mean specificity for cancer detection are reported.