The Clinical Problem Radiotherapy is a critical treatment for lung cancer, but response varies significantly between patients. Stereotactic ablative RT achieves over 70% 5-year local control in early NSCLC, yet many patients fail to respond. Without a reliable pre-treatment predictor, non-responders receive full courses of RT with side effects but without benefit.
The Novel Approach This study collected multimodal data one week before RT initiation - including demographic, radiological, biological (tumor markers), and physiological characteristics (respiratory patterns, heart rate variability, surface electromyography) - and trained a Back Propagation Neural Network (BPNN) to predict both partial response (PR) and progressive disease (PD).
Physiological Signal Innovation The most distinctive feature of this study is the inclusion of pre-treatment physiological signals: airflow, transthoracic impedance, diaphragm surface electromyography, and ECG collected during controlled breathing protocols. These signals encode autonomic nervous system function and respiratory physiology that may reflect tumor biology and systemic radiation sensitivity.
Key Results The comprehensive BPNN model achieved AUC 0.855 for PR prediction and AUC 0.929 for PD prediction, substantially outperforming the best individual features (AUC 0.699 and 0.750 respectively). Prediction MSE was 0.07 for PR and 0.01 for PD, confirming robust model calibration.
Patient Population 120 lung cancer patients scheduled for radiotherapy at a single Chinese military medical university hospital between May 2022 and December 2024 were enrolled. Of these, 41 achieved partial response, 69 had stable disease, and 10 experienced progressive disease. Treatment was IMRT delivered at 6 MV per ICRU Report 83 guidelines.
Multimodal Data Collection All data were collected one week before RT initiation (median 7 days). Demographic and radiological data were extracted from electronic medical records. Biological markers (CEA, CA15-3, CA19-9, NSE, SCC, CYFRA21-1, CA125, ProGRP) were obtained from pre-treatment blood tests. Physiological signals were acquired in a dedicated monitoring session.
Physiological Signal Acquisition Participants were instrumented with a thermal gas mass flowmeter for airflow and five Ag-AgCl electrodes for transthoracic impedance, diaphragm surface EMG, and ECG. Breathing rates were stepped through 10, 12, 14, and 16 breaths per minute over a 5-minute controlled session while patients lay supine. This multi-rate protocol captured respiratory mechanics across physiological frequencies.
Feature Extraction From the physiological signals, time-domain features (mean, standard deviation of respiratory rate), frequency-domain features (power spectral density), and nonlinear features (sample entropy, fractal dimension) were extracted per breathing rate segment. Combined with clinical and imaging variables, a total of 58 features showed significant differences between response groups.
BPNN Architecture A Back Propagation Neural Network with dynamic hidden layers was used as the primary model. BPNN learns nonlinear relationships between input features and treatment response through iterative forward propagation and backpropagation weight updates. The dynamic hidden layer configuration allowed the model to adapt its complexity to the training data rather than using a fixed architecture.
Feature Selection Of the collected features, 34 were identified as relevant for PR prediction and 36 for PD prediction through significance testing. This two-task setup reflects the different biological mechanisms of response and progression - a feature important for PR may be uninformative for PD, so separate feature sets were appropriate.
Class Imbalance Handling With only 10 PD patients versus 41 PR and 69 SD, class imbalance was a concern. Stratified 5-fold cross-validation ensured PD cases appeared proportionally in each fold. Oversampling or synthetic minority techniques may have been applied to prevent the model from being biased toward the majority stable disease class.
Model Validation The dataset was split 70/15/15 for training, validation, and testing using stratified random sampling. Model performance was evaluated on the held-out test set using ROC-AUC, calibration curves, and MSE. Stability was assessed across cross-validation folds to check for overfitting given the small patient count.
PR Prediction Performance For partial response prediction, the best single feature was maximum vertical tumor diameter with AUC 0.699 (95% CI: 0.630-0.757). The comprehensive multimodal BPNN raised this to AUC 0.855 (95% CI: 0.843-0.875) with MSE 0.07 - a substantial improvement demonstrating that no single feature captures the full complexity of RT response.
PD Prediction Performance For progressive disease prediction, the best single feature was zero-crossing ratio of surface EMG signals with AUC 0.750 (95% CI: 0.648-0.841). The BPNN model improved this to AUC 0.929 (95% CI: 0.900-0.960) with MSE 0.01 - an especially high performance that suggests physiological signals carry strong information about which patients will fail RT.
Physiological Feature Contributions Physiological features contributed significantly beyond clinical and radiological data alone. The zero-crossing ratio of surface EMG was the top single predictor for PD, indicating that diaphragm muscle activation patterns measured before treatment encode biologically relevant information about treatment failure risk - a finding with no prior literature precedent.
Clinical Value Pretreatment prediction of RT response with AUC 0.929 for PD is clinically actionable - patients predicted to progress could be directed toward alternative strategies (systemic therapy, combined modality, palliative care) rather than committing to a full RT course with high probability of failure.
Autonomic Nervous System Hypothesis Heart rate variability and respiratory pattern variability reflect autonomic nervous system tone. Cancer patients with more severely disrupted autonomic regulation - particularly reduced parasympathetic activity - have worse survival outcomes across multiple tumor types. Pre-treatment autonomic status may reflect the body's systemic capacity to mount an immune and repair response after radiation injury.
Respiratory Mechanics and Tumor Biology Diaphragm electromyography captures the neuromuscular control of breathing. Abnormal patterns may reflect tumor-related phrenic nerve involvement, pleural effusion, atelectasis, or paraneoplastic effects on the respiratory control system - all conditions associated with more advanced or aggressive disease.
Bioimpedance and Fluid Status Transthoracic impedance changes correlate with changes in thoracic fluid distribution and body composition. In cancer patients, altered fluid balance and cachexia are associated with worse treatment outcomes. Pre-treatment impedance measurements may capture these systemic effects before they become clinically apparent.
Practical Feasibility The physiological signal acquisition protocol required only 5 minutes of non-invasive monitoring with inexpensive electrode equipment. Unlike PET/CT or EBUS, this monitoring adds essentially no procedural risk, minimal patient burden, and low cost - making it potentially accessible even in resource-limited RT settings.
Small Sample Size 120 patients total with only 10 PD cases is severely limited for training and validating predictive models, particularly a BPNN with many parameters. The small PD group makes the AUC 0.929 estimate imprecise and potentially optimistic. External validation on a substantially larger independent cohort is essential.
Single-Center Retrospective Design All patients were treated at one institution with uniform RT protocol and monitoring procedures. Inter-institutional variation in equipment, clinical protocols, and patient characteristics could substantially affect model performance when deployed elsewhere.
Standardization of Physiological Signal Collection The physiological monitoring protocol requires controlled conditions (specific electrode placement, breathing guidance, multiple respiratory rates) that may be difficult to standardize across institutions with different clinical environments. Protocol standardization guidelines and training programs would be needed for multicenter adoption.
Future Directions Future work should validate this approach prospectively in a multicenter trial with a larger PD cohort, investigate which specific physiological features are most reproducible across sites, explore whether post-treatment physiological signals can predict response earlier than imaging, and test the model in SCLC and specific NSCLC subtypes separately.