TL;DR
This study explores the potential of combining traditional Chinese medicine (TCM) tongue diagnosis with Western medicine tumor markers to predict the stages of non-small cell lung cancer (NSCLC). By applying various machine learning models, the research found that integrating tongue features, tumor markers, and baseline patient data significantly improved prediction accuracy, with neural networks showing the most promise.
What Problem Does This Paper Address?
Accurately staging non-small cell lung cancer (NSCLC) is crucial for effective treatment and improving patient outcomes. Current methods, primarily imaging and histology, can be invasive, complex, and costly. This creates a need for non-invasive, reliable, and accessible diagnostic tools. Traditional Chinese Medicine (TCM) has long used tongue diagnosis to assess overall health, and emerging research suggests a correlation between tongue characteristics and various diseases. This paper investigates whether these tongue features, when combined with established tumor markers, can provide a more comprehensive and less invasive way to differentiate between early and advanced stages of NSCLC.
Methods
The researchers employed a multi-faceted approach:
- Tongue Feature Extraction: A specialized digital tongue and face diagnosis instrument (TFDA-1) was used to capture tongue images. These images were then analyzed using the Tongue Diagnosis Analysis System (TDAS) to extract objective features related to color, texture, and shape. Features were categorized into "Tongue Body" (TB) and "Tongue Coating" (TC) and analyzed across different color spaces (RGB, HIS, Lab, YCrCb) and texture indices.
- Tumor Marker Data: Standard tumor marker data (e.g., CEA, CA125, NSE) was collected from patient medical records.
- Machine Learning Models: Six common machine learning classifiers were utilized: Decision Tree, Support Vector Machine (SVM), Random Forest, Naive Bayes, Neural Network, and Logistic Regression.
- Dataset Construction: Models were trained and evaluated on various combinations of data: "baseline" (age, sex), "tumor marker," "tongue feature," "tongue feature and baseline," "tongue feature and tumor marker," and "tongue feature, tumor marker, and baseline."
- Evaluation Metrics: Model performance was assessed using Accuracy, Precision, F1-score, Sensitivity, Specificity, and Area Under the ROC Curve (AUC). Ten-fold cross-validation and repeated sampling were used to ensure robust evaluation.
Datasets
The study utilized data from 324 lung cancer patients at Longhua Hospital Affiliated to Shanghai University of Traditional Chinese Medicine. Of these, 219 NSCLC patients were included:
- Non-advanced NSCLC group: 109 patients (stages I, II, and III combined).
- Advanced NSCLC group: 110 patients (stage IV).
The datasets used for modeling included:
- Baseline data (age, sex)
- Tumor marker data
- Tongue feature data (derived from tongue images)
No new benchmark datasets were introduced; the study focused on analyzing existing data types in novel combinations.
Key Insights and Ideas
- Tongue as a Diagnostic Window: The study reinforces the idea that tongue appearance is not merely superficial but reflects underlying physiological changes, showing statistically significant differences between early and advanced NSCLC stages.
- Synergy of TCM and Western Medicine: The core insight is that combining TCM's tongue diagnostics with Western medicine's tumor markers offers a more powerful predictive capability than either alone. This bridges the gap between traditional and modern medical approaches.
- Multidimensional Data is Key: Relying on a single data source (tongue features or tumor markers) proved insufficient for accurate staging. The best results were achieved when integrating multiple data types, highlighting the complexity of disease and the need for holistic assessment.
- Neural Networks Shine: Among the tested classifiers, neural networks demonstrated superior performance, particularly when integrating all available data sources, suggesting their potential for complex pattern recognition in medical diagnostics.
Key Results
- Significant Tongue Differences: Statistically significant differences in tongue features were observed between non-advanced and advanced NSCLC groups, particularly in color space indices (TB-B, TB-H, TC-H, TB-L, TB-a, TC-b, TB-Y, and TC-Cb).
- Stronger Correlations in Advanced Stages: The advanced NSCLC group showed a higher number and stronger correlations between tongue features and tumor markers compared to the non-advanced group.
- Model Performance:
- Models relying solely on tongue features or tumor markers showed poorer classification results.
- The best performing models integrated multiple data sources.
- The neural network model using "tongue feature and tumor marker and baseline" data achieved the highest classification accuracy of 0.767 ± 0.081 and an AUC of 0.793 ± 0.086.
- Random Forest and Naive Bayes also performed well with combined data, achieving accuracies of 0.718±0.062 and 0.688±0.070, respectively.
- SVM, Decision Tree, and Logistic Regression generally performed less effectively.
Limitations and Future Work
- Pilot Study Limitations: The authors acknowledge that this is a pilot study with limited sample size. The findings require further verification with larger, more diverse cohorts.
- Single Data Source Insufficiency: The study concludes that single data sources (baseline, tongue feature, or tumor marker alone) are insufficient for accurately identifying NSCLC stages.
- Future Directions: The paper suggests that while combining tongue data with tumor markers and baseline information can improve differential diagnosis efficiency, further validation with larger sample sizes is essential. The potential of logistic regression, alongside other machine learning methods, for improving diagnostic efficiency by incorporating tongue image data warrants further investigation.
Summary updated