TLDR: A new explainable AI system combines textual clinical notes with numerical lab results to improve prostate cancer classification. Using BERT for text and Random Forest for numerical data, the model achieved 99% accuracy on a large NIH dataset. An ablation study showed textual features significantly boost recall for intermediate cancer stages, and SHAP values provide clear insights into feature contributions, making the model transparent and clinically valuable.
Prostate cancer remains a significant health concern, ranking as the second most common cancer among men globally. Accurate and timely diagnosis is crucial for effective treatment, yet traditional methods, often relying solely on numerical data like PSA levels and biopsy results, can sometimes lead to inaccuracies or incomplete understanding of a patient’s condition. This highlights a critical need for more refined diagnostic approaches.
A new research paper introduces an innovative explainable AI system designed to enhance prostate cancer classification. This system uniquely combines both numerical clinical features and the rich, often underexplored, textual information found in patient records, such as detailed clinical notes and symptom descriptions. The core idea is that textual data can offer a more nuanced understanding of a patient’s health, complementing the insights gained from numerical measurements.
A Novel Multimodal Approach
The proposed system integrates BERT (Bidirectional Encoder Representations from Transformers) for processing textual features and Random Forest for handling numerical features. BERT is a powerful natural language processing model capable of extracting meaningful context from complex text, while Random Forest is an ensemble machine learning method known for its strong predictive performance and interpretability. This combination allows the model to leverage the strengths of both data types.
The researchers trained this system on a substantial dataset from the PLCO National Institutes of Health (NIH), a part of the U.S. Department of Health and Human Services. This large dataset, containing both numerical and textual patient information, was crucial for developing a robust and reliable model.
Key Methodological Steps
Before feeding data into the model, extensive preprocessing was performed. Numerical data, including PSA levels, age, and BMI, were imputed for missing values using the median. Textual data, such as descriptions of prostate conditions and urination frequency, were concatenated and then processed by the BERT tokenizer to generate embeddings. To manage the high dimensionality of these textual embeddings and prevent overfitting, Principal Component Analysis (PCA) was applied, reducing the data while retaining 98% of its original variance.
Once processed, the numerical and PCA-reduced textual features were combined. The combined dataset was then split into training and testing sets. To address class imbalance, particularly for less common cancer stages, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training data. A Random Forest Classifier was then trained within a pipeline that included these preprocessing steps, ensuring balanced representation and effective learning.
Impressive Performance and Interpretability
The system demonstrated strong predictive performance. Through cross-validation, the model achieved a mean accuracy of 98%, indicating high consistency and reliability. When evaluated on combined features, the model achieved an impressive AUC (Area Under the Curve) of 99%, with precision, recall, and F1-scores of 98%, 84%, and 89% respectively, across multiple cancer stages. The final model on the test set achieved an accuracy of 99%.
A crucial aspect of this research is its focus on explainable AI (XAI). The SHapley Additive exPlanations (SHAP) framework was employed to provide insights into how each feature contributes to the model’s classification decisions. This interpretability is vital for building trust in AI systems, especially in clinical settings. SHAP analysis revealed that PSA levels, particularly from later screenings, were highly influential in the model’s predictions, alongside textual features related to prostate conditions.
An ablation study further underscored the value of multimodal integration. It showed that textual features significantly boosted recall for intermediate cancer stages (Classes 2 and 3). For instance, recall for Class 2 increased from 0.824 (numerical only) to 0.900 (combined features), and for Class 3 from 0.668 (numerical only) to 0.900 (combined features). This demonstrates the complementary nature of textual data in improving the detection of these critical stages.
Also Read:
- Advancing Thyroid Cancer Recurrence Prediction with Bayesian Neural Networks and Explainable AI
- PTCMIL: A Unified Approach to Analyzing Gigapixel Whole Slide Images for Cancer Diagnosis
Comparison and Future Outlook
The proposed model outperformed alternative ensemble models like averaging and stacking, as well as several existing state-of-the-art methods that typically rely on single data types or lack explainable AI components. This highlights the efficacy of fusing textual and numerical data for more accurate and transparent prostate cancer classification.
This work represents a significant step forward in leveraging multimodal AI for clinical decision-making in prostate cancer. The researchers suggest future work could involve exploring additional machine learning algorithms, integrating other diverse data sources such as radiological images (MRI scans) and pathological data, and conducting comprehensive tests on varied datasets to ensure broader applicability. For a deeper dive into the methodology and results, you can read the full research paper here.
