Developing Multimodal Artificial Intelligence-Based Tool for Enhancing Clinical Decision- Making in Personalised Prostate Cancer Treatment

Technical Summary: Multimodal AI-Based Tool for Personalised Prostate Cancer Treatment

Problem Statement
Prostate cancer (PCa) remains a leading malignancy with rising incidence, particularly in Asia. Current clinical decision-making relies heavily on staging and Prostate-Specific Antigen (PSA) levels, often resulting in suboptimal outcomes and high variability in treatment selection among oncologists. While guidelines exist (e.g., NCCN), the integration of complex, heterogeneous data—spanning clinical history, genomics, imaging, and longitudinal laboratory trends—into a cohesive, personalized treatment strategy remains a challenge. Existing AI tools often lack real-world applicability or fail to integrate multimodal data effectively to support precision oncology.

Methodology
This retrospective cohort study (2018–2024) involved 212 prostate cancer patients at the Brunei Cancer Centre (TBCC), generating 478 longitudinal observations. The study adhered to the TRIPOD-AI checklist and employed a comprehensive multimodal AI framework:

1. Data Integration and Preprocessing:

   • Data Sources: Integrated 82 variables including demographics, clinical history, genomic profiles (NGS), treatment records, and 160 whole-body PSMA PET/CT scans (PNG format).
   • Preprocessing: Missing values in static and longitudinal variables were imputed using an Extremely Randomized Trees-based iterative method. Categorical variables were encoded, and continuous variables were normalized.
   • Longitudinal Structuring: Patient data were organized into time-ordered sequences to capture temporal dynamics.

2. Unsupervised Learning:

   • Applied to explore latent structures and natural patient groupings without predefined outcomes. Techniques included time-independent and time-dependent clustering (e.g., DTWI-HAC). Results indicated a homogeneous structure with limited separation, reflecting the intrinsic similarity of the cohort but highlighting the difficulty of clustering with small sample sizes.

3. Supervised Learning (Tabular Data):

   • Representation Learning: A recurrent autoencoder based on Gated Recurrent Units (GRUs) was trained to model temporal patterns in laboratory and treatment trajectories.
   • Class Balancing: Synthetic Minority Oversampling Technique (SMOTE) was applied exclusively to the training embeddings to address class imbalance without distorting temporal structures.
   • Model Architectures: Evaluated classical models (Logistic Regression, KNN, SVM, Random Forest, XGBoost, LightGBM, CatBoost) and deep learning architectures (BiLSTM, Temporal Convolutional Networks, Transformers).
   • Hybrid Approach: A novel hybrid model combined an LSTM encoder (for temporal summarization) with a Random Forest classifier (for nonlinear decision boundaries).

4. Imaging Analysis:

   • A Simple Convolutional Neural Network (CNN) was developed to process PSMA PET/CT scan images.

5. Multimodal Late-Stage Fusion:

   • The best-performing tabular model (LSTM + Random Forest) and the CNN imaging model were integrated using decision-level fusion.
   • Fusion Strategies: Weighted averaging, majority voting, and stacking (meta-learner) were compared.

6. Interpretability:

   • SHAP (SHapley Additive exPlanations) values were used to visualize global feature importance and provide patient-level explanations for the best-performing models.

Key Results

• Unsupervised Learning: Models failed to form distinct, clinically useful clusters, dominated by a single large cluster, likely due to the small sample size and high feature correlation.
• Supervised Tabular Models:
  • The Hybrid LSTM–Random Forest architecture outperformed other models, achieving an overall AUC of 0.749, accuracy of 0.418, sensitivity of 32%, and specificity of 83%.
  • Performance varied by treatment category: The model showed strong discrimination for Active Surveillance (AUC=0.92) and ARPI-based combination therapy (AUC=0.92), but lower performance for ADT monotherapy (AUC=0.51), reflecting clinical ambiguity in that specific decision space.
  • Traditional models (e.g., Logistic Regression, SVM) performed near random classification levels.
• Multimodal Fusion:
  • Late-stage fusion significantly improved discriminative ability.
  • The Stacking (Meta-Learner) fusion model achieved the highest overall AUC of 0.971 with balanced sensitivity (0.429).
  • Weighted Averaging also improved performance (AUC=0.964) with very high specificity (0.929), offering a conservative alternative.
• Interpretability: SHAP analysis identified PSA-related metrics (mean PSA, post-treatment change), hemoglobin, phosphate, and inflammatory markers as dominant drivers. Tumor stage and risk groups were confirmed as key static predictors, aligning with established guidelines.

Significance and Claims
The paper claims that developing a multimodal AI-based Clinical Decision Support System (AI-CDSS) offers a promising strategy for enhancing personalized prostate cancer treatment in Brunei. Key contributions include:

• Temporal Modeling: Demonstrating that modeling the longitudinal nature of disease progression (via LSTM autoencoders) is essential for capturing evolving biomarkers like PSA trajectories, which cross-sectional models miss.
• Multimodal Synergy: Showing that late-stage fusion of sequential clinical data and imaging features (PSMA PET/CT) significantly enhances predictive discrimination compared to single-modality approaches.
• Explainability: Providing a framework for transparent decision-making through SHAP, which supports clinical interpretability by linking model outputs to specific patient characteristics and biomarkers.
• Role as Support Tool: The authors modestly position these models as decision support tools rather than replacements for clinician judgment. They emphasize that while the models show promise, they require further validation in larger, prospective, multi-institutional studies to address limitations such as small sample size, low sensitivity, and potential overfitting in late-stage fusion models.

The study concludes that while current limitations exist, the integration of longitudinal representation learning and multimodal fusion represents a viable path toward optimizing personalized oncology care, provided future work focuses on robust external validation and clinical implementation.