Trustworthy Medical AI: Predictive Modelling, Imaging & Reliability Assessment

Aim

Workshop Objectives

Workshop Structure

Workshop Structure

📅 Day 1: Clinical AI Foundations & Predictive Modelling

Focus: Building reproducible machine-learning workflows for clinical data.

• AI in Medicine; AI vs ML vs DL and healthcare data types.
• Clinical problem formulation, supervised learning and classification workflows.
• Training, validation, testing, overfitting and data leakage prevention.
• Clinical outcome definition, predictor availability and dataset limitations.
• Data exploration, missing-value handling, encoding and scaling.
• Reproducible preprocessing using Pipeline and ColumnTransformer.
• Stratified, patient-level, temporal and site-based validation.
• Logistic Regression, XGBoost, cross-validation and model comparison.
• Evaluation metrics: accuracy, precision, recall, sensitivity, specificity and confusion matrix.
• Clinical decision support and responsible interpretation of predictions.

🛠️ Hands-on: Clinical Prediction Modelling

Prepare the UCI Cleveland Heart Disease dataset, build Logistic Regression and XGBoost pipelines, and compare models using training-only cross-validation. Reserve the test set for Day 3, generate cross-validated predictions, calculate performance metrics, identify important features and discuss clinical suitability and leakage risks.

🧰 Tools Covered: Google Colab, Python, pandas, scikit-learn, XGBoost, matplotlib.

📅 Day 2: Medical Imaging & Explainable Deep Learning

Focus: Applying pretrained deep-learning models to medical images and interpreting their predictions.

• Medical imaging modalities, AI applications, classification, detection and segmentation.
• CNNs, visual feature learning, transfer learning and pretrained models.
• Self-supervised learning and foundation models in medical imaging.
• Challenges: limited datasets, class imbalance, overfitting, data leakage and domain shift.
• False positives, false negatives and explainability in medical diagnosis.
• Grad-CAM heatmaps, pretrained chest X-ray models and pathology scores.
• Limitations of visual explanations and AI-assisted versus autonomous diagnosis.

🛠️ Hands-on: Medical Image Prediction & Explainability

Explore PneumoniaMNIST through a guided classification workflow using a pretrained model. Compare predictions with actual labels and examine misclassifications. Use a separate suitable chest X-ray with TorchXRayVision to generate pathology scores and a Grad-CAM heatmap, interpreting results in light of domain shift and visualization limitations.

🧰 Tools Covered: Google Colab, Python, PyTorch, torchvision, MedMNIST, TorchXRayVision, pytorch-grad-cam.

📅 Day 3: Generative AI, Model Reliability, Bias Assessment & Governance

Focus: Building grounded healthcare AI applications and assessing model reliability, explainability and responsible deployment.

• Healthcare applications of Generative AI and LLMs: documentation, literature, patient information, decision support and research.
• Retrieval-Augmented Generation (RAG), grounding and hallucination risks.
• Probability calibration, validation-based thresholds and Brier score.
• Held-out evaluation, ROC-AUC, precision–recall and performance uncertainty.
• SHAP explanations, correlated features, and association versus causation.
• Subgroup assessment, sample-size limitations and model fairness.
• External validation, pre-market evaluation and post-market monitoring.
• Transparency, human oversight, patient-data privacy, WHO/FDA guidance, ethics and responsible deployment.

🛠️ Hands-on: Healthcare RAG Assistant & Clinical AI Reliability Assessment

Build a mini healthcare RAG assistant using public medical documents, embeddings, a searchable vector knowledge base and grounded answer generation. Identify unsupported answers and potential hallucinations. Calibrate the Day 1 model using development data, evaluate the locked model on the reserved test set, generate SHAP explanations, assess subgroup performance using Fairlearn, and draft a clinical AI governance checklist.

🧰 Tools Covered: Google Colab, Python, open-source embedding/LLM components, scikit-learn, CalibratedClassifierCV, SHAP, Fairlearn, pandas, matplotlib.

Who Should Enrol?

Important Dates

Workshop Outcomes