This practical course teaches medical researchers to apply AI to their work. You will learn to identify research problems where deep learning excels at extracting structure from unstructured data, such as patient images, raw clinical notes, or wearable sensor data. You will gain hands-on skills to apply or fine-tune powerful, pre-trained models (like CNNs and Transformers) with clinical or epidemiological case studies, with a focus on interpretation and validation. This includes using explainability techniques to understand why a model makes a prediction, identifying domain shift to see if a model will work on a new population, and testing for data biases that can perpetuate health disparities. You will leave with the ability to build/apply, interpret, and critically validate AI models for medical research.

The course alternates between lectures and practical exercises.

The course consists of 3 days.

Topics of Day 1:Foundations & Translation

• The “Why” & “What” – Study Design & Conceptual Translation
  • Translating Epidemiology to AI
  • The Mechanics of Learning
  • Practical application (afternoon)

Topics day 2: Unlocking Unstructured Data

• “How?” – What data and What methods ?
  • Handling Unstructured Data
  • Transfer Learning & Modern Architectures
  • Practical application (afternoon)

Topics day 3: Validation, Bias and Critical Appraisal :

• For Whom it works? How to improve?
  • Interpretation, Uncertainty & Trade-offs
  • Bias & External Validity
  • Simple nearest neighbor propensity score matching
  • Practical application (afternoon)

1. Identify research problems suitable for deep learning versus traditional machine learning.
   • The student can identify and justify which research problems are best addressed with traditional regression methods and which require Deep Learning approaches, based on the nature of the data (e.g., tabular vs. high-dimensional unstructured data).
   • The student is able to translate epidemiological concepts into machine learning terminology by mapping exposures to features (X) and outcomes to targets/labels (Y) in practical examples.
   • The student can select and apply the appropriate machine learning approach—Supervised Learning (prediction) or Unsupervised Learning (clustering)—to answer specific research questions.
   • The student is able to compare and evaluate the process of manual feature engineering in traditional machine learning with the automatic feature learning capabilities of Deep Learning models.
   • The student can analyze and articulate the data requirements and labeling standards (“Gold Standard” labels) necessary for effective Deep Learning, and explain why these are critical for model performance.
2. Fine-tune pre-trained deep learning models for tasks using health data.
   • The student can identify specific data types (such as raw text, medical images, and complex time-series) where Deep Learning holds a distinct advantage over regression methods.
   • The student is able to explain the specific workflows and pipelines for Natural Language Processing (NLP) and Computer Vision applications in healthcare.
   • The student can apply the principle of Transfer Learning to fine-tune models trained on general data (e.g., Wikipedia) for specific medical tasks.
   • The student is able to utilize domain-specific models (such as BioBERT) to address the challenge of limited labeled data in medical research.
   • The student can train a simple multi-layer neural network on a tabular health dataset (Lab 1).
   • The student is able to deploy a pre-trained Transformer model to extract structured variables from free-text clinical notes (Lab 2).
3. Identify the input features that drive a model’s prediction using explainability techniques.
   • The student can evaluate the trade-offs between model performance (e.g., Accuracy/AUC), interpretability (e.g., risk factors), and computational cost.
   • The student is able to apply explainability techniques, such as attention heatmaps, to identify which words or features influenced a model’s prediction.
   • The student can distinguish between Aleatoric uncertainty (caused by noisy data) and Epistemic uncertainty (caused by lack of knowledge or new patient types).
   • The student is able to estimate model uncertainty using techniques such as Monte Carlo Dropout to address the “overconfidence problem.”
4. Identify domain shift and bias to determine if a model can be applied to a new population.
   • The student can interpret training curves to detect overfitting and identify failures of generalizability.
   • The student is able to define domain shift (e.g., Hospital A vs. Hospital B) and explain its relationship to selection bias and transportability failure.
   • The student can audit a model for bias and generalizability to assess its performance on external populations (Lab 3).

Target audience

Medical and epidemiological researchers interested in using AI as a tool for their research
or
who are interested in the most common applications of AI in epidemiological and medical research

Course prerequisites

The following concepts are assumed known by participants at the start of this course:
basics of python would be an advantage for practicals but not strictly necessary

Course material

The course materials (lectures, assignments, feedback of the assignments etc) are available on Canvas, our digital learning environment. The documents will remain available on Canvas for at least one year.

Literature

Recommended reading

There is no exam for this course

A certificate of participation will be granted to all students who have attended at least 80% of the classes. The number of contact hours will be stated on the certificate.

Only for Dutch medical specialists!

On the final day of the course, you need to sign the attendance list if you wish to obtain KNMG accreditation credits (PE-points).

To qualify for these credits, there is an attendance requirement of 100%.