Capstone Assignment

Overview

This capstone course gives you the opportunity to bring everything you have learned in the Informed Clinical Decision Making using Deep Learning Specialization together in one hands-on, practical project. You will work with real-world critical care data from the MIMIC-III database and tackle a clinically meaningful prediction task from start to finish.

You will choose one of three advanced projects focused on explainable artificial intelligence in healthcare:

permutation feature importance, LIME, or Grad-CAM. Each project guides you through querying and preparing electronic health record data, building predictive models such as logistic regression or LSTM networks, and interpreting model predictions using state-of-the-art explainability techniques.

The focus is not only on model performance, but on understanding and communicating why a model makes its predictions. By completing this capstone, you will gain practical experience translating deep learning models into insights that support trustworthy and transparent Clinical Decision Support Systems.

This course is ideal for learners who want to demonstrate applied skills, build confidence working with clinical data, and showcase their ability to combine technical expertise with clinical reasoning.

Syllabus

• Permutation feature importance on the MIMIC critical care database
This is an advanced exercise/lesson that combines knowledge from the three earlier modules: 1) 'Data mining of Clinical Databases' to query the MIMIC database, 2) 'Deep learning in Electronic Health Records' to pre-process EHR and build deep learning models and 3) 'Explainable deep learning models for healthcare' to explain the models decision. In particular, permutation feature importance is implemented and applied on MIMIC-III extracted datasets. The technique is applied both on logistic regression and on an LSTM model. The explanations derived are global explanations of the model.
• LIME on the MIMIC critical care database
This is an advanced exercise/lesson that combines knowledge from the three earlier modules: 1) 'Data mining of Clinical Databases' to query the MIMIC database, 2) 'Deep learning in Electronic Health Records' to pre-process EHR and build deep learning models and 3) 'Explainable deep learning models for healthcare' to explain the models decision. In particular, LIME is applied on MIMIC-III extracted datasets. The technique is applied on both logistic regression and an LSTM model . The explanations derived are local explanations of the model.
• Grad-CAM on the MIMIC critical care database
This is an advanced exercise/lesson that combines knowledge from the three earlier modules: 1) 'Data mining of Clinical Databases' to query the MIMIC database, 2) 'Deep learning in Electronic Health Records' to pre-process EHR and build deep learning models and 3) 'Explainable deep learning models for healthcare' to explain the models decision. In particular, GradCam is implemented and applied on an LSTM model that predicts mortality based on MIMIC-III extracted datasets. The explanations derived are local explanations of the model.

---

Taught by

Fani Deligianni

---

Subjects

Computer Science