Project Overview
This project combines clinical neuropsychology and explainable AI to rethink how medical data is interpreted. Instead of treating datasets as static tables, the pipeline models them as dynamic patient profiles, preserving clinical meaning throughout the analysis.
The focus was on building a reliable and clinically grounded system. A LightGBM model validated as the final solution to reduce overfitting and ensure stable performance on unseen patients.
The result is a transparent Clinical Decision Support System (CDSS) prototype designed to assist medical judgment, with an emphasis on interpretability and real-world applicability.
Key Milestones
One of the most important outcomes of this project was achieving a strong balance between performance and reliability. The model reached a 94.4% AUC-ROC and a 93.7% F1-score, providing accurate dementia classification while maintaining consistent behavior on unseen patient data.
A key focus was making the system understandable, not just accurate. By integrating SHAP-based explanations, each prediction can be broken down into the specific biomarkers that influenced it, allowing clinicians to move beyond “black-box” outputs and actually interpret the reasoning behind each assessment.
To make these insights accessible in practice, the project includes an interactive Streamlit dashboard. Through visual tools such as radar charts, clinicians can explore how individual patients compare to broader population patterns, supporting a more intuitive and data-driven evaluation process.
Finally, the entire pipeline was built with a strong emphasis on robustness and reproducibility. From bias analysis to model comparison and statistical validation, each step was designed to ensure that the system performs reliably across different patient groups and remains grounded in real clinical scenarios.
For a deep dive into my analytical process, the GitHub repository includes all the Jupyter Notebooks, preprocessing steps, and detailed statistical analyses conducted throughout the project.