HEALTH AI / RESEARCH
Predictive Models and Explainable AI for Cardiovascular Telerehabilitation
This thesis project focused on the implementation of three predictive models to address a binary classification problem in the context of cardiovascular disease diagnosis and rehabilitation.
The research was developed in collaboration with Dedalus Italia within the RE-ACTS (REspiratory and Cardiac Telerehabilitation integrated home Services) project, supporting leading hospitals such as Maria Cecilia Hospital and Casa di Cura Sant’Eremo.
The core objective was to provide reliable tools to assist medical centers and clinicians in identifying and supporting patients at risk or affected by cardiovascular diseases, leveraging both clinical and non-clinical data for robust prediction and personalized rehabilitation planning.
Project Scope & Data Acquisition
The main challenge was acquiring and analyzing high-quality data to ensure robust model development.
- Significant effort was dedicated to statistical analysis of the available datasets to identify and resolve data quality issues.
- Three separate datasets were analyzed, including both clinical and non-clinical information, to ensure generalizable results.
- The work aimed to support both resource-efficient and data-rich healthcare environments, balancing the involvement of medical staff and institutional resources.
Model Development
Three predictive models were designed and developed, one for each dataset:
- Explored several machine learning algorithms, including Random Forest, XGBoost, Support Vector Machine (SVM), and AdaBoost.
- Selected and fine-tuned the best-performing configurations via hyperparameter optimization.
- The diversity of models allowed the analysis of clinical and non-clinical data separately, enhancing the applicability across various healthcare settings.
Explainable Artificial Intelligence
To promote model transparency and clinical adoption, Explainable Artificial Intelligence (XAI) techniques were integrated:
- Employed LIME and Counterfactual Explanations to make model predictions interpretable and trustworthy for clinicians.
- XAI methods strengthened confidence in the predictions and enabled new clinical applications, such as personalized rehabilitation plans based on 5-year risk scenarios.
- The use of counterfactual explanations facilitated the design of tailored interventions and actionable feedback for both patients and healthcare providers.
Results & Impact
The models demonstrated highly promising results:
- In the two main models, average accuracy, precision, and recall reached 95%, 94%, and 95%, respectively.
- The solutions offer a concrete tool for early risk assessment and personalized care in cardiovascular telerehabilitation.
- This work lays the foundation for further clinical validation and integration of explainable AI models in real-world healthcare workflows.