Peer-reviewed · Explainable ML
Diabetes Risk Prediction Web App
A diabetes risk prediction web app that estimates risk and explains every prediction it makes. The work is published in a peer-reviewed journal (IJSRED, October 2022) and runs as an interactive Streamlit app.
- Peer-reviewed (IJSRED, Oct 2022)
- Deployed as a Streamlit web app
By Adithya Reddy Nalla, Lead AI Engineer in Hyderabad.
Problem
A risk score without a reason is a number someone must take on faith. In healthcare that isn't enough: a clinician or patient needs to know why the model reached its answer.
For screening, the model also has to rank risk well across every threshold, not just score well at one.
Approach
- Analysed the clinical dataset, then trained a Random Forest, which captures non-linear interactions between clinical features and needs no feature scaling.
- Tuned hyperparameters with GridSearchCV using k-fold cross-validation, optimising ROC-AUC so the search rewards good risk ranking rather than raw accuracy.
- Explained predictions with SHAP: a global view of which features drive risk across the population, and a local view of which values raised or lowered one person's score.
- Deployed the model as an interactive Streamlit web app.
Tech stack
- Modelling
- Pythonscikit-learnRandom ForestGridSearchCV
- Explainability
- SHAP
- App
- Streamlit
Results
82%+
Accuracy
0.87
ROC-AUC
1
SHAP explanation per prediction
Research software, not a clinical diagnostic tool.
Links
More projects
- Research framework · Agentic AIGBNI: Generative Brand Narrative IntelligenceGBNI (Generative Brand Narrative Intelligence) is a state-aware multimodal AI agent framework for emotionally adaptive brand communication, on SSRN.
- Self-directed R&D · SecurityMITHRA: Cybercrime Rapid-Response and Intelligence PlatformMITHRA is Adithya Reddy Nalla's self-directed R&D platform for responding to cybercrime quickly and turning incident signals into usable intelligence.