Multi-Disease Prediction

About the multi-disease prediction M.Tech project

Diabetes, heart disease and kidney disease often appear together and share warning signs such as high blood sugar, blood pressure and age. This multi-disease prediction M.Tech project puts all three checks into one Flask web app. A doctor enters the patient's values once and gets a risk value for each disease, with a SHAP chart showing which values pushed it up or down.

The research angle goes past "which model scores highest". It asks whether a stacked ensemble can give risk values that behave like real probabilities, and whether its explanations stay the same when the training data changes slightly.

Research behind it

Base paper. The work builds on Enhancing heart disease prediction with stacked ensemble and MCDM-based ranking: an optimized RST-ML approach, published in Frontiers in Digital Health (2025). It reduces features with Rough Set Theory, builds several stacks grouped by correlation, ranks them with TOPSIS and explains the chosen stack with SHAP and LIME. Heart disease is the main dataset, with kidney and other datasets used as generalisation checks.

Research gap. That pipeline is tuned for one disease, and diabetes is not among its datasets. A clinic needs one tool for several linked diseases. A stack can also score well and still give over-confident probabilities, which makes a "70 out of 100 risk" misleading. Finally, a doctor needs reasons that do not change every time the model is retrained.

Proposed system and novelty. Five base learners are trained for each disease, and only those whose out-of-fold predictions disagree enough are kept, so the stack gains real diversity. Isotonic calibration then maps the output to a risk value that matches observed rates. SMOTE and imputation run inside each fold, which avoids data leakage. SHAP top features are compared across folds to measure explanation stability.

Models compared. Logistic Regression, Random Forest, XGBoost, a soft voting ensemble and the base paper stack re-implemented, against the proposed calibrated stack, on the Pima diabetes, heart disease and UCI CKD datasets with repeated stratified cross-validation and a Wilcoxon test. The ablation removes diversity selection, removes calibration and moves SMOTE outside the folds.

If you want the same diseases trained across hospitals without sharing records, look at the federated learning project for multi-hospital disease prediction.

How it works

• The doctor adds a patient and fills one form with routine health values.
• Values are checked for range and missing fields, then scaled as in training.
• Each disease stack returns a calibrated risk value and a band.
• SHAP shows the effect of each value as a bar chart.
• The doctor writes advice and downloads a PDF report.
• The admin views model comparison and calibration curves.

Users and modules

There are three roles. The doctor handles patients, runs the three checks, reads the SHAP reasons and prints reports. The patient logs in to see reports, reasons in plain words and check-up reminders. The admin manages users and records, reviews the dashboard, model comparison and calibration pages, and sets the risk band limits. Everything trains in minutes on a CPU laptop.

What you get

• Phase 1 Research: Base paper study: RST and TOPSIS-ranked stacked ensemble (Frontiers in Digital Health, 2025); Literature survey of 15-20 recent disease prediction papers; Research gap, novelty and objectives; EDA notebooks for diabetes, heart and kidney data; Synopsis and Phase-1 report.
• Phase 2 Implementation: Flask web app with doctor, patient and admin panels; Leakage-free imputation and SMOTE pipeline; Baselines, voting and base stack re-implementation; Proposed diversity-aware calibrated stack; SHAP module with stability check; MySQL database and trained models.
• Results: Accuracy, recall, F1, ROC-AUC and Brier score per disease; Calibration curves and Wilcoxon test; Ablation graphs and SHAP stability table; results.xlsx and all graphs in high resolution.
• Thesis and paper: Thesis written for you (about 90 pages); IEEE conference paper draft; Journal paper draft; Publication guide.
• Reviews, viva and support: PPTs for Review 1, Review 2 and final viva; Viva book with 50+ questions; A1 poster and project diary; Demo and code walkthrough videos; 1-on-1 setup, 2 revisions and support till viva.

Who can use this project

This project fits M.Tech and M.E. students of CSE and IT who want a healthcare machine learning dissertation without heavy image training. It also matches searches for multi-disease prediction using machine learning, a stacking ensemble XGBoost project or a SHAP explainable machine learning project. The pipeline moves cleanly from notebook to web app, so each step is easy to show your guide.