Rahik-516/clinic_assist_with_online_and_offline_support

Offline-first medical decision support for doctors in low-resource clinics. Diabetes and heart disease risk from clinical inputs, pneumonia patterns from chest X-rays, explained with SHAP, LIME and Grad-CAM. Gemini narrates the evidence in plain language and Bangla when a connection is available — it never makes the call. # Explainable AI for Black-Box Models in Medical Diagnosis Research-driven Streamlit MVP dashboard for clinician decision support across diabetes risk, heart disease risk, and pneumonia pattern detection from chest X-rays. ## Safety Disclaimer This prediction is for decision support only and must be reviewed by a qualified clinician. This system is not an autonomous diagnosis tool. ## Overview This application reuses trained research artifacts and serves inference-time outputs with explainability displays: - Diabetes prediction using saved stacking model + saved imputer/scaler. - Heart disease prediction using saved sklearn pipeline. - Pneumonia image prediction using saved DenseNet121 Keras model. - Rule-based natural-language summaries (deterministic, no LLM usage). - Static SHAP/LIME/Grad-CAM notebook outputs as fallback visual evidence. ## Supported Tasks 1. Diabetes risk prediction from tabular input. 2. Heart disease prediction from tabular input. 3. Pneumonia detection from chest X-ray upload. 4. Explainability rendering for each module. ## Research Background This dashboard is based on a full experimental research pipeline implemented in a Kaggle notebook. The notebook contains the model training code, feature engineering steps, and the explainability experiments (SHAP, LIME, Grad-CAM) together with XAI quality metrics and visual examples. Kaggle notebook (canonical reference): kaggle.com ## Features - Streamlit‑based web dashboard for interactive inference and explainability - Multi‑disease support: Diabetes, Heart Disease, Pneumonia - Rule‑based natural‑language clinical summaries for image/tabular predictions - SHAP explanations for tabular models (when environment supports SHAP) - LIME local explanations for selected examples - Grad‑CAM visualizations for chest X‑ray images (dynamic + static fallbacks) - Confidence‑based interpretation bands and decision thresholds - Static and dynamic …