NMS_AI / README.md
plutoa386's picture
fix
7860dc0
|
Raw History Blame Contribute Delete
4.21 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade
metadata
title: NMS AI
emoji: 🧠
colorFrom: red
colorTo: pink
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false

🧠 NMS Dementia Risk Assessment API

AI Backend for the NMS Mobile App & NMS Web Platform
3rd Year Uni Project

This repository contains the agentic AI backend powering dementia risk assessment and personalized news delivery for the NMS ecosystem. Built on CrewAI, it provides scalable, explainable, and audience-aware intelligence for patients, caregivers, and clinicians.

This is my contribution to the NMS group project.


βœ… Core Capabilities

1. Dementia Risk Prediction

  • Audio-Based Assessment: Analyzes speech patterns (MFCCs, prosody, articulation) from patient-recorded audio to predict dementia risk.
  • Lifestyle-Based Assessment: Evaluates 16+ clinical and behavioral factors (e.g., sleep, diet, cardiovascular health, education) for holistic risk profiling.
  • SHAP-Powered Explainability (XAI): Every prediction includes:
    • Top contributing features (e.g., "↓ SleepQuality (3/5) β†’ ↑ risk (+0.18)")
    • Plain-language, actionable insights for patients and clinicians.

2. Personalized News Curation

  • Patient-Focused News: Scrapes trusted caregiving/dementia advocacy sites (e.g., caregiving.org, alz.org) and generates empathetic, jargon-free summaries.
  • Clinician-Focused News: Retrieves and summarizes peer-reviewed research and policy updates (e.g., from open-access sources), emphasizing clinical relevance and evidence.

πŸ€– Agent Architecture (CrewAI)

The system uses 4 specialized agents, each with dedicated tools and LLMs:

Agent Role LLM Tools
Risk Calculator (Audio) Runs XGBoost model + SHAP on speech features gpt-oss-120b calculate_dementia_risk_audio, explain_dementia_risk_audio
Risk Calculator (Lifestyle) Runs XGBoost model + SHAP on lifestyle data gpt-oss-120b calculate_dementia_risk_lifestyle, explain_dementia_risk_lifestyle
Patient News Agent Fetches, scrapes, and summarizes news for patients/caregivers Qwen3-4B-Thinking-2507 gather_news_patient
Doctor News Agent Fetches, scrapes, and summarizes research for clinicians Qwen3-4B-Thinking-2507 gather_news_doctor

πŸ”‘ Why agents?

  • Separation of concerns (prediction vs. information retrieval)
  • Independent scaling & debugging
  • Audience-tailored language (no technical terms for patients; evidence-first for doctors)

πŸ› οΈ Key Technologies

  • Machine Learning: XGBoost (v3.0.5) for high-accuracy binary classification
  • XAI (Explainable AI): SHAP TreeExplainer for model-agnostic, global + local interpretability
  • Web Scraping: requests + BeautifulSoup4 (with fallbacks for dynamic sites)
  • LLM Orchestration: LiteLLM + Hugging Face Inference API (Qwen/Qwen3-4B-Thinking-2507, openai/gpt-oss-120b)
  • Data Pipeline: Firebase (Firestore for lifestyle, Storage for audio/transcripts)
  • Deployment: Hugging Face Spaces (Gradio frontend + CrewAI backend)

πŸ“‘ API Endpoints (via Gradio)

Endpoint Input Output
/assess_risk { "docID": "umWSG...", "mode": "audio"|\"lifestyle" } Risk level, confidence, SHAP explanations
/get_news { "mode": "patientnews"|\"doctornews", "count": 5 } List of {title, summary, url} articles

πŸ” Ethical & Technical Notes

  • All predictions undergo SHAP-based validation to ensure model fairness and transparency.
  • No personally identifiable information (PII) leaves Firebase β€” all processing is anonymized.

‼️ Disclaimer

  • This README was generated by AI
  • The webscraping is mid

πŸ’‘ Developed by plutoa386
🌐 Live Demo: Hugging Face Space
πŸ“¦ Models: modelAudiov8_ML.json, modelLifestyleV3_ML.json (XGBoost + SHAP-compatible)
πŸ‘Ύ Old Messy Model Training Repo: GitHub