Aoun (عون) — Intelligent Cancer Patient Monitoring System

AI-powered platform for continuous cancer patient monitoring, risk stratification, and intelligent medical assistance — fully local, no cloud dependency.

---

What is Aoun?

Cancer patients spend most of their time between clinic visits with no intelligent monitoring. Aoun closes that gap.

It is a mobile health platform that lets patients log daily symptoms, get AI-generated risk assessments, perform facial wellness checks, and consult an intelligent health chatbot — while their oncologist receives real-time alerts and monitors risk trends from a dedicated dashboard.

Everything runs locally. No cloud. No API keys. Patient data never leaves the device.

---

Features

Feature	Description
🧠 ML Risk Prediction	Random Forest evaluating 23 clinical features. 92% accuracy on a held-out test set of 148,746 records. Classifies risk as High / Medium / Low.
💬 RAG Medical Chatbot	Mistral-7B via Ollama + ChromaDB vector store with 254 indexed entries from a curated FAQ and NCI/CancerCare clinical PDFs. Feedback-weighted retrieval that improves with use.
📷 Facial Wellness Analyzer	OpenCV pipeline extracting pallor, eye fatigue, skin uniformity, and skin dullness from selfie photos to detect physical illness signs. No training data needed.
🔔 Real-time Doctor Alerts	Firebase Cloud Messaging pushes High-risk patient updates to the doctor's device instantly.
📊 Doctor Triage Dashboard	Patients sorted by ML risk level. Full symptom history. ML feedback buttons (correct / incorrect) for model improvement loop.
🚨 Emergency Screen	Symptom-specific guidance cards with one-tap emergency calling. Works offline.
🔁 Learning Loop	Doctor and patient feedback updates RAG source weights. Poorly rated sources are penalized or blocked.

---

Tech Stack

Layer	Technologies
Mobile	Flutter 3.41 · Dart · fl_chart · Firebase Messaging
Backend	FastAPI · Python 3.11 · PostgreSQL · SQLAlchemy · Uvicorn
ML Model	scikit-learn · Random Forest · pandas · NumPy
Computer Vision	OpenCV · Haar Cascade · LAB color space
RAG Chatbot	Ollama Mistral-7B · ChromaDB · all-MiniLM-L6-v2 (384-dim)
Knowledge Base	pypdf · NCI / CancerCare clinical PDFs · 254 indexed entries
Notifications	Firebase Cloud Messaging (FCM)

---

System Architecture

┌─────────────────────────────────────────────────┐
│                  Flutter App                    │
│  Patient · Doctor · Admin — 13 screens          │
└───────────────────┬─────────────────────────────┘
                    │ HTTP REST
┌───────────────────▼─────────────────────────────┐
│              FastAPI Backend  :8002              │
│                                                 │
│  ┌─────────────┐  ┌──────────────────────────┐  │
│  │ ML Service  │  │     RAG Chatbot           │  │
│  │ :8001       │  │  Mistral-7B + ChromaDB    │  │
│  │ Random      │  │  254 medical entries      │  │
│  │ Forest 92%  │  │  Feedback-weighted        │  │
│  └─────────────┘  └──────────────────────────┘  │
│                                                 │
│  ┌─────────────┐  ┌──────────────────────────┐  │
│  │   OpenCV    │  │      PostgreSQL           │  │
│  │   Facial    │  │      Database             │  │
│  │   Wellness  │  │      Port 5432            │  │
│  └─────────────┘  └──────────────────────────┘  │
└───────────────────┬─────────────────────────────┘
                    │
         ┌──────────▼──────────┐
         │   Firebase FCM      │
         │   Doctor Alerts     │
         └─────────────────────┘

---

ML Model Details

• Algorithm: Random Forest Classifier (100 estimators)
• Pipeline: SimpleImputer (median) → StandardScaler → RandomForestClassifier
• Features: 23 — Age, Gender, 21 symptom / risk-factor indicators
• Training data: 1,000 real Kaggle oncology records, class-balanced with SMOTE on the training split only → 594,982 train / 148,746 test (80/20 stratified). The test set is excluded from SMOTE.
• Accuracy: 92.0% (0.9201) · Precision: 0.9200 · Recall (macro): 0.9195 · F1: 0.9197
• Output: Risk level (High / Medium / Low) + confidence + top contributing features
• Top feature: Coughing of Blood
• Inference: <200 ms · Serialized to aoun_model.pkl

Note on scale: the source symptoms are scored 1–8. The patient-facing app sliders use a 0–10 scale for usability; a value of 0 is mapped to 1 at inference to stay within the model's valid input range.

Confusion Matrix (148,746 test patients)

	Predicted High	Predicted Medium	Predicted Low
Actual High	49,929	1,815	2,004
Actual Medium	2,223	41,371	1,966
Actual Low	2,094	1,782	45,562

---

Facial Wellness Analyzer

Four clinical features extracted from a selfie photo using OpenCV:

Feature	Method	Weight
Pallor Score	LAB color space A-channel mean	40%
Eye Fatigue	Haar cascade height/width ratio	35%
Skin Uniformity	LAB L-channel standard deviation	15%
Skin Dullness	LAB L-channel mean	10%

Output: Illness score 0–100 → Appears Well / Mild Fatigue / Appears Unwell

Evaluation: 88% face-detection success (44/50 test photos) · 100% EXIF orientation correction (12/12). The _fix_rotation() routine uses the Haar cascade as a rotation oracle, so no EXIF parsing library is required.

Grounded in: Forte et al. (2021). Deep Learning for Identification of Acute Illness and Facial Cues of Illness. Frontiers in Medicine, 8, 661309.

Proof-of-concept only — not clinically validated. Formal validation with oncology patients requires medical ethics board approval and is documented as future work.

---

RAG Chatbot Knowledge Base

Three source types are combined at query time and stored in a ChromaDB vector store (all-MiniLM-L6-v2, 384-dim). Every response carries citation chips tracing the answer back to its source.

Source	Type	Notes
Curated FAQ (knowledge_base.py)	Structured	20 evidence-based Q&A entries across 7 categories, adapted from ACS / Macmillan / Cancer Research UK
NCI Chemotherapy and You	Real PDF	Ingested page-by-page (~900-char chunks, 150-char overlap)
CancerCare Chemo Side Effects	Real PDF	Neutropenia monitoring, treatment-specific symptoms
Live patient context	PostgreSQL	Built per-message from 5 tables (profile, latest symptoms, 7-day trends, lifestyle averages, unread alerts)
Conversation memory	ChromaDB	Past turns retrieved semantically for multi-turn coherence
Total indexed entries		254

Retrieval: Over-fetch 3× → re-rank by feedback weights → inject into the Mistral-7B prompt alongside live patient data from PostgreSQL. Thumbs-up: +5% boost per vote; thumbs-down: −10% penalty per vote; net rating ≤ −3 permanently blocks a source.

---

Getting Started

Prerequisites

Python 3.11+
Flutter 3.41+
PostgreSQL 18
Ollama (https://ollama.com)

1. Clone and set up backend

git clone https://github.com/<your-org>/aoun.git
cd aoun/backend
python -m venv venv
venv\Scripts\activate        # Windows
pip install -r requirements.txt

2. Train the ML model

The trained model files are not included in the repo. Generate them from the notebook:

1. Download the "Cancer Patient Data Sets" CSV from Kaggle (see Dataset section).
2. Open gp2.ipynb (Google Colab or Jupyter) and point the read_csv path at your CSV.
3. Run all cells. This produces:
     aoun_model.pkl, aoun_scaler.pkl, aoun_imputer.pkl,
     aoun_label_encoder.pkl, aoun_features.csv
4. Copy those 5 files into the ml_service/ folder.

The ML service loads these on startup, so it will not run until they are present.

3. Set up database

psql -U postgres -c "CREATE DATABASE aoun_db;"
psql -U postgres -d aoun_db -f database/aoun_schema.sql
psql -U postgres -d aoun_db -f database/chat_feedback_migration.sql

4. Pull the LLM

ollama serve
ollama pull mistral:7b-instruct

5. Ingest knowledge base PDFs

python ingest_pdfs.py

6. Start all services

# Terminal 1 — ML service (port 8001)
cd ml_service && python -m uvicorn ml_main:app --reload --port 8001 --host 0.0.0.0

# Terminal 2 — Backend (port 8002)
cd backend && python -m uvicorn backend_main:app --reload --port 8002 --host 0.0.0.0

# Terminal 3 — Flutter
cd aoun_app && flutter run

---

Project Structure

aoun/
├── backend/
│   ├── backend_main.py          # FastAPI entry point
│   ├── models.py                # SQLAlchemy ORM models
│   ├── schemas.py               # Pydantic request/response schemas
│   ├── database.py              # DB connection and session
│   ├── chatbot_routes.py        # Chatbot endpoints + safety guardrails
│   ├── aoun_rag.py              # Feedback-weighted RAG pipeline
│   ├── knowledge_base.py        # Curated FAQ entries
│   ├── ingest_pdfs.py           # PDF ingestion into ChromaDB
│   ├── smart_recommendations.py # Lifestyle recommendation engine
│   └── wellness_analyzer.py     # OpenCV facial analysis
├── ml_service/
│   └── ml_main.py               # Random Forest inference API
├── aoun_app/                    # Flutter app (13 screens)
├── database/
│   ├── aoun_schema.sql
│   └── chat_feedback_migration.sql
└── requirements.txt

---

Functional Requirements

27 of 28 functional requirements implemented.

Status	Count	Notes
✅ Implemented	27	All core features
❌ Deferred	1	FR7: Wearable integration (hardware dependency)

GP2 additions: Emergency screen · Facial wellness · RAG chatbot · ML feedback loop

---

Testing

19 of 20 black-box test cases passed. TC19 (Firebase push alert) is deferred — it requires a live production Firebase key and is not testable locally.

---

Known Limitations

• Authentication uses standard password hashing — salted Argon2/bcrypt and JWT migration planned
• Session-based auth only — JWT access tokens planned
• Voice input works on Chrome/Edge only (Web Speech API)
• Chatbot feedback quality improvement requires scale (hundreds of ratings)
• Facial wellness analyzer is a proof-of-concept — not clinically validated
• Firebase push notifications require a production server key not bundled with the repo

---

Status

Completed ✅

---

Disclaimer

Aoun is an independent academic research project. It is designed to support clinical decision-making and does not replace medical diagnosis, clinical judgment, or professional oncology care. All risk predictions and wellness assessments are advisory only.

---

License

This project is licensed under the MIT License — see LICENSE for details.

---

Dataset

The ML model was trained on the publicly available Cancer Patient Data Sets (cancer_patient_data_sets.csv) from Kaggle with SMOTE applied to make it more authentic.

Property	Value
Raw records	1,000 patients (real)
Columns	26 — patient ID, age (14–73), gender, 21 symptom/risk-factor features, target Level
Raw class split	High 36.5% · Medium 33.2% · Low 30.3%
Balancing	SMOTE applied to the training split only
After SMOTE	743,728 records — 594,982 train / 148,746 test (80/20 stratified)
Target variable	Risk level — High / Medium / Low

Real-world EHR data was not used due to patient-privacy law, IRB approval timelines, and the sensitivity of oncology records. The Kaggle dataset provides a scientifically structured training environment grounded in real clinical symptom patterns. The raw CSV is subject to Kaggle's terms and is not redistributed here — see the source on Kaggle to obtain it.