HealthPlatform 🫀

Intelligent health analytics from your Apple Watch — powered by LSTM + XGBoost + GPT-4o

HealthPlatform is a full-stack AI health analytics application that ingests personal wearable data from Apple Watch, processes it through a multi-stage machine learning pipeline, and delivers personalised insights through an interactive Bento Grid dashboard and a GPT-4o conversational assistant.

---

✨ Features

Feature	Description
📱 Apple Health parser	Streaming XML parser handles export.xml files up to 800 MB without memory issues
🔄 Universal pipeline	Auto-detects columns from any health CSV via fuzzy matching (SequenceMatcher ≥ 75%)
🤖 Hybrid ML model	Bidirectional LSTM + XGBoost ensemble — 88.4% accuracy on real Apple Watch data
🛡️ Data validation	4-layer validation system with Trust Score (0–1) blocks bad data before prediction
💬 GPT-4o assistant	Health context + mood journal injected into LLM for personalised conversational advice
📊 Bento Grid dashboard	Glassmorphism UI with expandable bar charts, Activity Rings, BMI scale, and insights
👤 User profile	Body metrics, fitness goals, allergies, emotional wellness journal
💡 Health Tips	15 personalised tips across 5 categories with completion tracking
🔐 JWT auth	Secure login with bcrypt passwords and 7-day token sessions

---

🖥️ Screenshots

Dashboard · Import · Profile · Health Tips

---

🏗️ Architecture

HealthPlatform/
├── app/
│   ├── api/
│   │   └── routes/
│   │       ├── upload.py       
│   │       ├── predict.py      
│   │       ├── analysis.py    
│   │       ├── chat.py        
│   │       └── auth.py        
│   ├── core/
│   │   ├── parser.py           
│   │   ├── universal_data_pipeline.py 
│   │   └── data_validator.py  
│   ├── models/
│   │   ├── lstm_model.py      
│   │   ├── xgb_model.py      
│   │   └── hybrid_model.py    
│   └── services/
│       ├── auth_service.py    
│       └── gpt_service.py      
├── frontend/
│   └── src/
│       └── App.jsx            
├── data/
│   └── parsed_health_120days.csv 
└── run.py                    

---

🔄 Data Pipeline

Apple Health XML / CSV
        │
        ▼
┌───────────────┐
│  1. Parse     │  Streaming iterparse — constant ~50 MB memory for 800 MB files
└──────┬────────┘
       │
       ▼
┌───────────────┐
│  2. Validate  │  4-layer check → Trust Score 0.0–1.0
│               │  Types · Ranges · Semantics · Identity
└──────┬────────┘
       │  trust < 0.45 → BLOCKED
       ▼
┌───────────────┐
│  3. Pipeline  │  Schema detect → Column map (fuzzy) → Feature engineer
└──────┬────────┘
       │
       ▼
┌───────────────────────────────────┐
│  17-dimensional feature vector    │
│  HR · Steps · Sleep · Energy ·    │
│  BMI · fatigue_index · ...        │
└──────┬────────────────────────────┘
       │
       ├──────────────────┐
       ▼                  ▼
┌────────────┐    ┌────────────────┐
│  XGBoost   │    │  LSTM (30h     │
│  snapshot  │    │  window)       │
│  84.7%     │    │  86.2%         │
└──────┬─────┘    └──────┬─────────┘
       │                  │
       └────────┬─────────┘
                ▼
       ┌────────────────┐
       │  Hybrid 70/30  │  88.4% accuracy · F1 = 0.89
       └────────┬───────┘
                │
                ▼
       Dashboard + GPT-4o Chat

---

🧠 Multimodal Data Analysis

HealthPlatform fuses 6 data modalities into a unified 17-dimensional feature vector:

Modality	Signals	Role
Physiological	HeartRate, HRV, BloodOxygen	Primary LSTM temporal input
Activity	Steps, ActiveEnergy, Distance	XGBoost top features
Sleep	Awake / Light / REM / Deep	Encoded as sleep_stage + fatigue_index
Body composition	Weight, Height, BMI	Forward-filled, personalises advice
Workout events	Type, Duration, Energy	Encoded as workout_encoded, is_workout
User-reported	Mood emoji, journal, goal	Injected into GPT-4o system prompt

The last modality (qualitative) is not fed into the ML model — instead it enables language-model-level multimodal reasoning: physiological data + mood state → personalised natural language advice.

---

🛡️ Data Validation — Trust Score

Every uploaded file passes through 4 validation layers before the model runs:

Layer 1 — Type validation     (weight 20%)
  • All numeric columns parseable
  • datetime column valid

Layer 2 — Range validation     (weight 30%)
  • HeartRate: 28–220 bpm
  • BloodOxygen: 70–100 %
  • Steps: 0–80,000
  • Weight: 20–300 kg
  • ... 10 more physiological bounds

Layer 3 — Semantic validation  (weight 35%)
  • Minimum 24 rows (hourly records)
  • At least 6 distinct hours
  • HR standard deviation > 2 bpm
  • No impossible HR jumps > 60 bpm

Layer 4 — Identity check       (weight 15%)
  • Rejects research datasets (patient_id, os_months, fiber, collagen…)
  • Rejects multi-user data (> 5 unique IDs)

Trust Score = weighted average of all layers (0.0 – 1.0)
  ≥ 0.80  →  high    — full confidence predictions
  ≥ 0.60  →  medium  — predictions with warnings
  ≥ 0.45  →  low     — predictions with capped confidence
  < 0.45  →  BLOCKED — model does not run

---

📊 Model Performance

Dataset	Model	Accuracy	F1 Macro
Synthetic (15,000 rows)	Hybrid	99.8%	0.99
Weka Apple Watch (24 people)	LSTM	86.2%	0.87
Weka Apple Watch (24 people)	XGB	84.7%	0.85
Weka Apple Watch (24 people)	Hybrid	88.4%	0.89
Stress test (15% missing data)	Hybrid	43.9%	0.43

---

🚀 Getting Started

Prerequisites

• Python 3.10+
• Node.js 18+
• OpenAI API key (for chat feature)

Installation

# 1. Clone
git clone https://github.com/yourusername/HealthPlatform.git
cd HealthPlatform

# 2. Backend
pip install -r requirements.txt

# 3. Environment
cp .env.example .env
# Edit .env and add your OpenAI API key

# 4. Frontend
cd frontend
npm install
cd ..

Running

# Terminal 1 — Backend (port 8000)
python run.py

# Terminal 2 — Frontend (port 5173)
cd frontend && npm run dev

Open http://localhost:5173

Environment variables

SECRET_KEY=your-secret-key-here
OPENAI_API_KEY=sk-...
DATABASE_URL=sqlite:///./users.db
SEQUENCE_LENGTH=30

---

📡 API Endpoints

Method	Endpoint	Description
POST	/api/v1/auth/register	Create account → JWT token
POST	/api/v1/auth/login	Sign in → JWT token
PATCH	/api/v1/auth/update	Update name / email
POST	/api/v1/auth/change-password	Change password (bcrypt)
POST	/api/v1/upload	Upload health file → validate → parse
GET	/api/v1/analysis/{user_id}	Health stats + chart data
POST	/api/v1/predict/{user_id}	Run LSTM + XGB → activity prediction
POST	/api/v1/chat	GPT-4o health assistant
GET	/health	Server health check

Interactive docs: http://localhost:8000/docs

---

🧪 Supported File Formats

Format	Parser	Contents
export.xml	Streaming iterparse (800 MB+)	All Apple Health metrics
export_cda.xml	Regex chunked (multi-root XML)	HR, Weight, Height (HL7/CDA)
*.csv	pandas + fuzzy column mapping	Any smartwatch export

The universal pipeline accepts columns named HeartRate, heart_rate, Heart Rate (BPM), HR — fuzzy matching handles any naming convention with ≥ 75% similarity.

---

🛠️ Tech Stack

Layer	Technology
ML — Sequential	TensorFlow / Keras — Bidirectional LSTM
ML — Tabular	XGBoost
ML — Preprocessing	scikit-learn StandardScaler
Backend	FastAPI + Uvicorn
Auth	python-jose (JWT) + bcrypt
Database	SQLAlchemy + SQLite
AI Chat	OpenAI API — GPT-4o
Frontend	React 18 + Vite
Styling	Glassmorphism + Bento Grid + Playfair Display / Nunito
XML Parsing	xml.etree.ElementTree iterparse
Data	pandas + numpy

---

📁 Data

The repository includes data/parsed_health_120days.csv — 2,802 real hourly records from Apple Watch (December 2025 – April 2026) used for model fine-tuning and evaluation.

⚠️ This file contains anonymised personal health data. Do not use it for purposes other than model evaluation.

---

🗺️ Roadmap

• Fine-tune models on individual user data for personalisation
• Data Mining module — clustering, anomaly detection, weekly patterns
• Weekly trend analysis with longitudinal LSTM window
• Docker + cloud deployment (AWS / GCP) with PostgreSQL
• Mobile PWA with push notifications for daily health tips
• Sleep stage detailed breakdown and sleep debt tracker

---

📄 License

MIT License — see LICENSE for details.

---

👤 Author

Nataliia Nechyporenko

• GitHub: @Ffightik
• Email: [@nechyporenkonatalia31@gmail.com]

---

References

[1]-Smartwatch Health Data (Uncleaned), Mohammed Arfath R,2025. URL: https://www.kaggle.com/datasets/mohammedarfathr/smartwatch-health-data�uncleaned

[2] - Fitbit Fitness Tracker Dataset, Nneka Ekwemuk, 2023. URL: https://www.kaggle.com/datasets/nnekaekwemuka/fitbit-fitness-tracker-dataset

[3] - - iPhone User Guide, Download health records on iPhone, Apple Inc.,2026.URL: https://support.apple.com/en-om/guide/iphone/iphc30019594/ios

[4] - iPhone User Guide, Share your data in Health on iPhone, Apple Inc., 2026.URL:https://support.apple.com/en-om/guide/iphone/iph5ede58c3d/ios

---

_{Built with ❤️ and a lot of Apple Watch data}