added_lstm_model_alert_system

[pasanweerawardana]

📌 Summary

This PR introduces two major contributions to the Guardian Monitor AI module:

A robust data generation framework for synthetic patient monitoring.

A complete model formation pipeline covering data preparation, feature engineering, exploratory analysis, baseline models, and an advanced LSTM sequence model.

Together, these form a reproducible end-to-end workflow for generating, preparing, and modeling patient health and risk alert data.

🔹 Key Additions

1. Data Generation (Data_generation_code.ipynb)

Generates realistic multi-day patient datasets with 10-minute monitoring intervals.

Includes vitals (HR, SpO₂, BP, temperature, resp. rate), wearable signals (steps, accelerometer), and contextual features (activity level, meals, emotions).

Built-in rule-based logic for risk_alert labels, capturing anomalies (e.g., high HR, low SpO₂, abnormal meal/emotion effects).

Produces modular CSV batches for scalable experimentation.

2. Model Formation (Model_fomation_with_all_steps.ipynb)

EDA: Distribution checks, correlation analysis, and anomaly detection.

Feature Engineering: Rolling statistics (5-min, 60-min), HRV, step variability, and meal effects.

Data Prep: Scaling, encoding, and patient-level train/test split to avoid leakage.

Baseline Models: Logistic Regression and Random Forest with feature importance analysis.

Advanced Model: LSTM sequence model trained on 10-minute windows (~2 hours context).

Achieved ROC-AUC = 1.0 and PR-AUC ≈ 0.9999 on test data.

Very low false positives/negatives (≈ 4 misclassifications out of ~2000 samples).

Visualization: Training curves (loss, AUC, precision/recall), ROC/PR plots, and confusion matrix heatmap.

🔹 Why This Matters

Establishes a realistic, reproducible dataset for training/testing.

Provides a baseline-to-advanced modeling pipeline for Guardian Alert System development.

Lays the foundation for future experimentation (e.g., longer sequences, CNN-LSTM hybrids, real-time API integration).

✅ Next Steps

Extend generation to larger patient cohorts and longer monitoring periods.

Test alternative architectures (GRU, CNN-LSTM, Transformer).

Integrate trained models into the Guardian Alert System API for real-time monitoring.