Two-Stage Anomaly Detection in Personalized Healthcare using SVM & SVR

This repository contains the implementation of a two-stage anomaly detection model for wireless body area networks (WBANs), designed for personalized healthcare monitoring. The approach leverages Support Vector Machines (SVM) for anomaly classification and Support Vector Regression (SVR) for contextual anomaly detection. The method effectively distinguishes between sensor malfunctions and actual medical emergencies, ensuring high accuracy in anomaly detection.

📌 Key Features

Two-stage anomaly detection:

• SVM for point anomaly classification
• SVR for contextual anomaly detection

Feature Engineering:

• Average Heart Rate
• Heart Rate Difference
• ECG Rolling Mean & Standard Deviation
• MinMax Scaling for feature normalization
• Dynamic thresholding based on machine learning techniques

✔ Achieves up to 99.59% accuracy in anomaly detection
✔ Handles real-world physiological data for early medical intervention

📂 Dataset

The dataset consists of 72,000 physiological readings from 16 individuals (12 healthy participants and 4 patients), recorded thrice daily over five days. Data includes:

• Physiological attributes: Body Temperature, Heart Rate (from two sensors), SpO₂, ECG.
• Anomaly labels:
  • anomaly (overall anomaly classification)
  • first (point anomalies)
  • second (contextual anomalies)

📊 How It Works

1️⃣ Data Preprocessing

• Reads the modified_output.csv dataset.
• Assigns unique Person_IDs.
• Labels individuals as Healthy (Person_ID ≤ 12) or Patient (Person_ID > 12).
• Splits data into training (healthy individuals) and testing (patients + remaining healthy individuals).

2️⃣ Feature Engineering

Extracts relevant health metrics such as:

• Heart Rate Average & Difference
• ECG Rolling Mean & Standard Deviation

3️⃣ Anomaly Detection

The two-stage detection process includes:

• SVM Classification: Initially classifies anomalies using a radial basis function (RBF) kernel.
• SVR-Based Contextual Analysis: Builds regression models for physiological attributes and detects anomalies based on residual errors.

4️⃣ Evaluation

• Overall anomaly detection
• Point anomaly detection
• Contextual anomaly detection

📈 Performance Metrics

Model	Accuracy (%)
Proposed (SVM + SVR)	99.59%
DBSCAN + KMeans	60.01%
SVM (Standalone)	69.29%
Linear Regression	60.00%

🏛 Funding Agency

This research is supported by BITS Bio-CyTiH Foundation and the Department of Science and Technology (DST), Government of India. Their contributions have been instrumental in the development of this anomaly detection framework for wireless body area networks (WBANs) in personalized healthcare.