Mental Health Prediction Using Deep Learning

Project Overview

This comprehensive project implements multiple machine learning and deep learning approaches to predict mental health status based on technology usage patterns and lifestyle factors. The project explores various models including traditional ML algorithms, ensemble methods, and advanced neural networks to classify mental health into four categories: Excellent, Good, Fair, and Poor.

Dataset

Mental Health and Technology Usage 2024 dataset contains behavioral and lifestyle data:

Total samples: ~10,000 records
Features: Technology usage, lifestyle, and demographic factors
Target: Mental Health Status (4 classes)
Format: CSV with mixed numerical and categorical features

Feature Categories:

Technology Usage

Technology_Usage_Hours
Social_Media_Usage_Hours
Gaming_Hours
Screen_Time_Hours

Health & Lifestyle

Sleep_Hours
Physical_Activity_Hours
Stress_Level (Low/Medium/High)

Demographics & Support

Age, Gender
Support_Systems_Access (Yes/No)
Work_Environment_Impact (Positive/Neutral/Negative)
Online_Support_Usage (Yes/No)

Target Variable

Mental_Health_Status: Excellent(0), Good(1), Fair(2), Poor(3)

Model Architectures

1. Deep Neural Networks

Standard Neural Network

NeuralNet:
├── Linear(input_size → 128) + BatchNorm + ReLU + Dropout(0.3)
├── Linear(128 → 64) + BatchNorm + ReLU + Dropout(0.3)
├── Linear(64 → 32) + BatchNorm + ReLU + Dropout(0.3)
└── Linear(32 → 4)  # 4 mental health classes

Advanced Neural Network

AdvancedNN:
├── Linear(input_size → 384) + BatchNorm + ReLU
├── Linear(384 → 256) + BatchNorm + ReLU + Dropout(0.3)
├── Linear(256 → 128) + BatchNorm + ReLU + Dropout(0.3)
├── Linear(128 → 64) + BatchNorm + ReLU + Dropout(0.3)
├── Linear(64 → 32) + BatchNorm + ReLU
└── Linear(32 → 4)

2. Traditional ML Models

Random Forest: Ensemble of decision trees
Gradient Boosting: Sequential weak learner improvement
XGBoost: Optimized gradient boosting with GPU support
LightGBM: Fast gradient boosting framework
CatBoost: Categorical feature handling
SVM: Support Vector Machine with RBF kernel
Logistic Regression: Multinomial classification
KNN: K-Nearest Neighbors
Extra Trees: Extremely randomized trees

3. Ensemble Methods

Voting Classifier: Soft voting across multiple models
Feature Extraction + ML: Deep features + traditional ML
Advanced Ensemble: Best performing models combination

Implementation Details

Data Preprocessing Pipeline

1. Missing Value Handling

# Numerical features: Median imputation
# Categorical features: Mode imputation
# Advanced: Domain-specific imputation strategies

2. Feature Encoding

# Label Encoding: Categorical variables
# One-Hot Encoding: Gender, Work Environment Impact
# Ordinal Encoding: Stress Level (Low=0, Medium=1, High=2)
# Binary Encoding: Support access variables

3. Feature Scaling

# StandardScaler: Mean=0, Std=1 normalization
# MinMaxScaler: 0-1 range normalization
# Applied to: Age, usage hours, activity hours

4. Feature Engineering

# Derived Features:
├── Total_Screen_Time = Tech + Social + Gaming hours
├── Screen_to_Sleep_Ratio = Screen_Time / Sleep_Hours
├── Activity_to_Screen_Ratio = Physical / Screen_Time
├── Health_Balance = (Sleep + Activity) / (Total_Tech + 1)
├── Stress_WorkImpact_Interaction = Stress × Work_Impact
└── Age_Category = Binned age groups

Training & Evaluation

Training Configuration

Train/Test Split: 80/20 stratified split
Cross-Validation: 5-fold CV for model selection
Batch Size: 64 (Neural Networks)
Epochs: 100-1000 with early stopping
GPU Acceleration: CUDA support for compatible models

Hyperparameter Optimization

# Grid Search CV for traditional ML
# Optuna optimization for XGBoost
# Manual tuning for neural networks
# Ensemble weight optimization

Performance Metrics

Accuracy: Overall classification accuracy
F1-Score: Weighted F1 for class imbalance
Classification Report: Precision, recall per class
Confusion Matrix: Detailed error analysis

Key Features

1. Comprehensive Data Analysis

# Statistical analysis:
├── Correlation analysis
├── Chi-square tests for categorical variables
├── Feature importance ranking
├── Distribution analysis by mental health status
└── Visualization suite (heatmaps, bar plots, box plots)

2. Advanced Feature Selection

# Multiple selection strategies:
├── SelectKBest (f_classif)
├── Recursive Feature Elimination (RFE)
├── Random Forest feature importance
└── Correlation-based filtering

3. Model Comparison Framework

# Systematic evaluation:
├── Multiple algorithm comparison
├── Hyperparameter optimization
├── Cross-validation results
├── Performance visualization
└── Best model selection

4. Production-Ready Pipeline

# Complete workflow:
├── Data validation and cleaning
├── Automated preprocessing
├── Model training and selection
├── Prediction generation
└── Results interpretation

Project Structure

Titanic---Machine-Learning-from-Disaster/
├── Titanic---Machine-Learning-from-Disaster.ipynb
├── README.md
├── .gitignore
└── README.md

Usage

Prerequisites

pip install torch tensorflow scikit-learn xgboost lightgbm catboost
pip install pandas numpy matplotlib seaborn optuna

Running the Analysis

1. Data Exploration

# Comprehensive EDA:
├── Missing value analysis
├── Class distribution visualization
├── Feature correlation analysis
├── Statistical significance testing
└── Relationship exploration

2. Model Training

# Multi-model approach:
├── Traditional ML with GridSearchCV
├── Neural networks with PyTorch/TensorFlow
├── Ensemble methods
└── Performance comparison

3. Evaluation & Selection

# Model assessment:
├── Cross-validation scores
├── Test set evaluation
├── Feature importance analysis
└── Best model identification

Technical Highlights

1. Advanced Neural Architecture

Batch Normalization: Accelerated training and stability
Dropout Regularization: Overfitting prevention
Progressive layer sizing: Optimal information flow
GPU Acceleration: CUDA optimization

2. Sophisticated Feature Engineering

# Complex feature interactions:
├── Ratio features: Screen/Sleep, Activity/Screen
├── Interaction terms: Stress × Work_Environment
├── Polynomial features: Screen_Time²
├── Categorical interactions: Age × Screen_Time
└── Domain-specific indicators: Excessive_Screen_Flag

3. Ensemble Intelligence

# Multiple ensemble strategies:
├── Simple voting: Equal weights
├── Weighted voting: Performance-based weights
├── Stacking: Meta-learner combination
└── Feature-based ensemble: Different feature subsets

4. Robust Evaluation Framework

# Comprehensive assessment:
├── Stratified cross-validation
├── Multiple random seeds
├── Statistical significance testing
└── Confidence interval estimation

Results & Insights

Model Performance Ranking:

Ensemble Methods: Highest accuracy and robustness
XGBoost/LightGBM: Strong individual performance
Neural Networks: Good with proper regularization
Random Forest: Reliable baseline performance

Key Insights:

Technology usage patterns strongly correlate with mental health
Sleep and physical activity are crucial protective factors
Work environment significantly impacts mental wellbeing
Support systems show measurable benefits

Feature Importance:

Screen_Time_Hours
Sleep_Hours
Stress_Level
Physical_Activity_Hours
Social_Media_Usage_Hours

Future Improvements

1. Advanced Architectures

Transformer models for sequence modeling
Graph Neural Networks for relationship modeling
Attention mechanisms for feature importance
Multi-modal learning for diverse data types

2. Enhanced Features

Temporal patterns: Time-series analysis
Social network data: Relationship patterns
Biomarker integration: Physiological data
External factors: Weather, economic indicators

3. Deployment Optimizations

Model compression: Pruning and quantization
Edge deployment: Mobile/IoT optimization
Real-time inference: Streaming predictions
Explainable AI: SHAP/LIME integration

Dependencies

Core Libraries:

PyTorch/TensorFlow: Deep learning frameworks
Scikit-learn: Traditional ML algorithms
XGBoost/LightGBM/CatBoost: Gradient boosting
Pandas/NumPy: Data manipulation
Matplotlib/Seaborn: Visualization

Specialized Tools:

Optuna: Hyperparameter optimization
SHAP: Model interpretability
Imbalanced-learn: Class imbalance handling
PyCaret: Automated ML pipeline

Key Achievements

Multi-algorithm comparison across 10+ models
Advanced feature engineering with domain expertise
Comprehensive ensemble methods implementation
GPU-accelerated training for efficiency
Production-ready pipeline with full automation
Statistical rigor with proper validation
Interpretable results with feature importance analysis

Learning Outcomes

This project demonstrates:

End-to-End ML Pipeline: Complete workflow implementation
Model Selection Strategy: Systematic algorithm comparison
Feature Engineering Expertise: Domain-driven feature creation
Ensemble Methods: Advanced combination techniques
Deep Learning: Modern neural network architectures
Statistical Analysis: Rigorous evaluation methodology

Name		Name	Last commit message	Last commit date
Latest commit History 4 Commits
.gitignore		.gitignore
README.md		README.md
Titanic---Machine-Learning-from-Disaster.ipynb		Titanic---Machine-Learning-from-Disaster.ipynb

Folders and files

Latest commit

History

Repository files navigation