Autism Spectrum Disorder (ASD) Prediction using Machine Learning

Overview

This project focuses on building a machine learning model to predict whether an individual shows signs of Autism Spectrum Disorder (ASD) based on various behavioral and demographic attributes. The dataset consists of screening responses collected from adults through a questionnaire, with the target variable being Class/ASD (0: No, 1: Yes).

---

Tech Stack & Tools

• Language: Python
• Libraries: pandas, numpy, matplotlib, seaborn, scikit-learn, xgboost, imblearn, pickle
• Environment: Google Colab
• Version Control: Git & GitHub

---

Dataset

• Filename: train.csv
• The dataset contains features like age, gender, ethnicity, country of residence, and responses to 10 behavioral questions (A1_Score to A10_Score).
• The target column is Class/ASD (1 = ASD Positive, 0 = ASD Negative).

Key preprocessing steps:

• Dropped non-informative columns (ID, age_desc).
• Fixed inconsistent country names (Viet Nam → Vietnam, AmericanSamoa → United States, etc.).
• Replaced missing values in ethnicity and relation columns with "Others".
• Converted categorical features to numerical values using Label Encoding.
• Handled outliers in numerical columns (age, result) by replacing them with median values.
• Addressed class imbalance using SMOTE (Synthetic Minority Oversampling Technique).

---

Exploratory Data Analysis (EDA)

Performed detailed EDA to understand:

• Distribution of numerical features using histograms and box plots.
• Categorical feature counts using count plots.
• Correlation heatmap to detect relationships between features.

Insights from EDA:

• Some outliers in age and result columns.
• Noticeable class imbalance in the target variable (Class/ASD).
• No features showed strong multicollinearity.
• Ethnicity and relation columns required cleaning and normalization.

---

Data Preprocessing

Step	Description
Outlier Treatment	Replaced extreme values with median using IQR method
Label Encoding	Converted categorical columns to numeric
Class Balancing	Applied SMOTE to balance minority (ASD) class
Train-Test Split	80:20 ratio using train_test_split

---

Model Training

Three tree-based models were trained and evaluated:

1. Decision Tree Classifier
2. Random Forest Classifier
3. XGBoost Classifier

Each model was tuned using RandomizedSearchCV with 5-fold cross-validation.

---

Model Selection & Results

After tuning, the Random Forest Classifier achieved the best results.

Best Model:
RandomForestClassifier(bootstrap=False, max_depth=20, n_estimators=50, random_state=42)

📈 Model Performance

Metric	Score
Cross-Validation Accuracy (5-Fold)	0.93
Test Accuracy	0.82

Confusion Matrix

	Predicted No	Predicted Yes
Actual No	108	16
Actual Yes	13	23

Classification Report

Class	Precision	Recall	F1-Score
0 (No ASD)	0.89	0.87	0.88
1 (ASD)	0.59	0.64	0.61
Overall Accuracy	0.82

---

Key Insights

• SMOTE significantly improved recall for minority (ASD-positive) cases.
• The Random Forest model generalized well with strong test accuracy (0.82).
• The model demonstrates practical value for early ASD screening, especially for behavioral pattern-based datasets.

---

🧾 Files in Repository

File	Description
train.csv	Dataset used for model training
Autism_Prediction.ipynb	Complete Colab notebook with code
best_model.pkl	Serialized best-performing Random Forest model
encoders.pkl	Saved LabelEncoders for categorical variables
README.md	Project documentation

---

Author

Maitri Jain B.Tech in Computer Science (5th Semester) IIT Madras BS in Data Science (Foundation Level Completed)