XGBoost · SHAP Explainability · Apple-Health Inspired UI

1. Problem Statement
2. Why Autism Detection Matters
3. Dataset Description
4. Project Pipeline (ML Workflow)
5. System Architecture
6. Model Performance & Comparison
7. Explainability (SHAP)
8. Streamlit App UI Preview
9. Project Features
10. How to Run Locally
11. Folder Structure
12. Clinical Disclaimer
13. Limitations
14. Contact

Millions of individuals remain undiagnosed or diagnosed late for Autism Spectrum Disorder (ASD) due to:

• Limited access to clinical specialists
• Long waiting periods for assessments
• Lack of awareness or hesitation to seek help
• Resource constraints in low-income regions

The challenge:
How can we build a fast, transparent, accessible tool to support early ASD screening — without replacing clinical evaluation?

This project answers that by building a responsible, explainable ML-based support system using the AQ-10 screening questionnaire.

• ASD affects approximately 1 in 100 people globally
• Early identification improves communication, social, and learning outcomes
• Screening gaps exist in rural and low-resource healthcare systems
• Digital tools can help triage cases early
• Machine learning can support clinicians — not replace them

This project demonstrates how XAI + ML can enhance early screening accessibility.

Source:
UCI / Kaggle — Autism Screening Adults & Children Dataset

Dataset Type:
Questionnaire-based binary classification (ASD vs Non-ASD)

Contents:

• AQ-10 questionnaire (10 binary questions)
• Demographics:
  age, gender, ethnicity, country of residence
• Medical factors: jaundice at birth
• Social factors: relation (parent/self), used autism app before
• Target: class_asd

Size: ~700 samples
Features: 19
Label distribution: Balanced enough for supervised learning

📌 Note:
The dataset is small and highly separable because AQ-10 questions are directly diagnostic.
This explains the unusually high performance of ML models.

flowchart LR
    subgraph PREP[Data Preparation]
        A1[Raw Data]
        A2[Cleaning]
        A3[Feature Engineering]
        A4[Encoding]
        A5[Train-Test Split]
        A1 --> A2 --> A3 --> A4 --> A5
    end

    subgraph MODEL[Modeling]
        B1[XGBoost Training]
        B2[Evaluation Metrics]
        B3[SHAP Explainability]
        B1 --> B2 --> B3
    end

    subgraph APP[Application Layer]
        C1[Streamlit App]
        C2[PDF Report Generator]
    end

    A5 --> B1
    B3 --> C1
    C1 --> C2

flowchart LR

    subgraph UI[Streamlit Frontend]
        A[User Inputs & Dashboard]
    end

    subgraph BE[Backend Model Layer]
        B[Preprocessing → XGBoost Model → SHAP Explainability]
    end

    subgraph AR[Artifacts & Reports]
        C[Models • Processed Data • PDF Reports]
    end

    A --> B --> C

• Best trade-off between accuracy, stability, and interpretability
• Works extremely well on small structured datasets
• Fully compatible with TreeSHAP for transparent explainability
• Fast, robust, and highly generalizable

This project uses SHAP (SHapley Additive Explanations) to provide transparent, interpretable insights into why the model predicts ASD Positive or Negative.

Shows how each feature contributed to an individual prediction.

Example:

• Negative SHAP → pushes toward ASD Negative
• Positive SHAP → pushes toward ASD Positive
• Larger magnitude → stronger influence

Follow the steps below to run the Autism Diagnostic Support System on your machine.

1. Clone the Repository

git clone https://github.com/aparnaworkspace/autism-diagnostic-support-tool
cd autism-diagnostic-support-tool

2. Create and Activate a Virtual Environment

python3 -m venv venv
source venv/bin/activate

(Windows users: venv\Scripts\activate)

3. Install Dependencies

pip install -r requirements.txt

4. Run the Streamlit Application

streamlit run app/streamlit_app.py

5. Regenerate SHAP Explainability Visuals

If you want fresh SHAP plots (bar, beeswarm, waterfall):

python notebooks/04_Model_Evaluation.py

6. Jupyter Notebook Workflow

To explore EDA or model training:

jupyter lab

A well-structured, production-style codebase:

autism-diagnostic-support-tool/
│
├── app/
│   └── streamlit_app.py
│
├── assets/
│   ├── home.png
│   ├── prediction.png
│   ├── pdf_report.png
│   ├── shap_bar.png
│   ├── shap_beeswarm.png
│   ├── shap_waterfall_sample_0.png
│   ├── confusion_matrix.png
│   ├── calibration_curve.png
│   └── roc_curve.png
│
├── data/
│   ├── raw/
│   │   ├── autism_screening.csv
│   │   └── Autism-Child-Data.csv
│   └── processed/
│       ├── autism_combined.csv
│       ├── X_train.csv
│       ├── X_test.csv
│       ├── y_train.csv
│       └── y_test.csv
│
├── models/
│   ├── best_model.pkl
│   ├── scaler.pkl
│   ├── label_encoders.pkl
│   └── shap_explainer_and_values.pkl
│
├── notebooks/
│   ├── 01_EDA.ipynb
│   ├── 02_Feature_Engineering.ipynb
│   ├── 03_Model_Training.ipynb
│   └── 04_Model_Evaluation.ipynb
│
├── reports/
│   ├── confusion_matrix.png
│   ├── roc_curve.png
│   ├── calibration_curve.png
│   └── *.pdf
│
├── src/
│   ├── preprocess.py
│   ├── train_model.py
│   ├── risk_scoring.py
│   └── explainability.py
│
├── docs/
│   ├── MODEL_CARD.md
│   ├── DATA_CARD.md
│   ├── MODEL_COMPARISON.md
│   ├── SYSTEM_ARCHITECTURE.md
│   └── ETHICS_CARD.md
│
├── requirements.txt
└── README.md

⚠️ This tool is NOT a diagnostic system.

It is an educational machine-learning project designed for:

research demonstration

explainability exploration (SHAP)

portfolio and skill showcasing

Autism Spectrum Disorder (ASD) diagnosis requires trained clinicians and involves:

behavioural observation

developmental history

structured clinical interviews

neuropsychological assessments

multi-disciplinary evaluation

genetics & neurological analysis

No machine-learning model, screening questionnaire, or digital tool can replace professional evaluation. This project should not be used for medical, clinical, or therapeutic decision-making.

Despite strong performance, the project has important limitations that recruiters and reviewers should know:

Dataset Limitations

Small dataset (≈700 samples)

Questionnaire-based (AQ-10) → inherently diagnostic

Limited feature variety (binary responses, demographics)

May contain cultural or demographic biases

Does not include real-world behavioural, video, audio, MRI, or genetic data

Model Limitations

High accuracy partly due to dataset separability

May not generalize to unseen populations or clinical settings

No temporal, behavioural, or contextual signals

Risk of overfitting due to small sample size

Application Limitations

UI is for demonstration only (not medically approved)

SHAP helps explain decisions but does not guarantee model fairness

PDF reports are educational summaries, not clinical documents

If you’d like to connect or discuss this project: 📧 [email protected]

🔗 LinkedIn:www.linkedin.com/in/aparnasajeevan1610