NoShowAI — Clinical Appointment No-Show Prediction

An AI-powered healthcare analytics system that predicts patient appointment no-shows using machine learning and surfaces risk insights through an interactive Streamlit dashboard.

---

Problem

Missed clinical appointments lead to:

• Resource wastage and idle clinic capacity
• Increased patient waiting times
• Revenue loss for healthcare providers
• Disrupted continuity of patient care

Solution

NoShowAI estimates the probability that an appointment will result in a no-show using historical scheduling data. Clinics can identify high-risk appointments early and take preventive actions — sending reminders, rescheduling, or initiating social-work outreach.

---

Key Features

• Data Preprocessing & Feature Engineering — cleaned appointment records with lead-time, demographic encoding, and reminder flags
• Multi-Model ML Training — Logistic Regression, Decision Tree, and Random Forest trained with class balancing (SMOTE or class_weight='balanced')
• Best-Model Auto-Selection — model with highest F1 (No-Show) score is saved as model/best_model.pkl
• Risk Probability Estimation — per-appointment probability scores with configurable high-risk threshold
• Interactive Streamlit Dashboard — 3-page workflow: Upload → Risk Dashboard → Analytics
• Model Performance Comparison — Accuracy, F1, and R² metrics surfaced per model in the UI

---

Tech Stack

Layer	Libraries
Data	pandas, numpy
ML	scikit-learn, imbalanced-learn (optional, for SMOTE)
Serialisation	joblib
Visualisation	matplotlib, seaborn, Streamlit native charts
Dashboard	streamlit

Dependencies: requirement.txt

---

Architecture

Raw Data (CSV)
    │
    ▼
Preprocessing / Feature Engineering  ←── notebooks/
    │
    ▼
Model Training  (src/train.py)
    ├── Logistic Regression   →  model/logistic_regression.pkl
    ├── Decision Tree         →  model/decision_tree.pkl
    ├── Random Forest         →  model/random_forest.pkl
    └── Best Model (by F1)    →  model/best_model.pkl
    │
    ▼
Streamlit Dashboard  (ui/app.py)
    ├── Page 1: Upload Dataset
    ├── Page 2: Risk Dashboard  (predictions + colour-coded risk table)
    └── Page 3: Analytics       (distributions, model metrics, feature importance)

---

Project Structure

NoShowAI/
├── data/                   # Raw and processed datasets
├── notebooks/              # EDA, preprocessing, and modelling notebooks
├── src/
│   └── train.py            # Full training pipeline (3 models + best-model export)
├── model/                  # Saved model artefacts (.pkl files)
│   ├── best_model.pkl
│   ├── logistic_regression.pkl
│   ├── decision_tree.pkl
│   ├── random_forest.pkl
│   └── scaler.pkl
├── ui/
│   └── app.py              # Streamlit dashboard (Upload → Risk → Analytics)
├── docs/                   # Architecture notes and documentation
├── requirement.txt         # Python dependencies
└── README.md

---

Setup

1. Create and activate a virtual environment

python -m venv .venv
source .venv/bin/activate

2. Install dependencies

pip install -r requirement.txt

Optionally install imbalanced-learn for SMOTE oversampling during training:

pip install imbalanced-learn

3. Train the models

Place your cleaned dataset at data/noshow_cleaned (1).csv, then run:

python src/train.py

This will train all three models, print a comparison table, and save the best-performing model as model/best_model.pkl.

4. Launch the dashboard

streamlit run ui/app.py

Open http://localhost:8501 in your browser. Follow the 3-step in-app workflow:

1. Upload Dataset — upload the preprocessed CSV
2. Risk Dashboard — click Predict No-Show Risk to score all appointments; high-risk rows are highlighted in red
3. Analytics — view model performance (Accuracy / F1 / R²), no-show distribution, probability histogram, and feature importance

---

Dashboard Workflow

Step	Page	Action
1	Upload Dataset	Upload a preprocessed appointments CSV
2	Risk Dashboard	Run predictions; adjust high-risk threshold (default 0.70)
3	Analytics	Review per-model metrics, charts, and feature importances

---

Team Roles

Member	Responsibility
Harsh	Data preprocessing & feature engineering
Manmath	ML modelling & evaluation
Ansh	Analysis, visualisation & insights
Yash	UI development & system integration

---

Future Scope

• Automated intervention recommendations (reminders / rescheduling suggestions)
• Explainability layer (SHAP values for per-patient rationale)
• Deployment for clinic workflow integration (Docker / cloud)
• Real-time appointment ingestion via API

---

License

Developed for academic and research purposes.