Update README.md

README.md

@@ -1,19 +1,172 @@
 ---
-title: InjuryDetection
-emoji: 🚀
+title: 🏀 InjuryDetection
+emoji: 🏀
 colorFrom: red
 colorTo: red
 sdk: docker
 app_port: 8501
 tags:
-- streamlit
+  - streamlit
+  - transformers
+  - nlp
+  - pytorch
+  - nba
+  - healthcare
+  - sports
 pinned: false
-short_description: Streamlit template space
+description: Predict NBA injury type and duration using a fine-tuned DistilBERT + structured features.
+
+---
+
+# 🏀 Injury Detection & Recovery Duration Estimator
+
+A powerful **Streamlit app** powered by a fine-tuned **DistilBERT** transformer that predicts:
+
+- 🔍 **Injury Type** (e.g., bone, muscle, joint, illness, concussion)
+- ⏳ **Recovery Duration**:  
+  - `short` (< 7 days)  
+  - `medium` (7–45 days)  
+  - `long` (> 45 days)
+  
+---
+
+## 🚀 Why This Project?
+
+NBA injuries are unpredictable, and doctors often rely on vague reports or historical intuition. I wanted to go beyond zero-shot text classification by:
+
+- Cleaning and normalizing raw injury logs (1950–2022)
+- Labeling 8K+ examples by type and duration
+- Incrementally testing feature combinations
+- Analyzing attention weights and feature influence
+- Making predictions explainable and interactive via Streamlit
+
+---
+
+## 🧠 Model Highlights
+
+| Component         | Description                             |
+|------------------|-----------------------------------------|
+| Model            | `distilbert-base-uncased`               |
+| Task             | Dual classification                     |
+| Inputs           | Text + prior injuries, position, type ID|
+| Output Heads     | `label_type_id` and `label_duration_id` |
+| Loss             | Weighted cross-entropy (multi-task)     |
+| Extras           | Attention score input for interpretability|
+
 ---
 
-# Welcome to Streamlit!
+## 💡 Features
+
+- 📝 **Free-text input** for injury reports
+- 🧱 **Structured context**: prior injuries, position, injury category
+- 🎯 **Fine-tuned BERT** with dual-head classification
+- 🔍 **Live predictions** with **confidence scores**
+- 📊 **Built-in feature importance + attention hooks**
+- 🌐 **Streamlit UI** with dropdowns, metrics, and sample cases
+
+---
+
+## 🗂️ File Structure
+```text
+project/
+├── src/
+│   ├── streamlit_app.py        # main Streamlit UI app
+│   ├── predict_utils.py        # logic for prediction function
+│   └── final_injury_model.pt   # fine-tuned dual-head transformer model (DistilBERT)
+├── requirements.txt            # all dependencies (Torch, HF Transformers, Streamlit, etc.)
+├── modeling_notebooks/         # experimentation, feature importance, attention
+├── cleaned_data/               # cleaned dataset used for training
+├── raw_data/                   # original source CSVs
+└── README.md                   # this file
+```
+
+---
+
+## ⚙️ Setup
+
+1. **Install dependencies**
+
+```bash
+pip install -r requirements.txt
+```
+2. **Run App**
+```bash
+streamlit run app.py
+```
+---
+
+## ⚙️ Model Overview
+
+| **Component**          | **Description**                                           |
+| ---------------------- | --------------------------------------------------------- |
+| **Base Model**         | `distilbert-base-uncased`                                 |
+| **Input**              | Injury description text + structured inputs               |
+| **Structured Inputs**  | Prior injuries, position ID, injury type ID               |
+| **Output Heads**       | `label_type_id`, `label_duration_id`                      |
+| **Optimization**       | Multi-task cross-entropy loss                             |
+| **Performance Boosts** | Attention score injection, feature dropout, class weights |
+
+---
+
+## 📈 Results Summary
+
+| Metric              | Value   | Description                                                   |
+|---------------------|---------|---------------------------------------------------------------|
+| **Type Accuracy**   | 99.5%   | Nearly perfect prediction for general injury type            |
+| **Duration Accuracy** | 65.0% | More challenging task due to overlap in medium/long classes  |
+| **Macro F1 (Duration)** | ~0.64 | Balanced F1 across duration classes                          |
+| **Most Confused Pair** | `long` vs `medium` | Long and medium often overlapped in symptoms and context |
+| **Evaluation Set Size** | 200 samples | Held-out test subset from full dataset                     |
+| **Unknown Labels Removed** | ✅ | Improved class balance and duration accuracy                |
+| **Feature Importance** | moderate | Prior injuries most useful, attention least influential     |
+
+---
+
+## What I learned
+
+Structured + Textual fusion drastically improves performance over text-only
+
+Class balancing + weighted loss helped fix bias toward short injuries
+
+Adding features like injury type, position, prior injuries sequentially allowed modular experimentation
+
+Attention scores showed low influence, but modeling it validated model interpretability
+
+---
+
+## 💬 Sample Predictions
+
+"torn ACL expected to miss rest of season"  
+→ Type: **joint** (99%), Duration: **long** (91%)
+
+"minor hamstring strain"  
+→ Type: **muscle** (96%), Duration: **short** (87%)
+
+"fractured tibia, placed on IL"  
+→ Type: **bone** (98%), Duration: **long** (89%)
+
+---
+
+## Hugging Face Transformers & Datasets
+
+PyTorch, Scikit-learn, and Streamlit for all underlying tech
+
+---
+
+## Datasets
+- Player_Info: [NBA Players stats since 1950](https://www.kaggle.com/datasets/drgilermo/nba-players-stats?select=player_data.csv)
+- Player_Info:[NBA Players data (1950 to 2022)](https://www.kaggle.com/datasets/blitzapurv/nba-players-data-1950-to-2021?select=player_data.csv)
+- Injury Info: [🏀 NBA Injury Stats (1951–2023)](https://www.kaggle.com/datasets/loganlauton/nba-injury-stats-1951-2023)
+
+---
+
+## 🧠 Try it Out Now
+Enter a short injury description, and get:
+
+✅ Predicted injury type
+
+⏳ Estimated recovery time
 
-Edit `/src/streamlit_app.py` to customize this app to your heart's desire. :heart:
+📈 Confidence scores
 
-If you have any questions, checkout our [documentation](https://docs.streamlit.io) and [community
-forums](https://discuss.streamlit.io).
+✨ Fork this space and customize it to your league, team, or medical use case.