NeuroCEDT

# Cognitive Embedding Diagnostic Transformer (CEDT)


## 🧠 Project Overview

This repository implements the **Cognitive Embedding Diagnostic Transformer (CEDT)**, a deep learning architecture designed for **personalized diagnosis of psychiatric disorders** using **neuroimaging data**. It addresses limitations in traditional machine learning by integrating symbolic reasoning, probabilistic inference, and attention-based representation learning.

CEDT is accompanied by a novel calibration method, **Disorder-Aware Interpretive Calibration (DAIC)**, which enhances interpretability, uncertainty quantification, and clinical relevance.

---

## 🚀 Features

- 🧩 **Multimodal Symptom Embedding** with structured attention layers.- 🌐 **Latent Diagnostic Space** capturing diagnostic uncertainty.- 📊 **Multi-Hypothesis Output** with entropy-aware uncertainty estimation.- 📚 **Taxonomic Calibration** using domain knowledge from psychiatric ontologies.- 🧠 **Graph-based Comorbidity Regularization** for clinically aligned predictions.- 🛠️ Extensive ablation and benchmarking across 4 public datasets.

---

## 📁 Datasets

Experiments were conducted on:

- **ImageNet** (for transfer learning)- **Caltech-256** (object diversity)- **Oxford 102 Flowers** (fine-grained classification)- **Describable Textures Dataset (DTD)** (texture sensitivity)

Note: These datasets are used for architectural benchmarking. The clinical framework is designed for neuroimaging symptom data.

---

## 🏗️ Architecture

- **CEDT Core**: Attention-based encoder-decoder with multi-head attention for symptom-to-diagnosis mapping.- **Tok-KAN Blocks**: Tokenized knowledge-aware attention layers (see *Figure 1, Page 7*).- **Graphical Propagation Layer**: Incorporates co-diagnosis structure via Laplacian smoothing.- **DAIC Module**: Combines Auto-Fusion networks, taxonomic embeddings, and entropy gating for calibration (see *Figure 3, Page 11*).- **Prototype-Centered Refinement**: Aligns output with learned diagnostic prototypes (*Figure 4, Page 13*).

---

## 🧪 Training & Evaluation

- **Framework**: PyTorch- **Optimizer**: AdamW- **Learning Rate**: `1e-4` (cosine annealing)- **Batch Size**: 64- **Epochs**: 300- **Regularizations**: Label smoothing, dropout, contrastive loss- **Evaluation Metrics**: Accuracy, F1 Score, Recall, AUC

---

## 📈 Results Summary

| Dataset            | Accuracy | F1 Score | AUC     ||--------------------|----------|----------|---------|| ImageNet           | 83.90%   | 82.95%   | 88.62%  || Caltech-256        | 89.47%   | 88.74%   | 91.55%  || Oxford 102 Flowers | 93.66%   | 93.12%   | 95.45%  || DTD (Textures)     | 78.93%   | 78.01%   | 81.42%  |

See *Tables 1 and 2, Pages 16–17* for complete benchmarking.

---

## ⚙️ Installation

```bashgit clone https://github.com/your_username/cedt-framework.gitcd cedt-frameworkpip install -r requirements.txt