Disease Diagnosis with Deep Learning

A symptom-checker built around a Dense Neural Network implemented from scratch in MATLAB.

Project overview. This project addresses the critical issue of medical misdiagnosis by developing an AI-powered symptom checker. To ensure a deep understanding of the underlying mathematics, the Deep Neural Network (DNN) and backpropagation algorithms were implemented entirely from scratch in MATLAB, without relying on high-level deep-learning toolkits.

Introduction

Misdiagnosis is a pervasive issue in modern healthcare. Studies indicate that approximately 11% of medical problems result in a misdiagnosis, leading to nearly 795,000 instances of permanent disability or death annually in the United States alone.

The goal of this project was to create a decision-support system that can identify complex patterns between symptoms and diseases. By acting as a “second opinion,” such a system allows medical professionals (and patients) to explore a wider range of diagnostic possibilities, particularly for infrequent diseases that might otherwise be overlooked.

Methodology

1. Data preprocessing

The model was trained on a dataset of 41 diseases and 131 unique symptoms across 4920 datapoints. The raw data required significant preprocessing:

  • Mapping — diseases converted to index representations; symptoms encoded as numerical features.

  • One-hot encoding — labels (diseases) one-hot encoded. For disease index 2, the label vector becomes:

    \[y = [0, 1, 0, 0, \dots, 0]\]
  • Train/test split — randomized to ensure generalization.

Dataset dataframe and disease list

Figure 1: Snapshot of the raw dataframe (left) and the list of possible diseases (right).

2. Deep learning from scratch

To master fundamentals, I avoided standard libraries (PyTorch, TensorFlow) for the core model logic. Instead I implemented the network architecture manually in MATLAB — coding the matrix operations for the forward pass and deriving the gradients for backpropagation by hand.

The network is a Dense Neural Network with:

  • Input layer — 131 nodes (one per symptom)
  • Hidden layers — fully connected with learnable weights and biases
  • Output layer — 41 nodes (one per disease) with Softmax activation

Dense neural network architecture

Figure 2: The properties of the custom Neural Network class implemented from scratch.

3. Interactive application

The trained model was deployed into a MATLAB App Designer interface that lets users select symptoms from a checklist and receive a real-time diagnosis.

Disease diagnoser UI

Figure 3: The UI allowing users to select symptoms.

Results and discussion

Performance

The model achieved near-perfect accuracy on the testing set within the scope of the provided dataset. However, accuracy in a controlled environment can be misleading when applied to real-world medical diagnostics.

The Softmax-vs-Sigmoid dilemma

A significant finding was the limitation of using Softmax for this task. Softmax enforces that the sum of probabilities across all classes equals 1:

\[\sigma(z)_i = \frac{e^{z_i}}{\sum_{j=1}^K e^{z_j}}\]

This setup treats the problem as multi-class classification (the patient has exactly one disease out of 41) rather than multi-label classification (the patient could have comorbidities, or none of the above). To mitigate, I implemented a 70% confidence threshold: predictions below it are discarded.

Sensitivity issues

Due to thresholding and dataset structure, the model sometimes required an excessive number of symptoms to trigger a diagnosis.

Common cold diagnosis example

Figure 4: Diagnosing the common cold required selecting over 10 specific symptoms to cross the confidence threshold.

Key takeaways

  1. Fundamentals matter. Implementing backpropagation and layer logic from scratch provided deeper intuition than using high-level APIs.
  2. Architecture choice. For medical diagnosis, a multi-label approach (Sigmoid per node) is superior to multi-class (Softmax) — it allows independent probabilities per disease.
  3. Data limitations. Real-world deployment would require a vastly larger, more diverse dataset and a probabilistic framework that accounts for rare conditions.