HealthWatch at TechCorp

Project Overview

This project involves building a NLP and a CV model to generate diagnostics for patient using HealthWatch. The primary objective is to develop a model that can accurately classify illnesses according to their corresponding descriptions and to combine it with vizual input to confirm a diagnosis..

Data Source

The dataset used for this project was obtained from 9 different sources, all from huggingface:

• ashnaz/fine_tuned_symptoms
• ashnaz/symptoms_diagnose_doctors_data
• akhileshav8/json_symptom
• celikmus/symptom_text_to_disease_01
• venetis/symptom_text_to_disease_mk3
• celikmus/mayo_clinic_symptoms_and_diseases_v1
• ninaa510/diagnosis-text
• venetis/symptom_text_to_disease_mk4
• gretelai/symptom_to_diagnosis

After well processing the dataset and augmenting it, it contains a description and its classification, and I later added a label for each category so that the model can process it.

NLP

1. Model Selection:

   • Used the 'bert-base-uncased' model which is a transformer capable of understanding context.

2. Model Training:

   • Encapsulated training in a class and used some chosen configurations to train the model on a small training and testing set that is 10% the size of the original one.

3. Evaluation Metrics:

   • Accuracy was chosen as the primary metric because we had more than one class.

Computer Vision

1. Model Selection:

   • Used the 'MTCNN' model which is a CV model pre-trained on detecting faces.

2. Model Training:

   • Manual tuning on videos of the written detection methods until I got satisfactory results.

Key Insights

1. Strengths:

   • The model shows high precision in classification.

2. Weaknesses:

   • Has some bias towards hypertension.
   • CV sometimes wrongly detecs some signs because it needs to operate in perfect conditions.

Conclusion

The model developed in this project provides a reliable tool for diagnosing illnesses and for basic diagnostics.

Additional information

Adversities during preparation

• Finding a Pre-Trained CV model
• Google Colab GPU limit
• Training a transformer model

How to run the project

• Open the ipynb file in Google Colab
• Adjust the locations of the files in all the cells according to your drive
• Connect to your MongoDB database
• Run the cells in the order they're put in