Riskcovry-Hackathon

PROBLEM STATEMENT CHOSEN:

Valid Discharge Summary Prediction(Problem 2)

---

Working website on public gcp url: http://34.70.84.140:8000/ (If the link doesnt work please reach out to us)(jigyas15@gmail.com or nayak.amit.blr@gmail.com)

---

UPDATE => WORKS AS AN API

Note that the PDF should be sent as form-data with key set as 'file'

TECHSTACK

Django, HTML, CSS, JavaScript, Pytorch+FastAI, Google Cloud Platform (VM, Cloud Storage)

BASIC OVERVIEW

• Extraction of batched text from multipage PDF using Google's Vision API
• Implementing transfer learning on AWD-LSTM Text classification model
• Display results on an embedded Django Server running on a Google VM

DETAILS

• Trained on a handpicked minimal dataset with edges cases like health records and medical research papers. All stored on GCP bucket.
• ~89% accuracy attained on AWD-LSTM model after 3 epochs with a training set of ~50 PDF files

CODE OVERVIEW

• TextExtraction.ipynb: handpicked PDF dataset on Cloud Storage -> Vision API -> JSON on Cloud Storage
• Model.ipynb: JSON on Cloud Storage -> Training of AWD-LSTM -> exporting model
• scripts/testmodel.py: runs the fine-tuned model on django frontend
• bytes: django files

HAVE A LOOK

TEAM BYTES

• Jigya Shah
• Amitha Nayak