Presentation

Welcome to the CS685FinalProject wiki!

1-1.5 minutes for introducing the problem and what is/was the technical challenge.

• Problem: cancer diagnosis from clinical notes, unstructured data (narrative only)
  • get statistic on unstructured data in clinical notes
• Challenge: difficult to get good performance from NLP methods
  • data scarcity, availability of de-identified clinical text
  • labeled data scarcity
  • why is cancer diagnosis important?
  • how is cancer diagnosed?
  • why do we want to auto classify text?
  • why is hard to read text
• propose to use Pretrained models and fine tune on smaller dataset
• clinical text is very long, how to identify important portions
• project started big, had to narrow down

1-1.5 baseline approaches.

• get baselines from other clinical text cancer diagnosis
• look at difficulties from other papers
• show BERT model performance
• explain BERT, explain difference to GPT models

Approximately 5 minutes (i.e., 50% time) for your technical contribution in this work.

• data
• preprocessed data
• baseline / low effort training

Anchit:

• hyperparameter tuning
• data truncation at end (diff chunks, look at data)
• look at training loss / validation loss plots, analyze way to get better results from training

2-3 minutes for sharing your findings and results.