README.md · chufangao/CTO at main

metadata

annotations_creators:
  - other
language_creators:
  - other
language:
  - en
license: mit
multilinguality:
  - monolingual
size_categories:
  - 10K<n<100K
source_datasets:
  - original
task_categories:
  - text-classification
task_ids:
  - acceptability-classification
pretty_name: CTO (Clinical Trial Outcome Prediction Dataset)
config_names:
  - news
  - linkage
  - human_labels
  - pubmed_gpt
  - stocks_and_amendments
configs:
  - config_name: phase1_CTO_preds
    data_files:
      - split: test
        path: phase1_CTO_rf.csv
  - config_name: phase2_CTO_preds
    data_files:
      - split: test
        path: phase2_CTO_rf.csv
  - config_name: phase3_CTO_preds
    data_files:
      - split: test
        path: phase3_CTO_rf.csv
  - config_name: news
    data_files:
      - split: test
        path: news_lfs/*
  - config_name: linkage
    data_files:
      - split: test
        path: Merged_all_trial_linkage_outcome_df__FDA_updated/*
  - config_name: human_labels
    data_files:
      - split: test
        path: human_labels_2020_2024/*
  - config_name: pubmed_gpt
    data_files:
      - split: test
        path: pubmed_gpt_outcomes/*
  - config_name: stocks_and_amendments
    data_files:
      - split: test
        path: labels_and_tickers/*
tags:
  - clinicaltrials
  - healthcare

Dataset for predicting clinical trial outcomes in drug development. This dataset is part of the work presented in "Automatically Labeling Clinical Trial Outcomes: A Large-Scale Benchmark for Drug Development".

Website: https://chufangao.github.io/CTOD/

Paper: https://arxiv.org/abs/2406.10292

Code: https://github.com/chufangao/ctod

Descriptions:

human_labels contains the manually annotated subset. We follow the same rule-based termination of incomplete status and p-value < 0.05 as in the automated labeling step to remove the easy cases. The additional metadata is taken from studies.txt from CTTI.
linkage contains the weakly linked trials as predicted based on text similarity. E.g. if a similar trial is found in phase 3 from a trial in phase 2, then they are considered "linked". We use a reranking method after the initial text similarity retrieval to further refine relevance. "connected next phase" implies similar text similarity AND positive reranking. "weakly connected next phase" implies similar text similarity BUT negative reranking.
news contains the top 10 most similar (cosine text similarity) headlines to the trial, 0 being the most similar. "valid_sentiments" is a list of sentiments from a binary sentiment classifier, and "mode" is the most popular sentiment.
phase1_CTO_preds, phase2_CTO_preds, phase3_CTO_preds contain phase-specific CTO predictions. Note that each phase may have different columns due to the phase-specific thresholding of weak supervision sources. Some columns are duplicated (e.g. gpt1, gpt2) for dynamic programming purposes. Duplicated columns may be ignored in practice. The most relevant columns are ["nct_id", "pred", "pred_proba"], where "pred_proba" is the predicted probability of trial success.
pubmed_gpt contains linked GPT predictions on related trial publications
stocks_and_amendents contains scraped stock price Slope from a 5-day moving average. The Slope is calculated from: (trial completion date, trial completion date + 7 days). This file also contains the scraped number of amendments made to the trial on clinicaltrials.gov.

CTTI.zip is from: Aggregate Analysis of ClinicalTrials.gov (AACT) Database. Clinical Trials Transformation Initiative (CTTI). Available at: https://aact.ctti-clinicaltrials.org/ (Accessed: 10/2024).