HLMCC/TCGA-PAAD

Dataset Card for TCGA-PAAD Clinical Data

Dataset Summary

The TCGA-PAAD (The Cancer Genome Atlas - Pancreatic Adenocarcinoma) clinical dataset contains clinical data related to pancreatic adenocarcinoma patients. This dataset is part of the broader TCGA project, aimed at providing comprehensive genomic and clinical data for various types of cancer. The clinical data includes information such as patient demographics, treatment history, survival data, and other clinical attributes relevant to the study of pancreatic cancer. This dataset is used primarily for research and analysis purposes, specifically for building machine learning models to predict outcomes such as survival, recurrence, or response to treatment based on clinical attributes.

Hackathon Goal

Develop a machine learning model to predict the survival rate of pancreatic adenocarcinoma (PAAD) patients using clinical data. The model will be evaluated using the C-Index (Concordance Index), a metric specifically suited for measuring the accuracy of survival models.

Background

Pancreatic cancer has one of the highest mortality rates, and accurately predicting patient survival based on clinical data can assist in treatment planning and patient care. The dataset contains various clinical variables, such as demographic information, treatment history, and survival time, which will be used as predictors in the model.

Task Details

• —Type of Model: Regression
• —Target Variable: Patient risk score or relative survival prediction (often derived from survival time and censoring status)
• —Metric for Evaluation: C-Index (Concordance Index), which measures the predictive accuracy of survival models. It evaluates the concordance between the predicted survival times and the actual survival times in pairs of patients.
• —Key Columns: days_to_death, days_to_last_follow_up, and vital_status for survival analysis.

[!NOTE] For more information on the C-Index calculation, please refer to the Lifelines documentation.

Why C-Index?

The C-Index is commonly used for survival analysis because it reflects how well the model can distinguish between pairs of patients, predicting which one will survive longer. It provides an advantage over typical regression metrics like Mean Squared Error (MSE) when handling censored data, which is common in clinical survival datasets.

Dataset Considerations

• —Splits: The dataset is randomly split into training (70%), validation (15%), and test (15%) sets. The training and validation sets include all clinical variables, while the test set has clinical variables removed to simulate real-world testing scenarios.
• —Privacy: The dataset adheres to privacy standards as it is anonymized.
• —Data Imbalance: As survival rates and clinical outcomes are often skewed, it is essential to consider this during model training to avoid bias towards the majority class (patients with shorter survival times).

Columns Removed from the Test Set

To prevent data leakage, certain columns containing post-diagnosis, outcome-related, administrative, or pathology-related data are removed from the test set, as they would not be available at prediction time. These include:

• —Outcome-related Columns: Information directly related to patient outcomes, like days_to_death, vital_status, and progression_or_recurrence.
• —Administrative Identifiers: Technical identifiers irrelevant to prediction, such as sample_id, slide_id, and pathology_report_uuid.
• —Post-Analysis Pathology Data: Percentages of specific cell types (e.g., percent_tumor_cells, percent_normal_cells) derived after tumor analysis.
• —Cancer Staging: Pathologic staging data determined post-surgery (e.g., ajcc_pathologic_stage, ajcc_pathologic_n).

Usage Example

import pandas as pd

train_df = pd.read_csv("hf://datasets/HLMCC/TCGA-PAAD/train.csv")
validation_df = pd.read_csv("hf://datasets/HLMCC/TCGA-PAAD/val.csv")
test_df = pd.read_csv("hf://datasets/HLMCC/TCGA-PAAD/test.csv")

Submission Guidelines

[!IMPORTANT] Participants need to upload a csv file onto this submission portal (https://huggingface.co/spaces/HLMCC/PAAD-submission-portal).

[!CAUTION] Please make sure to have these columns : 'patientid' and 'predictedscores' in your submission.csv file. - patient_id: Unique identifier for each patient - predicted_scores: Predicted survival scores from your model

Citation

The data used in this study are from The Cancer Genome Atlas (TCGA) Research Network: https://www.cancer.gov/tcga. If you use this dataset in your work, please cite the TCGA Research Network:

The Cancer Genome Atlas Research Network. (2017). Comprehensive and Integrated Genomic Characterization of Pancreatic Ductal Adenocarcinoma. Cancer Cell, 32(2), 185-203.e13. https://doi.org/10.1016/j.ccell.2017.07.007

Dataset Curators

The dataset was curated and prepared by the TCGA consortium and processed for easy machine learning applications by Moffitt Cancer Center.