Challenge data

Description

Competitive challenge

Biology

Health

Regression

Tabular

Less than 10MB

Intermediary level

Dates

Started on Jan. 8, 2025

Challenge context

Challenge goals

In partnership with Institut Gustave Roussy, this year’s QRT data challenge focuses on **predicting disease risk for patients diagnosed with blood cancer**, specifically within subtypes of adult myeloid leukemias. For patients with these conditions, **risk assessment is measured by overall survival** — the period from the initial diagnosis to either the patient’s death or the last recorded follow-up.

**Why is this important?** Estimating a patient’s prognosis is vital for tailoring their treatment approach. Patients identified as having a low-risk profile may receive supportive therapies aimed at improving blood counts and overall quality of life, while those identified as high-risk candidates may be prioritized for more intensive treatment options, such as hematopoietic stem cell transplantation.

Accurate risk predictions could lead to better clinical decision-making, improved patient outcomes, and more efficient resource allocation within healthcare facilities.

This challenge gives participants a unique opportunity to work with **real data from 24 clinical centers** and contribute to a meaningful application of data science in the medical field.

Data description

The dataset is structured into **two zip files: X_train.zip** and **X_test.zip**, as well as one CSV file: **Y_train.csv**. The **training set** contains data on **3,323 patients**, and the **test set** contains data on **1,193 patients**. The zip files containing input data are separated into two sets: **Clinical Data** and **Molecular Data**.

The **ID** column is the **unique patient identifier** that links **Clinical Data**, **Molecular Data**, and **Y_train**.

---

### Outcome

The goal of the challenge is to predict **overall survival (OS)** for patients diagnosed with blood cancer. Two key outcomes for each patient are provided in **Y_train**:

- **OS_YEARS**: The overall survival time in years from diagnosis.
- **OS_STATUS**: Indicator of survival status, where 1 denotes death and 0 denotes that the patient was alive at the last follow-up.

**The expected output is a CSV file of patient IDs with a prediction of disease risk. As explained below in the Metric, the scale of the predictions doesn't matter, only the order. For two patients `i` and `j`, a lower predicted risk for `i` than for `j` indicates that we predict that `i` will survive longer than `j`. **

**The submitted file must have the column named "ID" as index, containing the patient IDs, and a column named "risk_score", containing the predicted risk of death.**
A submission file containing random predictions is provided as an example in the Files section.

---

### Loss Metric: Concordance Index for Right-Censored Data with IPCW

The **loss metric** is the **Concordance Index for Right-Censored Data with IPCW (IPCW-C-index)**.

#### Concordance Index (C-index)

The **C-index** measures how well a predictive model can correctly rank survival times. **It calculates the proportion of all comparable pairs of individuals for which the model's disease risk predictions are in the correct order relative to their actual survival times.**

For example a patient `j` who has died at time `T_j` cannot form a “comparable” pair with a patient `i` who has survived up to a time `T_i<T_j`, because we do not know if this patient `i` would have survived up to `T_j`.

However, it can be determined that this patient `j` survived longer than all patients `i` who died before `T_j` (`T_i<T_j`). Similarly, it can be determined that patient `j` died before all patients who either survived or died at a time `T_i` when `T_i>T_j`.

For a "comparable" pair of individuals `i` and `j` with `T_i<T_j`, a pair is correctly ordered by a model when predicted disease risks are in the right order `R_i>R_j`.

For all these “comparable” pairs, the C-index evaluates the number of pairs correctly ordered by a model divided by the total number of comparable pairs:

```math
C=\frac{\text{Number of Concordant Pairs}}{\text{Total Number of Comparable Pairs}}
```

For a model with survival predictions, the **C-index** ranges between 0 and 1:
- **1**: Perfect concordance (ideal ranking).
- **0.5**: Random model (no predictive power).

#### IPCW-C-index

The **IPCW-C-index** extends the traditional C-index to better handle **right-censored data** by applying **inverse probability of censoring weights (IPCW)**. Censorship is defined by the fact that for individuals, data only contains partial survival time, and that there is no information about total survival time (this is the case for patients who are still alive at the time of the last observation). This approach modifies the C-index calculation by assigning different weights to data pairs, depending on the likelihood that each observation would have been observed (i.e., not censored).

The metric is imported from **scikit-survival** ([scikit-survival documentation](https://scikit-survival.readthedocs.io/en/stable/api/generated/sksurv.metrics.concordance_index_ipcw.html)) in the benchmark and is **clipped at looking at a maximum horizon of 7 years**.

---

### Clinical Data (One Row per Patient)

Each patient is associated with a unique identifier and detailed clinical information:

- **ID**: Unique identifier for each patient.
- **CENTER**: The clinical center where the patient is treated.
- **BM_BLAST**: Bone marrow blasts percentage, indicating the proportion of abnormal blood cells found in the bone marrow of patients.
- **WBC**: White Blood Cell count, measured in Giga/L (blood count parameter).
- **ANC**: Absolute Neutrophil Count in the blood, in Giga/L (blood count parameter).
- **MONOCYTES**: Monocyte count in the blood, in Giga/L (blood count parameter).
- **HB**: Hemoglobin level, measured in g/dL (blood count parameter).
- **PLT**: Platelet count, in Giga/L (blood count parameter).
- **CYTOGENETICS**: Detailed description of the patient’s karyotype, representing chromosomal abnormalities detected in blood cancerous cells. Notations follow the **ISCN standard**, with typical markers like **“46,XX”** for a normal female karyotype or **“46,XY”** for a normal male karyotype. Abnormalities like monosomy 7 (loss of chromosome 7) indicate higher-risk disease.

---

### Gene Molecular Data (One Row per Somatic Mutation per Patient)

Somatic (acquired) mutations are mutations specific to cancerous cells, not found in the patient’s normal cells. This data captures genetic mutations at the cellular level:

- **ID**: Unique identifier for each patient.
- **CHR, START, END**: Chromosomal position of the mutation on the human genome.
- **REF, ALT**: The reference nucleotide and the alternate (mutant) nucleotide.
- **GENE**: The gene affected by the mutation.
- **PROTEIN_CHANGE**: The impact of the mutation on the protein produced by the gene.
- **EFFECT**: Broad classification of the mutation’s impact on gene function.
- **VAF**: Variant Allele Fraction, representing the proportion of cells carrying the mutation.

Benchmark description

The given benchmark shows two models:

1. **A simple LightGBM (Light Gradient-Boosting Machine)** model using only clinical data, not accounting for censoring.
2. **A Cox Proportional Hazards Model**, which considers both clinical data and (limited) gene mutation information.

The first benchmark (LGBM model) is provided as an example. The score obtained with the second benchmark (Cox model) is the one that is considered for the ranking. A benchmark notebook, containing the code used to generate the benchmarks, is provided in the additional files section.

### About Cox Proportional Hazards Model
A Cox Proportional Hazards Model is a statistical technique, close to a linear regression, used in survival analysis to examine the effect of several variables on the time to a specified event. It models the hazard function, which represents the instantaneous risk of the event at a given time, assuming that the effect of each predictor is multiplicative and remains constant over time (hence **"proportional hazards"**). The model is semi-parametric, meaning it doesn’t assume a specific baseline hazard function, making it flexible for analyzing time-to-event data, especially with censored observations. *(See [Wikipedia](https://en.wikipedia.org/wiki/Proportional_hazards_model))*

Feel free to build your own features, draw inspiration from the existing benchmark, and build your own models.

**Good luck for this challenge!**

Files

Files are accessible when logged in and registered to the challenge

The challenge provider

Qube Research & Technologies (QRT) is a global quantitative and systematic investment manager, operating in all asset classes across the world. Driven by a shared passion for data, research, technology and trading expertise, we strive to deliver high-quality returns for our investors. Established in 2016, QRT can rely on its 1,400+ employees across 11 offices in Europe, Middle East and Asia Pacific. QRT supports various coding initiatives as well as academic projects developing and promoting maths and science education.

PROVIDER WEBSITE

Challenge Data

Overall Survival Prediction for patients diagnosed with Myeloid Leukemia
by QRT

Description

Dates

Challenge context

Challenge goals

Data description

Benchmark description

Files

The challenge provider

Challenge Data

Overall Survival Prediction for patients diagnosed with Myeloid Leukemia by QRT

Description

Dates

Challenge context

Challenge goals

Data description

Benchmark description

Files

The challenge provider

Overall Survival Prediction for patients diagnosed with Myeloid Leukemia
by QRT