electricsheepafrica/ssa-chemotherapy-response

SSA Chemotherapy Response Dataset (Women, Multi-ancestry, Synthetic)

Dataset summary

This dataset provides a synthetic cohort of women receiving first-line chemotherapy for common solid tumors across multiple ancestry groups with a focus on sub-Saharan Africa (SSA). It captures:

• Cancer type and subtype (breast, colorectal, cervical, ovarian).
• First-line chemotherapy regimen type.
• Best treatment response (CR, PR, SD, PD) and derived objective response and disease-control endpoints.
• Grade 3–4 toxicity profiles (neutropenia, febrile neutropenia, neuropathy, GI and hepatic AEs).

The dataset is designed for methods development and education around treatment response modeling and toxicity prediction; all individuals and outcomes are synthetic.

Cohort design

Sample size and populations

• Total N: 10,000 synthetic women.

• Populations:

  • SSA_West: 2,000
  • SSA_East: 2,000
  • SSA_Central: 1,500
  • SSA_Southern: 1,500
  • AAW (African American women): 1,500
  • EUR (European reference): 1,000
  • EAS (East Asian reference): 500

• Age range: 18–85 years, with modest differences in mean age by population (slightly older ages in high-income/AAW groups), approximating typical oncology cohorts.

Population labels align with other Electric Sheep Africa synthetic datasets to facilitate multi-modal integration (e.g., combining treatment response with genomics, stage, or body composition).

Cancer types, subtypes, and regimens

Cancer types

• breast
• colorectal
• cervical
• ovarian

Approximate cancer-type mix:

• Breast: 40%
• Colorectal: 30%
• Cervical: 15%
• Ovarian: 15%

Subtypes

• Breast:

  • TNBC – triple-negative.
  • HRpos_HER2neg – hormone-receptor–positive, HER2-negative.
  • HER2pos – HER2-positive.

• Colorectal:

  • left – left-sided colon.
  • right – right-sided colon.
  • rectal – rectal primaries.

• Cervical:

  • squamous
  • adenocarcinoma

• Ovarian:

  • HGSOC – high-grade serous ovarian carcinoma.
  • endometrioid_CCC – endometrioid/clear cell.
  • mucinous

These are simplified subtype groupings meant to retain major clinical distinctions without attempting full real-world heterogeneity.

Regimens

Regimens are specified by cancer type:

• Breast:

  • AC_T – anthracycline → taxane.
  • taxane_only – taxane-based regimen without anthracycline.
  • HER2_targeted – chemotherapy plus HER2-directed therapy for HER2-positive cancers.

• Colorectal:

  • FOLFOX – infusional 5-FU/leucovorin + oxaliplatin.
  • FOLFIRI – 5-FU/leucovorin + irinotecan.
  • FOLFOXIRI – 5-FU/leucovorin + oxaliplatin + irinotecan.

• Cervical:

  • cisplatin_5FU_CRT – cisplatin/5-FU concurrent with radiation.
  • cisplatin_single – single-agent cisplatin.

• Ovarian:

  • carboplatin_paclitaxel – platinum-taxane doublet.
  • carboplatin_single – single-agent carboplatin.

Regimen distributions per cancer type reflect common practice (e.g., ovarian cancer dominated by carboplatin/paclitaxel).

Response endpoints

Variables

• best_response – best observed response during first-line treatment:

  • CR – complete response.
  • PR – partial response.
  • SD – stable disease.
  • PD – progressive disease.

• objective_response – boolean (CR or PR).

• disease_control – boolean (CR, PR, or SD).

Probability model

Best response is sampled from regimen- and cancer-type–specific probabilities informed by the chemotherapy literature (anthracycline+taxane, FOLFOX/FOLFIRI/FOLFOXIRI, carboplatin/paclitaxel, cisplatin/5-FU chemoradiation):

• Breast, anthracycline + taxane (AC_T) (first-line metastatic):

  • ORR (CR+PR) ~55%, SD ~25%, PD ~20%.

• Breast, HER2-targeted (HER2pos):

  • Higher ORR (CR+PR ~65%) with lower PD fraction.

• Breast, taxane_only:

  • Slightly lower ORR than combination.

• Metastatic colorectal cancer:

  • FOLFOX and FOLFIRI: ORR ~40–50%; SD ~30%; PD ~20–30%.
  • FOLFOXIRI: higher ORR (~50%+) at expense of more toxicity.

• Cervical cancer chemoradiation:

  • cisplatin_5FU_CRT: high CR rates (~55%) with additional PR and SD.
  • cisplatin_single: lower CR and ORR.

• Ovarian cancer:

  • carboplatin_paclitaxel: ORR ~60–70% (CR+PR) with some SD and PD.
  • carboplatin_single: lower ORR and more PD.

These probabilities are encoded in chemotherapy_response/configs/chemotherapy_response_config.yaml and used directly by the generator.

Toxicity profiles

Variables

Grade 3–4 toxicity indicators (CTCAE-like):

• any_grade3_4_toxicity
• neutropenia_g3_4
• febrile_neutropenia
• neuropathy_g3_4
• GI_toxicity_g3_4 (grade 3–4 GI events, including severe nausea/vomiting/diarrhea)
• hepatic_toxicity_g3_4 (e.g. transaminitis/bilirubin elevation)

Regimen-specific probabilities

Per-regimen grade 3–4 AE probabilities are set based on published trials and reviews:

• Breast (AC_T): higher neutropenia and some FN; moderate neuropathy.

• Taxane-only: more neuropathy, somewhat less neutropenia.

• HER2-targeted: moderate hematologic and GI AEs.

• FOLFOX/FOLFIRI: intermediate neutropenia (more with oxaliplatin/FOLFOX), GI AEs higher with irinotecan (FOLFIRI).

• FOLFOXIRI: highest grade 3–4 neutropenia (~45–50%) and notable GI toxicity.

• Cisplatin/5-FU chemoradiation: GI and hematologic AEs consistent with chemoradiation data.

• Carboplatin/paclitaxel: grade 3–4 neutropenia in ~35% and neuropathy in ~10–15% of patients.

The generator samples each AE as Bernoulli(p) per regimen and derives any_grade3_4_toxicity accordingly.

Files and schema

chemotherapy_response_data.parquet / chemotherapy_response_data.csv

One row per woman-regimen:

• Demographics

  • sample_id
  • population, region, is_SSA, is_reference_panel
  • sex – Female
  • age

• Oncology structure

  • cancer_type
  • subtype
  • regimen

• Response

  • best_response – CR, PR, SD, PD.
  • objective_response – True if CR/PR.
  • disease_control – True if CR/PR/SD.

• Toxicity

  • any_grade3_4_toxicity
  • neutropenia_g3_4
  • febrile_neutropenia
  • neuropathy_g3_4
  • GI_toxicity_g3_4
  • hepatic_toxicity_g3_4

Generation and validation

Generation

• Script: chemotherapy_response/scripts/generate_chemotherapy_response.py
• Config: chemotherapy_response/configs/chemotherapy_response_config.yaml
• Literature: chemotherapy_response/docs/LITERATURE_INVENTORY.csv

Key steps:

1. Generate multi-ancestry sample with age distribution.
2. Assign cancer type, subtype, and regimen using configured distributions.
3. Sample best response from regimen- and cancer-type–specific CR/PR/SD/PD probabilities.
4. Derive ORR and disease-control flags.
5. Sample grade 3–4 toxicities based on regimen-specific probabilities.

Validation

• Script: chemotherapy_response/scripts/validate_chemotherapy_response.py
• Report: chemotherapy_response/output/validation_report.md

Checks:

• C01–C02 – Cohort size and population counts vs config.
• C03 – Cancer type, subtype, and regimen distributions vs config.
• C04 – Response distributions (CR/PR/SD/PD) by cancer type and regimen.
• C05 – Grade 3–4 toxicity rates by regimen.
• C06 – Missingness in key variables.

For the released version, the validator reports overall status PASS, with minor WARN-level deviations where Monte Carlo sampling produces small differences from target probabilities.

Intended use

This dataset is intended for:

• Methods development in:
  • Treatment-response modeling.
  • Toxicity prediction and risk stratification.
  • Regimen selection algorithms in synthetic environments.
• Teaching on:
  • CR/PR/SD/PD definitions.
  • ORR and disease-control endpoints.
  • Regimen-specific toxicity profiles.

It is not suitable for:

• Clinical decision-making or individual patient management.
• Estimating real-world response or toxicity rates.
• Regulatory or guideline development.

All records are synthetic.

Ethical considerations

• No real patient data are used.
• Population and cancer subtype labels are for simulation realism only.
• Analyses should be interpreted as methodological exercises rather than statements about specific populations or regimens.

License

• License: CC BY-NC 4.0.
• Free to use for non-commercial research, teaching, and methods development with attribution.

Citation

If you use this dataset, please cite:

Electric Sheep Africa. "SSA Chemotherapy Response Dataset (Women, Multi-ancestry, Synthetic)." Hugging Face Datasets.

and, where appropriate, cite key chemotherapy and toxicity trials (anthracycline+taxane breast cancer, FOLFOX/FOLFIRI/FOLFOXIRI in colorectal cancer, carboplatin/paclitaxel in ovarian cancer, cisplatin/5-FU chemoradiation in cervical cancer) that informed the configuration.