stripeddonkey-data/nairobi-longitudinal-healthcare-utilization

Nairobi Longitudinal Healthcare Utilization

Synthetic longitudinal population data for predicting healthcare use across time.

100% synthetic. No real patient records. Every resident, household, insurance record, employment state, education history, encounter, diagnosis, prescription, condition, and healthcare outcome in this release is synthetic. No real individual's data was used to produce it.

What is this?

This dataset is a balanced longitudinal healthcare-utilization panel following the same:

• 10,000 synthetic Nairobi residents
• across 3 annual observation dates
• for 30,000 person-period rows
• with 38 columns

Observation dates are:

• Train — 2023-09-01
• Validation — 2024-09-01
• Test — 2025-09-01

Each row contains point-in-time context available at or before its ObservationDate, including:

• demographics;
• geography;
• household context;
• education history;
• employment state;
• insurance coverage;
• prior healthcare utilization;
• prior diagnosis, prescription, and chronic-condition aggregates;
• maternity history.

The row also contains outcomes observed during the following 12 months.

This makes the release suitable for temporal forecasting and longitudinal analysis, rather than as a single disease table or three unrelated cross-sectional samples.

Research question

Who is most likely to use healthcare — or require inpatient care — during the next 12 months?

Typical uses include:

• 12-month healthcare-utilization prediction
• 12-month inpatient-admission prediction
• future encounter-count modeling
• temporal-generalization research
• longitudinal resident-level analysis
• insurance and healthcare-utilization analysis
• cross-domain synthetic-data research

Load from Hugging Face

from datasets import load_dataset

ds = load_dataset(
    "STRIPEDDONKEY/nairobi-longitudinal-healthcare-utilization"
)

train = ds["train"]
validation = ds["validation"]
test = ds["test"]

Expected rows:

train        10,000
validation   10,000
test         10,000

The same 10,000 synthetic residents appear in all three splits.

Primary targets

AnyHealthcareUseNext12M

Binary primary target indicating whether the resident has qualifying healthcare utilization during the 12 months after ObservationDate.

Observed positive rates:

• Train — 24.86%
• Validation — 26.46%
• Test — 36.05%

InpatientAdmissionNext12M

Binary primary target indicating whether the resident has a qualifying inpatient admission during the 12 months after ObservationDate.

Observed positive rates:

• Train — 4.99%
• Validation — 4.68%
• Test — 6.87%

EncounterCountNext12M

Auxiliary count target representing qualifying healthcare encounters in the forward 12-month window.

Observed range:

0–14

UtilizationIntensityBand

Derived secondary target grouping EncounterCountNext12M into:

• 0
• 1
• 2+

DiedDuringOutcomeWindow

Diagnostic only.

This field represents future-window information and must never be used as a prediction-time feature.

See:

• docs/DATA_DICTIONARY.md
• docs/TEMPORAL_LEAKAGE.md
• docs/MODELING_GUIDE.md

for complete role and leakage guidance.

A genuine balanced panel

This release follows the same residents through time.

The residents in Train are the same residents in Validation and Test.

Independent validation confirmed:

• 10,000 distinct residents
• identical resident membership across all three periods
• 0 duplicate (CitizenId, ObservationDate) rows
• 0 duplicate CandidateKey values

This allows longitudinal questions such as:

How did a resident's prior healthcare use change from one annual observation period to the next?

or:

Does earlier insurance or employment context help predict later healthcare utilization?

Temporal split

The split is intentionally temporal, not random:

2023 → Train
2024 → Validation
2025 → Test

Do not randomly reshuffle the published person-period rows for the default benchmark.

Doing so would:

• mix earlier and later population states;
• weaken temporal evaluation;
• allow later information to enter training;
• break the intended balanced-panel structure.

See docs/SPLIT_METHODOLOGY.md.

Temporal feature contract

For each row:

• historical features use information available at or before ObservationDate;
• prior healthcare windows end at ObservationDate;
• forward targets begin strictly after ObservationDate.

Historical utilization windows use:

(ObservationDate - N months, ObservationDate]

The 12-month forward outcome window uses:

(ObservationDate, ObservationDate + 12 months]

This prevents overlap between historical encounter features and future encounter targets.

See docs/TEMPORAL_LEAKAGE.md.

Feature groups

The panel contains point-in-time information from the same synthetic resident across multiple domains.

Demographics

• CurrentAge
• Sex

Geography

• WardId
• WardName
• SubCountyName

Household

• HouseholdSize

Education

• HasK12EnrollmentByObservationDate
• EverEnrolledTertiary

Employment

• IsEmployedFormal
• IsEmployedInformal
• IsEmployed

Insurance

• IsCoveredAsOfObservation
• CoverageAcquisitionType
• CoveragePlanTier

Prior healthcare history

• PriorEncounterCount12M
• PriorEncounterCount24M
• PriorInpatientCount12M
• PriorEmergencyCount12M
• PriorDiagnosisCount12M
• PriorPrescriptionCount12M
• PriorChronicConditionCount
• HasRecentMaternityRecord24M

See docs/DATA_DICTIONARY.md for the complete 38-column definition.

Companion files

The companion/ folder provides additional resident-level evidence and examples from the same corrected frozen Health state.

companion/health_encounters.csv

Encounter-level history used to independently reconstruct the principal visit-derived panel fields.

companion/patient_journeys.csv

Selected longitudinal resident journeys for inspection, demonstration, and notebook examples.

companion/health_population_profile.csv

Population-level synthetic Health context supporting interpretation of the release.

The companion files support transparency and evidence inspection but do not replace the main longitudinal panel.

Exact companion reconciliation

Across all 30,000 person-period rows, the following published fields were independently reconstructed from the encounter companion with:

0 discrepancies

• PriorEncounterCount12M
• PriorEncounterCount24M
• PriorInpatientCount12M
• PriorEmergencyCount12M
• EncounterCountNext12M
• AnyHealthcareUseNext12M
• InpatientAdmissionNext12M
• PriorPrescriptionCount12M

UtilizationIntensityBand also derives exactly from EncounterCountNext12M.

The public companion does not expose the complete diagnosis- and condition-level fact sets required to independently reconstruct:

• PriorDiagnosisCount12M
• PriorChronicConditionCount

No stronger companion-reconstruction claim is made for those fields.

Corrected v1.2 release

The current release reflects a corrected post-remediation Health state.

Compared with the earlier panel:

• 2,524 person-period rows changed
• 942 distinct residents were affected
• Train — 762 changed rows
• Validation — 837 changed rows
• Test — 925 changed rows

The deterministic publication cohort itself remained unchanged.

The earlier compressed future-encounter distribution no longer applies.

The corrected EncounterCountNext12M range is:

0–14

See:

• docs/RELEASE_NOTES.md
• docs/EVIDENCE_PACK.md
• docs/CERTIFICATION.md

for the correction and reconciliation record.

Temporal drift

Healthcare-utilization prevalence increases in the Test period.

Observed synthetic panel rates are:

Train — 2023

AnyHealthcareUseNext12M

24.86%

InpatientAdmissionNext12M

4.99%

Validation — 2024

AnyHealthcareUseNext12M

26.46%

InpatientAdmissionNext12M

4.68%

Test — 2025

AnyHealthcareUseNext12M

36.05%

InpatientAdmissionNext12M

6.87%

This increase is a real property of the synthetic release.

It should not be interpreted as evidence of a real Nairobi epidemiological or healthcare-utilization trend.

Treat the dataset as a temporal-generalization / distribution-shift benchmark, not as a source of empirical year-over-year trend estimates.

Balanced-panel survivorship

The requirement that residents appear at all three observation dates introduces a known survivorship characteristic.

A resident who does not survive to a later required observation date cannot belong to the balanced publication cohort, even at an earlier period.

Consequently:

• Train has DiedDuringOutcomeWindow = 0
• Validation has DiedDuringOutcomeWindow = 0
• Test may contain outcome-window deaths

In Test:

66 / 10,000 rows (0.66%)

have DiedDuringOutcomeWindow = 1.

Do not use this panel to estimate mortality.

See docs/LIMITATIONS.md.

Birth-cohort floor

Because residents must already exist at the first observation date, residents born after:

2023-09-01

are absent from the balanced panel.

This affects age-sensitive distributions and contributes to the high prevalence of historical K-12 enrollment in the sample.

Do not treat sample education-history rates as citywide empirical estimates.

External calibration

The panel is internally consistent and encounter-reconciled, but absolute whole-resident healthcare-utilization levels have not yet been fully externally calibrated against an applicable empirical Nairobi benchmark.

Therefore, do not use this release to estimate observed Nairobi healthcare-utilization levels directly.

This is an external-realism limitation, not an internal-consistency defect.

See:

• docs/GROUNDING.md
• docs/LIMITATIONS.md

Full certified parent product

The larger certified parent product contains:

• 12,177,749 person-period rows
• 4,224,188 distinct synthetic citizens
• all three annual observation periods

The Hugging Face release is a deterministic, balanced 10,000-resident publication cohort derived from that larger certified synthetic product.

The public panel receives additional release-level validation, including exact encounter-companion reconciliation for the reconstructable utilization measures.

See docs/CERTIFICATION.md.

Repository contents

README.md
LICENSE
CITATION.cff
REPAIR_LOG.md

data/
  longitudinal_panel/
    train.csv
    validation.csv
    test.csv

companion/
  README.md
  health_encounters.csv
  patient_journeys.csv
  health_population_profile.csv

docs/
  CERTIFICATION.md
  DATA_CARD.md
  DATA_DICTIONARY.md
  EVIDENCE_PACK.md
  FEATURE_DICTIONARY.csv
  GENERATION_AND_CURATION.md
  GROUNDING.md
  LIMITATIONS.md
  LINEAGE.md
  MODELING_GUIDE.md
  PROVENANCE.md
  RELEASE_NOTES.md
  SPLIT_METHODOLOGY.md
  TEMPORAL_LEAKAGE.md
  VALIDATION_REPORT.md

notebooks/
  01_quickstart.ipynb
  02_citizen_health_journey.ipynb
  03_ml_baseline.ipynb
  04_cohort_utilization_trends.ipynb

metadata/
  certification_summary.json
  dataset_manifest.json
  grounding_manifest.csv
  release_manifest.json
  sample_manifest.csv
  sha256sums.txt

samples/
  preview_15.csv

verify_release.py

Notebooks

The repository contains four public notebooks:

01_quickstart.ipynb

Fastest path to loading the panel, inspecting splits, and understanding the primary targets.

02_citizen_health_journey.ipynb

Worked longitudinal example following one synthetic resident through healthcare history.

03_ml_baseline.ipynb

Leakage-aware baseline models for the two primary targets.

04_cohort_utilization_trends.ipynb

Descriptive cohort analysis across geography, insurance, and observation periods.

Validation

The release passed checks covering:

• panel grain
• identical resident membership across periods
• observation-date / split alignment
• target/count consistency
• non-negative utilization measures
• utilization-band consistency
• point-in-time feature semantics
• temporal leakage controls
• encounter-level reconciliation
• prescription-count reconciliation
• file integrity

Selected results:

• 30,000 person-period rows
• 10,000 distinct residents
• 0 duplicate person-period keys
• 0 / 30,000 discrepancies across reconstructable visit-derived measures
• 0 / 30,000 discrepancies for prior prescription count
• future encounter range 0–14

See docs/VALIDATION_REPORT.md.

Modeling guidance

For the default forecasting benchmark:

• preserve the published temporal split;
• never use CitizenId as a feature;
• never use CandidateKey as a feature;
• never use SplitName as a feature;
• never use DiedDuringOutcomeWindow as a prediction-time feature;
• never use future targets as predictors;
• use only information available at or before ObservationDate.

See:

• docs/MODELING_GUIDE.md
• docs/TEMPORAL_LEAKAGE.md
• docs/DATA_DICTIONARY.md

Known limitations

Important limitations include:

• balanced-panel survivorship;
• birth-cohort exclusion after 2023-09-01;
• synthetic temporal distribution shift;
• incomplete external calibration of absolute healthcare-utilization levels;
• approximately 7.6% unresolved HouseholdSize;
• synthetic insurance and education prevalence values should not be treated as empirical Nairobi estimates;
• complete diagnosis and condition fact sets are not exposed through the public companion;
• single-city scope.

See docs/LIMITATIONS.md for the full disclosure.

License

CC-BY-4.0

This dataset is fully synthetic and is not derived from real personal or medical records.

See LICENSE for the complete licensing terms.