Mattysmittttt/crashrisk-qld-severity

CrashRisk-QLD — Severity classifier (4 classes)

LightGBM classifier predicting the severity of a Queensland road crash given pre-/at-crash conditions. Classes: fatal, hospitalisation, medical treatment, minor injury.

© State of Queensland (Department of Transport and Main Roads) 2026. Licensed CC-BY 4.0. Source: data.qld.gov.au, dataset 'Crash data from Queensland roads', version rqC45037 (2025-06), retrieved 2026-04-30.

Disclaimer. This is a population-level statistical model trained on publicly reported crash data. It is NOT suitable for individual driver risk assessment, insurance underwriting, pre-incident law enforcement targeting, or any decision with legal or financial consequence to an individual. Use it for research, road-safety analysis, and education.

Quick start

from huggingface_hub import hf_hub_download
import joblib, lightgbm as lgb, json

REPO = "Mattysmittttt/crashrisk-qld-severity"
booster = lgb.Booster(model_file=hf_hub_download(REPO, "model.txt"))
pre = joblib.load(hf_hub_download(REPO, "preprocessor.joblib"))
classes = json.loads(open(hf_hub_download(REPO, "label_classes.json")).read())
# `pre` is a fitted sklearn ColumnTransformer; pass a DataFrame whose
# columns match the dataset's ✅ feature columns. See dataset card for
# the schema, and feature_importance.csv for the names this model uses.

Intended uses

• Road-safety research and descriptive analytics.
• Education on tabular ML, class imbalance, and leakage prevention.
• As a baseline severity model in road-safety pipelines.

Out of scope

• Insurance underwriting — using population-level statistical patterns to set individual premiums creates fairness concerns and is outside this model's intended scope.
• Individual driver risk assessment — these features describe road conditions and aggregate vehicle context, not driver behaviour or identity.
• Pre-incident law enforcement targeting — geographic patterns may reflect reporting biases as much as actual risk; using them to pre-target locations creates feedback loops.
• Any decision with legal or financial consequence to a single individual — full stop.

Training data

See the dataset card for full provenance, cleaning, schema, and split definition.

• Train (2011–2022): 152,842 rows
• Validation (2023): 13,622 rows (used for early stopping and Optuna HPO)
• Test (2024): 14,358 rows

Training details

• Model: LightGBM multiclass objective, multi_logloss metric
• Class imbalance: balanced sample weights (sklearn.utils.class_weight.compute_sample_weight('balanced', y))
• Hyperparameter tuning: Optuna TPE, 20 trials, minimising val log-loss with early stopping
• Best iteration on val: 1686
• Best HPO config:

{
  "num_leaves": 227,
  "learning_rate": 0.010205401359666888,
  "feature_fraction": 0.7235023898845259,
  "bagging_fraction": 0.6068744601075835,
  "bagging_freq": 6,
  "min_data_in_leaf": 76,
  "lambda_l1": 1.4239280268649237e-08,
  "lambda_l2": 1.2540363726263013e-08
}

• Features after preprocessing: 173 (low-card → OHE, high-card → TargetEncoder, numerics passthrough)

Evaluation (held-out test = 2024)

• Accuracy: 0.4404 (sanity check: must NOT exceed 0.70 — observed: ok)
• Macro-F1: 0.3259 [95% bootstrap CI 0.3164, 0.3351]
• Log-loss: 1.1820 [1.1680, 1.1965]

Per-class metrics

class	precision	recall	f1	support
fatal	0.077	0.117	0.093	273
hospitalisation	0.609	0.580	0.594	7,051
medical treatment	0.402	0.225	0.288	4,409
minor injury	0.255	0.462	0.328	2,625

Confusion matrix

Rows = actual, columns = predicted.

—	pred fatal	pred hospitalisation	pred medical treatment	pred minor injury
actual fatal	32	198	11	32
actual hospitalisation	269	4088	949	1745
actual medical treatment	79	1565	990	1775
actual minor injury	34	866	512	1213

Top 20 feature importances (LightGBM gain)

1. loc_suburb_0 (gain = 1574725)
2. loc_suburb_3 (gain = 574071)
3. crash_hour (gain = 562661)
4. loc_abs_statistical_area_2_0 (gain = 468933)
5. count_unit_car (gain = 466195)
6. crash_year (gain = 464291)
7. loc_suburb_2 (gain = 403835)
8. loc_suburb_1 (gain = 391710)
9. crash_longitude (gain = 367704)
10. crash_latitude (gain = 332986)
11. loc_abs_statistical_area_2_3 (gain = 300528)
12. loc_post_code_0 (gain = 260922)
13. crash_speed_limit (gain = 254960)
14. loc_post_code_3 (gain = 242106)
15. loc_abs_statistical_area_2_2 (gain = 240559)
16. loc_abs_statistical_area_2_1 (gain = 235112)
17. loc_post_code_2 (gain = 199437)
18. loc_post_code_1 (gain = 194552)
19. loc_state_electorate_3 (gain = 183491)
20. loc_state_electorate_0 (gain = 183178)

SHAP explainability

Global and local SHAP plots for this model are in the reports/shap_severity/ folder of the source repository:

• summary_bar.png — mean |SHAP| across the test sample
• summary_beeswarm.png — per-feature SHAP distribution
• dependence_*.png — top-6 features
• force_plot_examples.png — TP / TN / FP local explanations for the fatal class
• values.parquet — raw SHAP values, reproducible

Geographic surface

Per-LGA mean predicted P(fatal) under a fixed condition grid (wet/dry × day/night × intersection/no-feature × 60 km/h vs 110 km/h) is published in reports/maps/ of the source repository as both interactive HTML and static PNG.

Limitations & biases

• Trained on reported crashes only — under-reporting is real.
• Possible feedback loops if used to allocate enforcement or infrastructure spend (geographic risk concentrates because past incidents concentrated, not necessarily because future ones will).
• Geographic coverage variance across QLD: urban LGAs are very well-represented; remote LGAs much less so.
• Temporal drift: 2011–2024 includes vehicle-safety improvements, regulatory changes, and recording-system migrations.

Ethical considerations

• Population-level, not causal: The model encodes correlations between pre-crash conditions and recorded outcomes. It does not assign fault and cannot be read as a statement about individual responsibility.
• Geographic predictions can stigmatise: We publish per-LGA aggregates only, never per-address. Even at LGA level, higher predicted risk reflects historical reporting and demographics as much as it does inherent road danger.
• Demographic features deliberately excluded: We do not include gender, age, or any demographic field, even though some are present in the casualties aggregates. This is to avoid encoding protected-class proxies. Vehicle-type counts are kept because they describe the crash configuration, not the people involved.
• Reporting bias: This is a model of reported crashes, not true crashes. Under-reporting is differential by severity (PDO under-reported, fatal generally fully reported) and by region.

Citation

@software{crashrisk_qld_severity_2026,
  title  = {CrashRisk-QLD severity classifier},
  author = {Mattysmittttt},
  year   = {2026},
  url    = {https://huggingface.co/Mattysmittttt/crashrisk-qld-severity},
  note   = {Trained on Mattysmittttt/qld-traffic-crashes-clean; source data CC-BY 4.0 © State of Queensland (Department of Transport and Main Roads).}
}

License

Released under CC-BY 4.0. Attribution must be preserved.