Project_​v.01

Builds the Decision Tree model

Measures the overall classification performance

evaluates how well the model performed

Balance classes in training folds

Uses the trained Decision Tree model to make predictions

Combines the results from all folds into one final output

Measures the overall classification performance

Splits the data into training and testing folds for cross-validation

Balance classes in training folds

Combines the results from all folds into one final output

evaluates how well the model performed

Accident

vehicle

person

distract

Binned roadway surface condition into dry, wet, snow/ice/slush, other surface, and non-trafficway/driveway categories; not reported and unknown left missing.

Created binary rollover variable: coded no rollover versus any rollover involvement.

Created binned trafficway-description variable: grouped original VTRAFWAY codes into non-trafficway/driveway, two-way not divided, two-way divided, one-way trafficway, and entrance/exit ramp categories; not reported and unknown values were left missi

Binned number of roadway lanes into 1–2 lanes and 3+ lanes; unknown or invalid lane values left missing.

Binned pre-event vehicle movement into broader movement categories; no-driver and unknown values were left missing.

Binned roadway alignment into straight versus curve categories; non-trafficway/driveway was retained and not reported/unknown values were left missing.

Created binary vehicle-removal variable indicating whether the vehicle was towed or not towed.

Created binary driver-distraction variable: coded 0 as not distracted and all specific distraction-related codes as distracted; not reported and unknown values left missing.

Measures the overall classification performance

Combines the results from all folds into one final output

Binned roadway grade into level and non-level grade categories; not reported and unknown values were left missing.

Binned traffic control devices into no controls, traffic signals, regulatory signs, and other controls; not reported and unknown values were left missing.

Created binary fire-occurrence variable: coded 0 as “No Fire” and 1 as “Fire Occurred.”

Created binary interstate-highway variable: coded whether the crash occurred on an interstate highway

Binned traffic control functioning into no controls, functioning properly, and not functioning properly; not reported and unknown values were left missing.

Binned vehicle damage severity into no damage, minor, functional, and disabling damage; not reported and unknown values were left missing.

INJ_SEV1 to 4 only

Handled missing predictor values before modeling. Numeric variables were replaced using the median,

INJ_SEVHigh vs Low

Final variables in the model

Splits the data into training and testing folds for cross-validation

valid male/female onlyRemoved not reported/unknown sex codes before recoding.

Created REST_USE_binnedCollapsed valid restraint-use codes into restrained vs. not restrained; other/not reported/unknown codes were left as missing.Created binary restraint-use variable. Valid restraint-use codes were grouped as “Restraint Used,”

Created vehicle-count category from VE_TOTAL: coded crashes as single-vehicle or multiple-vehicle to capture crash complexity without using injury-outcome information.

Created binary alcohol-involvement variable: coded 0 as “No Alcohol” and 1 as “Alcohol Involved”; not reported and unknown values were left missing/unmatched.

Cleaned numeric number-injured variable by retaining valid values 0–97 and leaving special/non-informative codes missing before modeling.

Created binary police-reported drug involvement variable: coded 0 as “No Drug Involvement” and 1 as “Drug Involvement”; not reported and unknown values were left missing/unmatched.

Created binary air bag deployment variableGrouped all air bag deployment codes as “Air Bag Deployed” and code 20 as “Air Bag Not Deployed”; non-informative codes such as not reported, not applicable, or unknown were left missing/unmatched.

Compares the ROC performance of all models in one view

Created binary ejection variable: coded 0 as “Not Ejected” and codes 1–3 as “Ejected”; not reported, not applicable, and unknown codes were left missing/unmatched.

Created season variable from crash month: grouped months into Winter, Spring, Summer, and Fall to capture seasonal crash patterns.

Created light-condition variable: grouped crash lighting into Daylight, Dark, and Dawn/Dusk; other, not reported, and unknown light-condition codes were left missing/unmatched.

Binned initial impact location into front, side, rear, top, undercarriage, and non-collision/no-impact categories; unknown and not reported values were left missing.

Created time-of-day variable from crash hour: grouped valid 0–23 hour values into late night, morning, afternoon, and evening/night; unknown hour values were left missing/unmatched.

Created binned crash-type variable from ACC_TYPE: grouped detailed crash-type codes into broader crash-mechanism categories including single-driver, same-direction, opposite-direction, intersecting-paths, and backing/other crashes.

Created binned first-harmful-event variable: grouped detailed harmful-event codes into non-collision, collision with motor vehicle, collision with non-fixed object, and collision with fixed object categories; unknown or unmatched codes were left miss

evaluates how well the model performed

Grouped critical precrash event into broader crash-mechanism categories to capture the event that immediately preceded the crash.

Created weather-condition variable: grouped clear/cloudy conditions separately from adverse weather conditions; other, not reported, and unknown weather codes were left missing/unmatched.

Splits the data into training and testing folds for cross-validation

Grouped original collision-orientation codes into interpretable categories; not reported and unknown codes were left missing/unmatched.

Balance classes in training folds

Created binary hit-and-run variable: coded 0 as “No Hit-and-Run” and 1 as “Hit-and-Run”; any unmatched values were left missing.

Measures the overall classification performance

Created binned specific junction-location variable: grouped RELJCT2 into non-junction, intersection-related, driveway access, ramp/acceleration-lane related, railway crossing, crossover, and other junction/interchange locations; not reported and unkn

Combines the results from all folds into one final output

Created binary work-zone variable: coded 0 as “No Work Zone” and codes 1–4 as “Work Zone” to capture whether the crash was work-zone related.

Splits the data into training and testing folds for cross-validation

Combines the results from all folds into one final output

evaluates how well the model performed

Balance classes in training folds

Measures the overall classification performance

evaluates how well the model performed

Measures the overall classification performance

Balance classes in training folds

Splits the data into training and testing folds for cross-validation

evaluates how well the model performed

SEAT_POSDriver/11 only

Balance classes in training folds

Combines the results from all folds into one final output

BODY_TYPPassanger Car ONLY

Splits the data into training and testing folds for cross-validation