Predicting distal tibia fracture type using demographic, vehicle, and crash factors via a random forest classification algorithm

Distal tibia fractures occur in approximately 4% of police-reported crashes where at least 1 vehicle was towed in the U.S., and frequently result in complications like infection, nonunion, and osteoarthritis. This study used real-world crash data to train a random forest algorithm to predict distal tibia fracture types, identifying key demographic, vehicle, and crash factors contributing to the model predictions. Crash Injury Research and Engineering Network (CIREN) cases (2005–2024) with computed tomography (CT) radiology of a distal tibia fracture were identified. Frontal, non-rollover crashes with non-pregnant, non-ejected occupants older than 13 years were selected, resulting in 94 cases (90 unilateral, 2 bilateral fractures) included for analysis. Fractures were labeled by 3 trained researchers as extraarticular, partial articular, or complete articular using the AO Foundation-Orthopedic Trauma Association (AO/OTA) compendium, which has an established relationship with clinical outcomes. An orthopedic surgeon graded a subset (<i>N</i> = 20) for validation. CIREN injury mechanisms and contributing factors - delta-v, vehicle model year, body type, and safety restraints, as well as occupant age, sex, BMI, and comorbidities - were collected to train a random forest classifier to predict AO/OTA fracture type. SHapley Additive exPlanations were used to examine how demographic, vehicle, and crash factors influenced the probability of random forest predictions, which allowed for the association of key factors from real-world crash data with AO/OTA fracture types. The agreement between the 3 graders (Fleiss’ Kappa of 0.78) and between the ground truth classifications and the orthopedic surgeon (Cohen’s Kappa of 0.76) were substantial. Random forest correctly determined the AO/OTA distal tibia fracture type in 75.5% of cases. Toe-pan intrusion as a contributing factor and delta-v greater than 30 kph were associated with increased random forest prediction probability of complete articular fractures. Knee bolster airbag deployment was associated with decreased probability of partial articular fractures, often linked with rotational injury mechanisms, suggesting they induce increased mechanisms of compression. Concomitantly, the random forest found that rotational mechanisms were associated with increased probability of partial articular fractures. Despite not using radiology, the random forest predicted distal tibia fracture type with similar accuracy to human graders, demonstrating the potential to support in areas like post-crash triage. Finally, the random forest characterized underlying relationships between factors from real-world crash data and AO/OTA fracture types, demonstrating its usefulness as an explanatory model.