Multifactor analysis of cost of injury using classification and regression trees

In this study we have two objectives: 1. To establish a dominant cost of injury structure. We perform an analysis of full medical cost for crash injuries including hospital charges and professional fees, to determine a cost structure model that can be extrapolated to higher level datasets. 2. To build multifactor models for cost of injury based on Classification and Regression Trees (CART) machine learning technique. This type of analytical tool gives us many advantages compared to other methods.

We use two sources of data: the Maryland statewide hospital population data for 2017-2022, which includes hospital charges, and the trauma registry data from the R. Adams Cowley Shock Trauma Center in Baltimore, MD for 2016-2021, which includes hospital charges and professional fees. The hospital charges comprise of hospital bed occupancy fees, nursing support, medications and supplies, etc. Professional fees include doctors' fees at the hospital, rehabilitation, and therapy charges after hospital discharge. The injury severity is measured by the Abbreviated Injury Score (AIS) by body region and the maximum AIS when there is more than one injury in the same body region. The second injury severity measure is the Injury Severity Score (ISS). Overall, in the state of Maryland, hospital charges are reimbursed at 87%, while the professional fees are reimbursed at about 30%. We use the hospital charges and professional fee charges to estimate the cost of injury.

Most studies include only hospital costs, which underestimates the total injury cost. We discovered that hospital charges constitute on average 71% of the cost, while professional fees constitute 29%. We develop injury group classification, with four injury cost models based on CART: (i) with injury groups; (ii) without injury groups; (iii) full injury cost; and (iv) professional fees only. Each model identifies specific cost injury groups.

The study finds that injury cost is the highest for polytrauma patients, followed by isolated injuries, and minor injuries. The machine learning analysis points to three unique findings: (1) The best predictor of cost is not a single factor but a combination of factors. (2) Injuries in Lower Extremities and Abdomen are very costly, while injuries to the Head region don't have immediate high cost but may have long-term cost effect. (3) Medical professional fees constitute about a third of injury costs.