Artificial intelligence-driven body surface area (BSA) estimation using computed tomography: Comparative evaluation with existing formulae

Total body surface area (BSA) is thought to be a better predictor of metabolic mass than body weight because it is less affected by abnormal adipose mass, and as a result, it is often used for drug dosing in clinical medicine. There are several formulae used for estimating BSA, however, most are based on 2-dimensional and not 3-dimensional BSA measurements or used naïve statistical techniques to create the estimating formulae. This study analyzed 3-dimensional whole-body positron emission tomography/computed tomography studies from 698 patients. A limitation of this dataset was the imbalance in racial distribution. BSA was obtained from the computed tomography component of these studies using TotalSegmentator software. There was a positive correlation between BSA and height (R = 0.49, P < .001), and between BSA and weight (R = 0.83, P < .001). Multivariable ridge regression with 5-fold cross-validation was used to create new formulae to estimate BSA by adjusting for sex, race, height, and weight. The optimal formula to estimate BSA in m², is (Formula presented), where H is height in meters and W is weight in kilograms. Our formulae were significantly more accurate than all other formulae and had a lower mean absolute error and mean squared error than the Dubois (P < .001, P < .001), Mosteller (P < .001, P < .001), Haycock (P < .001, P < .001), Gehan (P < .001, P < .001), Boyd (P < .001, P < .001), Fujimoto (P < .001, P < .001), Takahira (P < .001, P < .001), Shuter and Aslani (P < .001, P < .001), Lipscombe (P < .001, P < .001), and Schlich (P < .001, P < .001) BSA formulae respectively.