Preliminary regression-based feature selection for optimizing spatial frequency domain imaging in burn-severity assessment

We identified key spatial frequency domain imaging (SFDI) features for assessing burn wound severity. We built upon our previous two-step regression framework, using SFDI-calibrated reflectance ( $R_d$ ) maps to predict collagen denaturation depth (CDD) and healing outcome of burn wounds. We used feature selection methods to identify key parameters from a large data set (40 features, including eight wavelengths, each at five spatial frequencies).

We aimed to identify a reduced set of key SFDI features (i.e., pairs of spatial frequencies and wavelengths) for assessing burn wound severity.

We utilized a graded-severity porcine burn model ( $n=4$ ) with eight burn contact durations (5 to 40 s). Multiple linear regression (MLR) was employed to predict CDD from $R_d$ obtained within 24 h post-burn. Subsequently, logistic regression (LogR) was used to predict day-28 healing outcomes based on the predicted CDD. In this study, we fielded three feature selection algorithms (i.e., standardized ordinary least squares, least absolute shrinkage and selection operator, and minimum redundancy maximum relevance) to identify key SFDI features in the MLR model. The resulting prediction then dictates the LogR outputs. We assessed the models for accuracy and fit performance against those of the 40-feature original model.

A subset of 13 LASSO features matched the predictive performance of the 40-feature baseline (RMSE 98 versus $105\mu m$ ; within fold-to-fold noise), and a minimum set of eight LASSO features (minimum features with acceptable errors) closely approximated it. Across all three methods, the consistently top-ranked features were planar ( $0{\mathrm{mm}}^{-1}$ ) reflectance at 526, 591, and 621 nm.

Using feature selection algorithms, we identified preliminary candidate features that may be suitable for developing a future SFDI system to assess burn wound severity with reduced complexity. This has the potential to enable a more compact design compared with the current research-grade SFDI system.