In Vivo Raman Spectroscopy Combined With Long Short‐Term Memory Networks for Predicting Transdermal Absorption of Active Cosmetic Ingredients

ABSTRACT Efficient and noninvasive in vivo quantification of the transdermal absorption of active cosmetic ingredients is essential for formulation optimization and efficacy evaluation. This study presents a quantitative model for predicting later‐time Raman spectra and estimating transdermal absorption using Ceramide NP (Cer‐NP) as a representative active ingredient by integrating in vivo confocal Raman spectroscopy (CRS) with deep learning. At each measurement time point, including baseline, 1, and 4 h after application, 600 in vivo Raman spectra were collected across different skin depths. Among them, the 1‐h spectra were used as inputs and the corresponding 4‐h spectra as target outputs to optimize and train a long short‐term memory (LSTM)‐based spectral generation model. The predicted 4‐h spectra were then used to calculate the corresponding transdermal absorption at 4 h, and the prediction performance was evaluated by comparing these values with the transdermal absorption calculated from the experimentally measured 4‐h spectra. Compared with autoregressive integrated moving average (ARIMA) and back propagation neural network (BPNN) models, the LSTM achieved the highest coefficient of determination ( R 2 ) for absorption prediction, despite exhibiting a slightly higher RMSE. Furthermore, a linear regression calibration applied after initial prediction increased R 2 from 0.7225 to 0.8034 (11.20% increase) and reduced RMSE from 0.0394% to 0.0332% (15.74% decrease), thereby enhancing both predictive accuracy and stability. The results demonstrate that deep learning can effectively model spectral changes between two time points based on in vivo Raman measurements, providing a practical framework for rapid evaluation of the transdermal absorption of cosmetic active ingredients, with potential applications in rapid formulation screening and performance assessment.