Detection and Quantification of Caffeine in Creatine Supplements Using Mid-Infrared Spectroscopy and AI Algorithms
Name: MARIA CLARA DA CRUZ PIRES
Publication date: 03/12/2025
Examining board:
Name |
Role |
|---|---|
| DIEGO FORTES DE SOUZA SALGUEIRO | Examinador Externo |
| FABIANO KENJI HARAGUCHI | Presidente |
| JACKLINE FREITAS BRILHANTE DE SAO JOSE | Examinador Interno |
Summary: Due to the economic importance and popularity of dietary supplements (DS), combined with weak regulatory oversight, DS have become susceptible to fraud, particularly adulteration, reinforcing the need for alternative methods to assess their quality and safety. This study aimed to develop machine learning algorithms for the detection and quantification of caffeine (CAF) in creatine (CRE) supplements using Mid-Infrared Spectroscopy associated with Artificial Intelligence (AI) models. Five commercial brands of CRE were adulterated with CAF (2–20%), totaling 50 samples. The mixtures were analyzed using a Bruker® ALPHA II FTIR spectrometer (4 cm¹ resolution, 32 scans, 4000–400 cm¹) in triplicate, totaling 168 spectra, which were processed in Orange Data Mining® (v. 3.38.1). The data were divided into calibration (CAL = 70%) and prediction (PRED = 30%) sets and assigned to class 0 (n = 18) for pure samples and class 1 (n = 150) for adulterated ones. The following multivariate analyses were applied: Principal Component Analysis (PCA), Support Vector Machine (SVM), and Partial Least Squares (PLS). Two additional blind prediction tests were performed: one with three known CRE brands (18 adulteration levels, 0–20%) and another with one CRE brand unknown to the models (six adulterations), both prepared by independent researchers. The PCA of pure CRE and CAF samples showed explained variance (EV) of 98.5% (PC1 = 97.4%; PC2 = 1.1%), while that of all samples showed EV = 86.75% (PC1 = 68.61%; PC2 = 18.14%). In the test, the SVM achieved Sensitivity (SEN) = 100% and Specificity (SPEC) = 75%. The PLS showed a Coefficient of Determination (R²) = 0.75 and a Root Mean Square Error of Prediction (RMSEp) = 3.06%. The limits of detection (LoD) and quantification (LoQ) were 0.55% and 1.82%, respectively. In the first blind test, the SVM reached 100% SEN starting at 1.71% adulteration; the PLS, for samples above the LoQ, obtained R² = 0.67 and RMSE = 2.84%. In the second blind test, SEN was 80%, and the PLS achieved R² = 0.73 and RMSE = 2.11%. The study demonstrated that the AI algorithms were effective in detecting and quantifying CAF in CRE supplements, representing an applicable and scalable tool for verifying their quality and authenticity.
