Publicación: PREDICTION OF GESTATIONAL DIABETES MELLITUS WITH MACHINE LEARNING TECHNIQUES: A COMPARISON BETWEEN NEAR-INFRARED SPECTRA AND MATERNAL DATA BASED-MODELS
Fecha
2022
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
PLACENTA
Resumen
OBJECTIVE: TO COMPARE THE PERFORMANCE OF NEAR-INFRARED (NIR) SPECTRA AND MATERNAL DATA BASED-MACHINE LEARNING (ML) MODELS FOR GESTATIONAL DIABETES MELLITUS (GDM) PREDICTION IN CHILEAN PREGNANT WOMEN.
METHODOLOGY: PREGNANT WOMEN WITH ? 12 GESTATIONAL WEEKS AND WITHOUT PREGESTATIONAL DIABETES WERE RECRUITED IN CONCEPCION, CHILE. GDM DIAGNOSIS WAS PERFORMED AT 24-28 GESTATIONAL WEEKS, WITH FASTING GLYCEMIA 100-125 MG/DL OR POST-LOAD GLYCEMIA (75 G, 2 H) ? 140 MG/DL. DURING THE FIRST TRIMESTER OF PREGNANCY, SERA WERE COLLECTED, AND 63 CLINICAL AND BIOCHEMICAL MATERNAL VARIABLES WERE REGISTERED. FOR EACH SERUM SAMPLE, 5 NIR SPECTRA (RANGE 4000-10500 CM-1, RESOLUTION 4 CM-1) WERE RECORDED AND AVERAGED. FOR NIR SPECTRA, 80 DIFFERENT COMBINATIONS OF TRANSFORMATIONS (SAVITZKY-GOLAY SMOOTHING OR FIRST/SECOND DERIVATIVE WITH VARYING FILTER WIDTH, STANDARD NORMAL VARIATE SCATTERING CORRECTION, AUTOMATIC WEIGHTED LEAST SQUARES BASELINE CORRECTION, 2-NORM NORMALIZATION) WERE TESTED. NIR AND MATERNAL DATA WERE PREPROCESSED BY MEAN CENTERING AND AUTOSCALING, RESPECTIVELY. FOR GDM PREDICTION, THE CLASSIFICATION ML TECHNIQUE PARTIAL LEAST SQUARES-DISCRIMINANT ANALYSIS (PLS-DA) WAS EMPLOYED. EVERY MODEL WAS SUBJECTED TO LEAVE-ONE-OUT CROSS-VALIDATION.
RESULTS: THE BEST NIR DATA-BASED MODEL WAS OBTAINED WITH SAVITZKY-GOLAY SMOOTHING (FILTER WIDTH 15, POLYNOMIAL ORDER 2) AND 2-NORM NORMALIZATION. IT ACHIEVED A CROSS-VALIDATION NON-ERROR RATE (CV-NER) OF 68% AND A CROSS-VALIDATION AREA UNDER THE RECEIVER OPERATING CHARACTERISTIC CURVE (CV-AUC) OF 0.669. THE MATERNAL DATA BASED-MODEL ACHIEVED A CV-NER OF 81% AND A CV-AUC OF 0.852.
CONCLUSIONS: CLINICAL AND BIOCHEMICAL MATERNAL PARAMETERS ARE MORE USEFUL TO PREDICT GDM IN CHILEAN PREGNANT WOMEN THAN NIR SPECTRAL DATA.
