2.49 Transfer Learning of Scanner-Generalization Neural Networks for Predicting General Psychopathology Factor (p factor) in Adolescents Based on Resting-State Functional Connectivity

ObjectivesTransfer learning is a method to utilize pretrained deep neural networks (DNNs) to a new data set. We were motivated to examine whether the transfer learning can be gainfully applied to the prediction of the psychopathology factor (p factor), a single general factor of psychopathology, in Korean adolescents with MDD and healthy controls (HCs) using the DNN pretrained on the adolescent data set from the ABCD study.MethodsWe used 6095 adolescents (9.9 ± 0.6 years, 47.8% girls) from the ABCD study who were scanned using either Siemens or GE scanners across 18 sites. We trained a DNN named scanner-generalization neural networks (SGNN) to predict the p factor using resting-state functional connectivity (RSFC), while simultaneously alleviating scanner effects. We applied confirmatory factor analysis to obtain the p factor using Child Behavior Checklist (CBCL) scores. We evaluated the p factor prediction performance via Pearson’s correlation coefficient (CC) based on leave-one-site-out cross-validation (CV). The trained SGNN, using the ABCD sample, was applied to p factor prediction for the adolescent data set acquired at Seoul National University Hospital (SNUH) (n = 125; 79 MDD, 14.9 ± 1.6 years, 67.1% girls; 46 HC, 14.3 ± 1.4 years, 55.6% girls). We evaluated prediction performance for the SNUH sample based on repeated Monte Carlo CV (n = 20). The SGNN was compared to the kernel-based ridge regression (KRR) and randomly initialized DNN. We analyzed the weight feature of trained SGNN for SNUH to interpret important functional networks (FNs) of the brain.ResultsOur SGNN showed superior performance for the ABCD study (CC = 0.16 ± 0.06; p < .05; two-tailed paired t test) compared to KRR (CC = 0.15 ± 0.07). Pretrained SGNN achieved significant prediction performance for the SNUH sample (0.25 ± 0.15; p < .05) compared to without transfer learning (0.22 ± 0.15) and KRR (0.17 ± 0.15). Default mode network (DMN), cingulo-opercular network (CON), and visual network (VIS) were identified as important FNs to predict the p factor of Korean adolescents.ConclusionsTransfer learning of SGNN enhanced the p factor prediction using RSFC for Korean adolescents with MDD and HCs. Furthermore, alterations within DMN, CON, and VIS may serve as potential biomarkers for the p factor.DDD, ADOL, PSP ObjectivesTransfer learning is a method to utilize pretrained deep neural networks (DNNs) to a new data set. We were motivated to examine whether the transfer learning can be gainfully applied to the prediction of the psychopathology factor (p factor), a single general factor of psychopathology, in Korean adolescents with MDD and healthy controls (HCs) using the DNN pretrained on the adolescent data set from the ABCD study. Transfer learning is a method to utilize pretrained deep neural networks (DNNs) to a new data set. We were motivated to examine whether the transfer learning can be gainfully applied to the prediction of the psychopathology factor (p factor), a single general factor of psychopathology, in Korean adolescents with MDD and healthy controls (HCs) using the DNN pretrained on the adolescent data set from the ABCD study. MethodsWe used 6095 adolescents (9.9 ± 0.6 years, 47.8% girls) from the ABCD study who were scanned using either Siemens or GE scanners across 18 sites. We trained a DNN named scanner-generalization neural networks (SGNN) to predict the p factor using resting-state functional connectivity (RSFC), while simultaneously alleviating scanner effects. We applied confirmatory factor analysis to obtain the p factor using Child Behavior Checklist (CBCL) scores. We evaluated the p factor prediction performance via Pearson’s correlation coefficient (CC) based on leave-one-site-out cross-validation (CV). The trained SGNN, using the ABCD sample, was applied to p factor prediction for the adolescent data set acquired at Seoul National University Hospital (SNUH) (n = 125; 79 MDD, 14.9 ± 1.6 years, 67.1% girls; 46 HC, 14.3 ± 1.4 years, 55.6% girls). We evaluated prediction performance for the SNUH sample based on repeated Monte Carlo CV (n = 20). The SGNN was compared to the kernel-based ridge regression (KRR) and randomly initialized DNN. We analyzed the weight feature of trained SGNN for SNUH to interpret important functional networks (FNs) of the brain. We used 6095 adolescents (9.9 ± 0.6 years, 47.8% girls) from the ABCD study who were scanned using either Siemens or GE scanners across 18 sites. We trained a DNN named scanner-generalization neural networks (SGNN) to predict the p factor using resting-state functional connectivity (RSFC), while simultaneously alleviating scanner effects. We applied confirmatory factor analysis to obtain the p factor using Child Behavior Checklist (CBCL) scores. We evaluated the p factor prediction performance via Pearson’s correlation coefficient (CC) based on leave-one-site-out cross-validation (CV). The trained SGNN, using the ABCD sample, was applied to p factor prediction for the adolescent data set acquired at Seoul National University Hospital (SNUH) (n = 125; 79 MDD, 14.9 ± 1.6 years, 67.1% girls; 46 HC, 14.3 ± 1.4 years, 55.6% girls). We evaluated prediction performance for the SNUH sample based on repeated Monte Carlo CV (n = 20). The SGNN was compared to the kernel-based ridge regression (KRR) and randomly initialized DNN. We analyzed the weight feature of trained SGNN for SNUH to interpret important functional networks (FNs) of the brain. ResultsOur SGNN showed superior performance for the ABCD study (CC = 0.16 ± 0.06; p < .05; two-tailed paired t test) compared to KRR (CC = 0.15 ± 0.07). Pretrained SGNN achieved significant prediction performance for the SNUH sample (0.25 ± 0.15; p < .05) compared to without transfer learning (0.22 ± 0.15) and KRR (0.17 ± 0.15). Default mode network (DMN), cingulo-opercular network (CON), and visual network (VIS) were identified as important FNs to predict the p factor of Korean adolescents. Our SGNN showed superior performance for the ABCD study (CC = 0.16 ± 0.06; p < .05; two-tailed paired t test) compared to KRR (CC = 0.15 ± 0.07). Pretrained SGNN achieved significant prediction performance for the SNUH sample (0.25 ± 0.15; p < .05) compared to without transfer learning (0.22 ± 0.15) and KRR (0.17 ± 0.15). Default mode network (DMN), cingulo-opercular network (CON), and visual network (VIS) were identified as important FNs to predict the p factor of Korean adolescents. ConclusionsTransfer learning of SGNN enhanced the p factor prediction using RSFC for Korean adolescents with MDD and HCs. Furthermore, alterations within DMN, CON, and VIS may serve as potential biomarkers for the p factor.DDD, ADOL, PSP Transfer learning of SGNN enhanced the p factor prediction using RSFC for Korean adolescents with MDD and HCs. Furthermore, alterations within DMN, CON, and VIS may serve as potential biomarkers for the p factor.