Deep neural networks predict future overweight/obesity and obesity-related major psychiatric disorde
Poster No:
1407
Submission Type:
Abstract Submission
Authors:
HEEHWAN WANG1, Yoonjung Joo2, Eun-Ji Lee1, Bo Gyeom Kim1, Gakyung Kim1, Jiook Cha3
Institutions:
1Seoul National University, Seoul, Other, 2Sungkyunkwan University, Seoul, Other, 3Seoul National University, Seoul, Seoul
First Author:
Co-Author(s):
Introduction:
Childhood overweight/obesity is a global health issue affecting life-long risks for chronic diseases and mental health. Although the role of the brain in the development of overweight/obesity is well recognized, the potential of the brain as a predictor for future childhood overweight/obesity remains unclear. We hypothesize that the brain structure can accurately predict whether children become overweight/obese within 1 and 2 years via deep neural networks.
Methods:
We used brain structural MRI data and behavioral assessment data of 11,316 children (ages 9 to 10 years) from the Adolescent Brain Cognitive Development (ABCD) study. We calculated Body Mass Index standard deviation score (BMI-sds) of children based on the Grwoth Chart released by the Center for Disease Control and Prevention, used for an index of weight status at the baseline year. We respectively assigned children who were in the normal weight status at the baseline year but became overweight/obese and excessively gained weight (the change of BMI-sds > 0.2) within 1-year follow-up and 2-year follow-up to become overweight within 1 year and become overweight within 2 years group. If children did not satisfy those requirements, they were assigned to still normal group. We employed transfer learning approach to predict future overweight/obesity. We pre-trained deep neural networks to predict BMI-sds at the baseline year from brain structure MRI at the baseline year and respectively fine-tuned them to predict whether children become overweight/obesity within 1-year follow-up and 2-year follow-up based on brain structure MRI at the baseline year. Furthermore, we employed explainable AI (XAI) algorithms to investigate brain structures attributing to deep neural networks' predictions, i.e., the combination of SmoothGrad and Integrated Gradients. Using XAI-driven attribution scores, we performed clustering analysis and statistical analyses to test the difference in polygenic risk and the development of behavioral problems between children who became overweight/obese. To investigate the utility of deep neural networks and transfer learning approach, we fine-tuned those pre-trained model, learned to predict BMI-sds at the baseline year from brain structure MRI at the baseline year, to predict the diagnoses of major psychiatric disorders which are frequently comorbid with obesity, such as Major Depressive Disorder and Anxiety disorders.
Results:
Our results showed that deep neural networks accurately predict current weight status (R2 = 0.49 土 0.01; Mean Squared Error = 1.97 土 0.06) and transfer learning approach significantly improved future overweight/obesity prediction within 1 year (ACC = 0.64 土 0.03; AUROC = 0.70 土 0.02) and 2 years (ACC = 0.61 土 0.01; AUROC = 0.66 土 0.03). When investigating brain structures attributing to the prediction of current weight status and future overweight/obesity, we found the brain stem, the pituitary gland, the cerebellum, various temporal lobe regions, and the amygdala were important for deep neural network's prediction. Additonally, XAI-driven the amygdala subnuclei attribution scores for future overweight/obesity revealed individual differences in the development of CBCL anxiety problem (Cohen's d = 0.14, FDR corrected p value < 0.05). One step further, we investigated the potential of pre-trained deep neural networks for current weight status by demonstrating significant improvements in classifying current diagnoses of major psychiatric disorders frequently comorbid with obesity, such as Major Depressive Disorder (ACC = 0.60 土 0.04; AUROC = 0.67 土 0.05).
·Model performance for predicting future overweight/obesity within follow-up years from brain data at the baseline year
·Model performance for predicting the diagnoses of obesity-related major psychiatric disorders at the baseline year from brain data at baseline year
Conclusions:
We showed the utility of deep neural networks and transfer learning approach in early detection of childhood overweight/obesity. Using XAI algorithms, we revealed the individual differences in the development of anxiety problems. One step further, we also demonstrated the potential of our novel transfer learning approach in psychiatric research.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 2
Psychiatric (eg. Depression, Anxiety, Schizophrenia)
Genetics:
Genetics Other
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Neuroanatomy Other
Keywords:
Data analysis
Development
Machine Learning
MRI
Open Data
Open-Source Code
Psychiatric Disorders
1|2Indicates the priority used for review
Provide references using author date format
Freemark, M (2010). Pediatric Obesity: Etiology, Pathogenesis, and Treatment. (Springer Science & Business Media)