Functional Imaging Derived ADHD Biotypes Based on Deep Clustering May Guide Personalized Medication

415 

Abstract Submission 

Aichen Feng¹, Dongmei Zhi², Yuan Feng³, Rongtao Jiang⁴, Zening Fu⁵, Ming Xu⁶, Shan Yu⁷, Michael Stevens⁸, Li Sun³, Vince Calhoun⁹, Jing Sui¹⁰

¹Institute of Automation, Chinese Academy of Sciences, Beijing, China, Beijing, China, ²Beijing Normal University, Beijing, Select a State, ³Peking University Sixth Hospital/Institute of Mental Health, National Clinical Research Center for M, Beijing, China, Beijing, China, ⁴Yale School of Medicine, New Haven, CT, ⁵Tri-institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS) Georgi, Atlanta, GA, ⁶Institute of Automation, Chinese Academy of Sciences, Beijing, Select a State, ⁷Brainnetome Center and National Laboratory of Pattern Recognition, Institute of Automation, Chinese, Beijing, Beijing, ⁸Department of Psychiatry, Yale University School of Medicine, New Haven, CT, United States., New Haven, CT, United States., New Haven, CT, United States., ⁹GSU/GATech/Emory, Decatur, GA, ¹⁰Beijing Normal University, Beijing, China

Aichen Feng  
Institute of Automation, Chinese Academy of Sciences
Beijing, China, Beijing, China

Dongmei Zhi  
Beijing Normal University
Beijing, Select a State

Yuan Feng  
Peking University Sixth Hospital/Institute of Mental Health, National Clinical Research Center for M
Beijing, China, Beijing, China

Rongtao Jiang  
Yale School of Medicine
New Haven, CT

Zening Fu  
Tri-institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS) Georgi
Atlanta, GA

Ming Xu  
Institute of Automation, Chinese Academy of Sciences
Beijing, Select a State

Shan Yu  
Brainnetome Center and National Laboratory of Pattern Recognition, Institute of Automation, Chinese
Beijing, Beijing

Michael Stevens  
Department of Psychiatry, Yale University School of Medicine, New Haven, CT, United States.
New Haven, CT, United States., New Haven, CT, United States.

Li Sun  
Peking University Sixth Hospital/Institute of Mental Health, National Clinical Research Center for M
Beijing, China, Beijing, China

Vince Calhoun  
GSU/GATech/Emory
Decatur, GA

Jing Sui  
Beijing Normal University
Beijing, China

Attention deficit hyperactivity disorder (ADHD) is one prevalent neurodevelopmental disorder with childhood onset, however, there is no clear correspondence established between clinical ADHD subtypes and primary medications. Identifying objective and reliable neuroimaging markers for categorizing ADHD biotypes may lead to more individualized, biotype-guided treatment.

A population graph was first constructed based on functional network connectivity (Fig. 1a) and phenotypic information (age, gender) to build individual mappings; where FNCs serve as the feature of the nodes and the similarities between two subjects, which were abstracted from gender and age serve as edges. Then we applied GCN-BSD to learn embeddings that are both group-discriminative between ADHD and controls, as well as adapted to the clustering constraint through K-Means loss (Fig. 1b). We selected 1069 ADHD patients from the ABCD study as the discovery dataset, identified K=2 for biotype division via the cluster sum of square (CSS) using the elbow method31, and evaluated the clustering performance of 4 popular algorithms, including (1) agglomerative clustering, (2) conventional K-Means, (3) DNN with deep K-Means, and (4) autoencoder GCN with K-Means with GCN-BSD based on Davies-Bouldin Index (DBI) and Calinski-Harabasz Index (CHI)32 (Fig. 1c). As a result, two ADHD biotypes were identified, manifesting with different FNC patterns and distinguishing cognitive abilities. Then we used 130 ADHD and 105 controls collected from Peking University Sixth Hospital as validation dataset to test the generalizability and potential clinical use of the identified biotypes. Interestingly, we found that ADHD biotypes identified in ABCD and PKU showed high similarity and replicability in FNC patterns(Fig. 1d). The most contributing FNCs and clinical records were compared dedicatedly between two biotypes. Specifically, biotype 1 presented milder symptoms while biotype 2 manifested more severe hyperactivity/impulsivity symptoms and worse cognitive levels. Finally, we compared the symptom relief and treatment outcome of two biotypes from 44 out of 130 ADHD patients either treated by MPH or ATX at PKU6 according to our division (Fig. 1e).

Biotype 1 contained more patients, presented milder symptoms, and overrepresented several wildly recognized brain aberrations including frontal gyrus and cerebellum. In contrast, biotype 2 included fewer patients, presented more severe symptoms especially hyperactive/impulsive, and showed greater degrees in regions from DM to SM, as well as the connectivity between the cerebellum and fusiform gyrus. Interestingly, in addition to differences in cognitive performance and hyperactivity/impulsivity symptoms, biotype 1 treated with methylphenidate demonstrated significantly better recovery than biotype 2 treated with atomoxetine (p<0.05, FDR corrected).

Collectively, in this study, we proposed a novel framework, GCN-BSD, that can jointly characterize brain imaging data and phenotypic association and further use this knowledge to guide disease biotype detection. Importantly, the identified two ADHD biotypes exhibit significant group differences in functional networks and multiple cognitive abilities and symptoms, especially in fluid intelligence and hyperactive/impulsive. All the above findings indicate the validation of the frontoparietal circuits to serve as a key signature to ADHD and provide the first evidence for the connection from the cerebellum to the fusiform gyrus to be used as a biomarker in the uncommon subgroup. This study helps move forward from a conventional biotype detection approach to the use of a more flexible deep learning-based analysis.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ¹

fMRI Connectivity and Network Modeling ²

Attention Deficit Disorder

Other - deep learning