Multi-domain and Uni-domain Fusion for domain-generalizable fMRI-based phenotypic prediction

1392 

Abstract Submission 

Pansheng Chen¹, Lijun An¹, Naren Wulan¹, Chen Zhang¹, Shaoshi Zhang¹, Leon Ooi¹, Ru Kong¹, Jianxiao Wu², Sidhant Chopra³, Danilo Bzdok⁴, Simon Eickhoff⁵, Avram Holmes⁶, B. T. Thomas Yeo¹

¹National University of Singapore, Singapore, Singapore, ²Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich, Jülich, Germany, ³Yale University, New Haven, CT, ⁴McConnell Brain Imaging Centre (BIC), Montreal Neurol, McGill Universityogical Institute (MNI), Montreal, Quebec, ⁵Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf, Düsseldorf, North Rhine–Westphalia Land, ⁶Department of Psychiatry, Brain Health Institute, Rutgers University, Piscataway, NJ

Pansheng Chen  
National University of Singapore
Singapore, Singapore

Lijun An  
National University of Singapore
Singapore, Singapore

Naren Wulan  
National University of Singapore
Singapore, Singapore

Chen Zhang  
National University of Singapore
Singapore, Singapore

Shaoshi Zhang  
National University of Singapore
Singapore, Singapore

Leon Ooi  
National University of Singapore
Singapore, Singapore

Ruby Kong  
National University of Singapore
Singapore, Singapore

Jianxiao Wu  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich
Jülich, Germany

Sidhant Chopra  
Yale University
New Haven, CT

Danilo Bzdok  
McConnell Brain Imaging Centre (BIC), Montreal Neurol, McGill Universityogical Institute (MNI)
Montreal, Quebec

Simon Eickhoff  
Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf
Düsseldorf, North Rhine–Westphalia Land

Avram Holmes  
Department of Psychiatry, Brain Health Institute, Rutgers University
Piscataway, NJ

B. T. Thomas Yeo  
National University of Singapore
Singapore, Singapore

Resting-state functional connectivity (RSFC) is widely used to predict phenotypes in individuals[1,2]. Due to unavoidable small sample size issues[3-5], recent work has sought to translate models trained from large-sized neuroimaging datasets to predict phenotypes on small target datasets[5,6]. However, predictive models may fail to generalize to new datasets due to differences in population, data collection, and processing across datasets[7,8]. Here we proposed a method named Multi-domain and Uni-domain Fusion (MUF) to enhance model generalizability, which outperformed 4 strong baselines on 6 target datasets.

We used 419 X 419 RSFC matrices to predict phenotypes[5] from 7 datasets: UK Biobank[10,11], ABCD[12], GSP[13], HBN[14], eNKI[15], HCP-YA[16] and HCP-Aging[17]. We did a leave-one-dataset-out test on each dataset, except UK Biobank as we need the largest dataset and for training. Each dataset was iteratively used as the target dataset and all others were used as source datasets (Fig 1A). Predictive models trained from source datasets were adapted to K participants (K-shot) in the target dataset to predict target phenotypes. The adapted models were evaluated in the remaining test participants. This procedure was repeated 100 times for stability. We compared MUF against 4 baselines (Fig 1B): classical Kernel Ridge Regression (KRR), meta-matching with stacking[5], and our previous work: meta-matching with dataset stacking and multilayer meta-matching[9].

Classical KRR models were trained on K-shot to predict target phenotypes[5]. For meta-matching with stacking[5], a feedforward deep neural network (DNN) was trained on UK Biobank to predict 67 phenotypes. The base DNN was applied to K-shot and DNN predictions were used as features to train a KRR model on K-shot to predict target phenotypes (i.e. stacking). Meta-matching with dataset stacking[9] extends meta-matching with stacking by training separate KRR/DNN models for each source dataset and then performing stacking on K-shot to predict target phenotypes. To resolve the sample size imbalance across source datasets, multilayer meta-matching[9] gradually applied stacking from larger source datasets to smaller datasets to boost prediction accuracies. These predictive models then underwent another round of stacking using K-shot to predict target phenotypes.

Of note, the above 4 baselines trained base models independently on each dataset. Fig 1C shows our new proposed MUF method, which combines both cross-domain and intra-domain training to learn domain-general and domain-specific information. On the basis of multilayer meta-matching, we also employed a multi-domain strategy that trains a joint multi-task DNN using all source datasets together, enabling it to offset site differences to some extent and thus generalize better on unseen datasets. Predictions from the above joint DNN were concatenated together with predictions from the multilayer meta-matching method and then used as features of the stacking model to predict target phenotypes.

Fig 2 shows the prediction accuracy (Pearson's correlation) in 6 target datasets. All reported p values survived a false discovery rate of q < 0.05.

We found that original meta-matching with stacking[5] was not better than classical KRR on some datasets (e.g., HBN), but our proposed methods improved it by incorporating more diverse source datasets. In all test datasets, MUF consistently outperformed all other baselines, indicating that MUF can better generalize from different source datasets and is more robust on new target datasets. Using Wilcoxon signed-rank test, MUF was better than meta-matching with dataset stacking[9] (p < 0.02 for all K) and multilayer meta-matching[9] (p < 0.02 for K > 20).

We propose a method using both multi-domain and uni-domain training, to translate phenotypic prediction models from multiple source datasets to small-sized target datasets. We found that our MUF performed the best on 6 test datasets.

Classification and Predictive Modeling ¹

Connectivity (eg. functional, effective, structural) ²

Methods Development

Computational Neuroscience

FUNCTIONAL MRI

Machine Learning

Modeling