Deep Learning Models for Predicting fMRI Time Series in Autistic Adults
Poster No:
321
Submission Type:
Abstract Submission
Authors:
Moses Sokunbi1, Oumayma Soula2
Institutions:
1De Montfort University, Leicester, United Kingdom, 2University ElManar, Tunis and University of Sfax, Sfax, Tunisia
First Author:
Co-Author:
Introduction:
Understanding and modelling the complex dynamics of functional magnetic resonance imaging (fMRI) time series is crucial for advancing brain research in autism spectrum disorder (ASD). ASD is characterised by atypical functional connectivity (Tyszka et. Al. 2014) but capturing and interpreting these temporal patterns remains a challenge. Deep learning models such as long short-term memory recurrent neural networks (LSTM-RNNs) are well-suited for sequential data, such as fMRI time series, as they effectively capture long-term dependencies (Hochreiter and Schmidhuber, 1997). This study aims to explore the use of LSTM-RNNs to predict the blood-oxygen-level-dependent (BOLD) fMRI signal time series in autistic adults and evaluates the accuracy and reliability of these predictions. We hypothesise that the LSTM-RNN model will predict the BOLD fMRI time series in autistic adults with high accuracy and temporal coherence.
Methods:
Nineteen high-functioning adults with an ASD [15 male, mean age (27.40 ± 10.3)] were extracted from the California Institute of Technology's (Caltech) resting state dataset made publicly available in the Autism Brain Imaging Data Exchange I (ABIDE I) project. Diagnosis was confirmed by DSM-IV-TR. Ethical approval was given by the Human Subjects Protection Committee of Caltech. Resting-state fMRI data and T1 weighted images were acquired on a 3 Tesla Magnetom Trio (Siemens Medical Solutions, NJ, USA) with an 8-channel phased array head receive coil and body coil transmission. The preprocessing steps included discarding the first 5 volumes for signal conditioning, realignment, segmentation, co-registration and band-pass filtering (0.008 – 0.1Hz) to reduce noise and low-frequency drifts.
We employed an LSTM-RNN model with 128 LSTM layers to predict the next time point of the BOLD signal. The fMRI time-series was partitioned into a "training dataset" which included 90% of the whole time-series and a "test dataset",which was the remaining 10% of the time-series. The "training datasets" was used to train the developed LSTM using a mean squared error (MSE) loss function. Predictions was made to forecast future time-series using the "test datasets". To evaluate the accuracy for each "test dataset", the root mean squared error (RMSE) was calculated. Statistical analyses (e.g., paired t-tests and correlation) were used to compare observed and predicted time series to assess model performance.
We employed an LSTM-RNN model with 128 LSTM layers to predict the next time point of the BOLD signal. The fMRI time-series was partitioned into a "training dataset" which included 90% of the whole time-series and a "test dataset",which was the remaining 10% of the time-series. The "training datasets" was used to train the developed LSTM using a mean squared error (MSE) loss function. Predictions was made to forecast future time-series using the "test datasets". To evaluate the accuracy for each "test dataset", the root mean squared error (RMSE) was calculated. Statistical analyses (e.g., paired t-tests and correlation) were used to compare observed and predicted time series to assess model performance.
Results:
The LSTM-RNN successfully learned temporal dependencies within the BOLD fMRI signal (see Figure 1), achieving a mean RMSE of 0.017± 0.008. A paired t-test comparing the observed and predicted time series revealed no significant difference in mean signal amplitude (t = 1.395, p = 0.18), suggesting that the model adequately captures the observed BOLD dynamics.
Additionally, a significantly (p < 0.001) positive (r=1) paired temporal correlation between the observed and predicted time series was computed. This indicates that the predicted time series strongly mirrored the temporal structure of the observed signals. Visual inspection of observed versus predicted time series (see Figure 2) highlighted the capacity of the LSTM-RNN to capture key signal patterns, particularly during periods of high signal variance. These results indicate that the LSTM layers are well-suited to model complex, non-linear dynamics in neural data and effectively leveraged long-term dependencies to enhance prediction accuracy.
Additionally, a significantly (p < 0.001) positive (r=1) paired temporal correlation between the observed and predicted time series was computed. This indicates that the predicted time series strongly mirrored the temporal structure of the observed signals. Visual inspection of observed versus predicted time series (see Figure 2) highlighted the capacity of the LSTM-RNN to capture key signal patterns, particularly during periods of high signal variance. These results indicate that the LSTM layers are well-suited to model complex, non-linear dynamics in neural data and effectively leveraged long-term dependencies to enhance prediction accuracy.
·Figure 1: LSTM-RNN Training progress for a representative autistic adult (age: 22.8 years)
·Figure 2: Observed time series and Predicted time series of a representative autistic adult (age: 22.8 years) with an RMSE of 0.0134
Conclusions:
This study demonstrates the utility of LSTM-RNNs in predicting BOLD fMRI signal time series in autistic adults, offering a powerful framework for analysing the temporal dynamics of neural activity in ASD. Future work will expand this approach to larger cohorts and incorporate advanced methodologies such as entropy and connectivity metrics to the LSTM-RNN model. This work underscores the potential of deep learning in advancing computational neuroimaging and improving our understanding of neurodevelopmental disorders.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Methods Development
Task-Independent and Resting-State Analysis
Novel Imaging Acquisition Methods:
BOLD fMRI
Keywords:
ADULTS
Autism
Computational Neuroscience
Computing
Data analysis
FUNCTIONAL MRI
Machine Learning
MRI
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2. Tyszka, J.M., Kennedy, D.P., Paul, L.K., Adolphs, R. (2014). Largely typical patterns of resting-state functional connectivity in high-functioning adults with autism. Cereb Cortex. 24(7):1894-905.