Behavior Score Prediction in Resting-State Functional MRI by Deep State Space Modeling

Early clinical assessment of Alzheimer’s disease relies on behavior scores that measure a subject’s language, memory, and cognitive skills. On the medical imaging side, functional magnetic resonance imaging has provided invaluable insights into the neural pathways underlying Alzheimer’s disease. While prior studies have used resting-state functional MRI by extracting functional connectivity matrices, these approaches neglect the temporal dynamics inherent in functional data. In this work, we present a deep state space modeling framework that directly leverages the blood-oxygenation-level-dependent time series to learn a sparse collection of brain regions to predict behavior scores. Our model extracts temporal features that encapsulate nuanced patterns of intrinsic brain activity, thereby enhancing predictive performance compared to traditional connectivity methods. We identify specific brain regions that are most predictive of cognitive impairment through experiments on data provided by the Michigan Alzheimer’s Disease Research Center, providing new insights into the neural substrates of early Alzheimer’s pathology. These findings have important implications for the possible development of risk monitoring and intervention strategies in Alzheimer’s disease.

💡 Research Summary

Alzheimer’s disease (AD) diagnosis increasingly relies on brief cognitive assessments such as the Montreal Cognitive Assessment (MoCA) and its sub‑domains (memory, language). While resting‑state functional MRI (rs‑fMRI) has been used to predict these behavior scores, prior work has almost exclusively reduced the data to static functional connectivity matrices, thereby discarding the rich temporal dynamics inherent in the BOLD signal. This paper introduces a deep state‑space modeling (D‑SSM) framework that directly consumes the full BOLD time series from each brain region, learns a sparse set of predictive regions, and extracts temporally informed features for behavior‑score regression.

The authors used the Michigan Alzheimer’s Disease Research Center (MADRC) cohort (N = 281; 203 cognitively normal, 53 amnestic MCI, 25 AD dementia). Each subject underwent a 3 T multi‑band rs‑fMRI scan (TR = 0.8 s, 570 volumes). After rigorous preprocessing (BIDS conversion, fmriprep, CONN toolbox denoising, band‑pass filtering, and parcellation into 272 regions of interest—264 from the Power atlas plus eight manually added hippocampal/amygdalar ROIs), the data for each subject is a matrix (X_i\in\mathbb{R}^{570\times272}).

Two conventional baselines were implemented: (1) Pearson correlation functional connectivity matrices (FCM) vectorized and fed to kernel ridge regression (KRR) with an RBF kernel, and (2) subject‑wise independent component analysis (I‑ICA) followed by the same KRR. Both yielded modest correlations with MoCA (R ≈ 0.07–0.15), confirming that collapsing the temporal dimension limits predictive power.

The proposed D‑SSM treats the BOLD series as observations generated from latent states (z_t). A recurrent encoder (GRU/LSTM) maps the sequence (x_t) to a hidden representation, which is then pooled and passed through fully‑connected layers to predict the three scores simultaneously. Crucially, an L1‑based sparsity penalty is applied to the region‑wise input weights, encouraging the model to select a compact subset of brain regions. The loss combines mean‑squared error with the sparsity term, and the network is trained with Adam using five‑fold cross‑validation and grid‑searched hyperparameters.

Results show that D‑SSM achieves a mean Pearson correlation of ≈ 0.42 ± 0.03 for MoCA, substantially outperforming the FCM‑KRR (R ≈ 0.15) and I‑ICA‑KRR (R ≈ 0.07) baselines. Similar gains are observed for the memory (R ≈ 0.38) and language (R ≈ 0.35) sub‑scores. Ablation studies reveal that removing the sparsity constraint reduces performance by ~10 %, while replacing the recurrent encoder with a simple temporal average drops accuracy by ~15 %, underscoring the importance of both temporal modeling and region selection.

The sparsity analysis highlights a set of regions most influential for prediction: core nodes of the default mode network (medial prefrontal, posterior cingulate, lateral parietal cortices), bilateral hippocampi, amygdalae, and left ventrolateral prefrontal cortex. These areas are well‑known to be affected early in AD, providing biological plausibility to the model’s selections and suggesting potential biomarkers for early intervention.

Limitations include reliance on a single institutional dataset, which may restrict generalizability, and the use of ROI‑based parcellation that could miss finer‑grained spatial patterns. The authors propose future extensions such as graph neural networks to model inter‑regional interactions directly, Bayesian state‑space formulations for uncertainty quantification, and attention mechanisms to further improve interpretability. Validation on multi‑site cohorts and integration with electronic health records are also suggested to move toward clinical deployment.

In summary, this work demonstrates that deep state‑space models leveraging the full temporal richness of rs‑fMRI can markedly improve the prediction of cognitive behavior scores in Alzheimer’s disease. By simultaneously delivering higher accuracy and interpretable region‑level insights, the approach offers a promising avenue for developing non‑invasive, imaging‑based risk monitoring tools and for informing targeted neuromodulation strategies in early AD.