Method comparisons for differentiation of Schizophrenia and Bipolar based on rs-fMRI Intrinsic and Functional Networks

Technical Summary: Method Comparisons for Differentiation of Schizophrenia and Bipolar Disorder based on rs-fMRI

Problem Statement
Schizophrenia (SZ) and Bipolar Disorder (BP) with psychosis are distinct clinical entities that share significant overlapping symptoms, genetic risk factors, and cognitive deficits. This overlap creates substantial challenges for accurate differential diagnosis, where misdiagnosis can lead to ineffective or exacerbating treatments (e.g., antipsychotics worsening mania in BP). While resting-state functional Magnetic Resonance Imaging (rs-fMRI) offers a promising avenue for identifying objective biomarkers, the field lacks consensus on the most effective representation of brain connectivity for distinguishing these disorders. Specifically, it remains unclear whether intrinsic connectivity network (ICN) temporal profiles or various functional network connectivity (FNC) approaches (static, dynamic, and high-order) provide superior discriminative power.

Methodology
The study evaluated multiple rs-fMRI representations on a cohort of 786 subjects (371 for training, 315 for a held-out test set), comprising 288 SZ and 183 BP patients. The data consisted of 105 intrinsic connectivity networks (ICNs) derived from rs-fMRI using a multi-scale spatially constrained Independent Component Analysis (ICA) template. The methodology was divided into two primary streams:

1. ICN Temporal Profile Analysis:

   • Preprocessing: ICN time courses were zero-padded to a uniform length of 234 samples. Signals were processed via three bandpass filters (Low: 0.01–0.15 Hz, Mid: 0.15–0.35 Hz, High: 0.35–0.5 Hz).
   • Time-Frequency Representations: Spectrograms were generated using Short-Time Fourier Transform (STFT) with a Tukey window, and scalograms were generated using Continuous Wavelet Transform (CWT) with a Morlet wavelet.
   • Classification: Deep learning architectures were employed. A 1D Convolutional Neural Network (CNN) was trained on raw and filtered time courses. A simplified 3D CNN was applied to the volumetric spectrogram and scalogram features to manage computational constraints.

2. Functional Network Connectivity (FNC) Analysis:

   • Feature Extraction: FNC matrices were computed as Pearson correlations between pairs of ICN time courses.
   • Representations:
     • Static FNC (sFNC): Computed over the full duration (and a standardized shortened length of 100 samples for comparison).
     • Dynamic FNC (dFNC): Computed using sliding windows with varying lengths ($L$) and step sizes ($s$).
     • High-Order FNC (hFNC): Derived by computing correlations between the correlation time series of dFNC clusters (using K-means with varying cluster counts $U$).
   • Classification: Seven classical machine learning algorithms (SVM, Random Forest, Logistic Regression, KNN, Naïve Bayes, Decision Tree, LDA) were trained on flattened connectivity features. Logistic Regression was specifically used for the high-dimensional sFNC, dFNC, and hFNC features derived from shortened time courses.

Key Results
The evaluation, primarily measured by the Area Under the Curve (AUC) on the held-out test set, yielded the following findings:

• ICN Temporal Profiles: The 1D CNN trained on raw ICN time courses achieved the highest overall performance with an AUC of 0.67. This outperformed models trained on band-limited signals (Mid-frequency: 0.54, Low: 0.546, High: 0.53) and 3D CNNs applied to spectrograms (AUC 0.49) and scalograms (AUC 0.53).
• Connectivity-Based Methods: Among connectivity approaches, Static FNC (sFNC) generally outperformed dynamic and high-order representations.
  • sFNC achieved an AUC of 0.570.
  • dFNC configurations ranged from 0.548 to 0.569, showing no clear advantage over static methods.
  • hFNC configurations performed the poorest, with AUCs ranging from 0.436 to 0.565, showing no consistent trend relative to parameter choices (window length, step size, cluster count).
• Classical ML Performance: Among the seven classifiers applied to full FNC features, Logistic Regression achieved the highest ROC-AUC (0.640), followed by SVM (0.632) and LDA (0.620). Random Forest achieved the highest raw accuracy (0.68) but exhibited a very low F1 score (~0.20), indicating a bias toward the majority class (SZ) due to class imbalance.

Significance and Claims
The paper posits that while the classification performance remains modest, the results offer critical insights into the stability of neuroimaging biomarkers for psychosis differentiation:

1. Superiority of Raw Temporal Profiles: The finding that raw ICN temporal profiles processed by a 1D CNN outperform complex time-frequency representations suggests that diagnostically relevant information is distributed across the full frequency spectrum. Decomposing signals into narrow bands or complex 3D structures may discard complementary cross-band dynamics or introduce overfitting risks given the limited cohort size.
2. Stability of Low-Order Connectivity: The consistent outperformance of static functional connectivity over dynamic and high-order representations challenges the intuition that increased temporal complexity yields better diagnostic utility. The authors suggest that the standardization of time course lengths (to 100 samples) for methodological comparability likely reduced the temporal resolution necessary for dFNC and hFNC to capture meaningful fluctuations.
3. Challenges in Differentiation: The modest overall performance underscores the inherent difficulty of differentiating SZ and BP using rs-fMRI alone, given their substantial genetic, neurobiological, and symptomatic overlap. The study emphasizes that increased representational complexity does not necessarily translate to improved generalization in this context.
4. Methodological Implications: The results highlight the importance of acquisition duration, noting that sFNC computed from full-length time courses in previous work yielded better scores than the shortened versions used here. The study also flags the impact of class imbalance on metrics like F1 scores and accuracy, suggesting that future work must address this to ensure reported metrics reflect genuine discriminative ability.

In conclusion, the study contributes to the ongoing effort to develop reproducible neuroimaging biomarkers by demonstrating that simple, low-order representations (raw temporal profiles and static connectivity) currently offer the most robust, albeit still limited, discriminative power for distinguishing schizophrenia from bipolar disorder.