2026 ECNP Congress

PS04-4043

Transfer learning of shared structure for robust psychosis classification across cohorts

Wiem Idoudi — D. Popovic, K. Schiltz, H. Dobrowolny, P. Falkai, N. Koutsouleris, J. Steiner · Max Planck Institute of Psychiatry, International Max Planck Research School for Translational Psychiatry IMPRS-TP, München, Germany

2026 ECNP Congress

Generating and validating machine learning models across datasets from diverse cohorts is challenging due to heterogeneity in data distributions, modalities, and measurement protocols[1],[2]. These discrepancies often render models non-applicable across cohorts[1,][2]. Furthermore, harmonisation strategies typically constrain analyses to a reduced common feature space, leading to a substantial loss of information and variance[3]. Transfer learning provides a potential solution by allowing models to incorporate common structures across datasets[4]. This study investigated if cross-modal transformer pretraining enhances downstream prediction and uncovers relationships among diverse datasets.

A cross-modal transformer model was pretrained on the PRONIA cohort (n = 828), comprising young (15–40 years old), minimally medicated individuals in the earliest stages of mental disorders, using diffusion tensor imaging (DTI) and neurocognitive features, primarily derived from the MATRICS battery. The pretraining task aimed to differentiate healthy controls (HC) from individuals with recent onset of psychosis (ROP). Features were preprocessed and grouped into biologically meaningful tokens. The model integrated modality-specific self-attention with cross-attention to learn shared representations across modalities.

The pretrained DTI encoder was applied for transfer learning to an independent German cohort from the Magdeburg area (n = 238; 132 controls (C), 106 schizophrenia (Sz)), comprising older individuals at first-episode and chronic stages of schizophrenia. This cohort included neurocognitive, blood-based, and physical features with a distinct feature structure. Fine-tuning used an adapter-based approach with staged unfreezing. Performance was compared to an identical model trained from scratch using repeated nested cross-validation (10 repeats, 5×5 folds). Statistical analyses comprised paired comparisons, Kolmogorov–Smirnov tests, Wasserstein distance, and Wilcoxon tests.

Pretraining achieved a mean balanced accuracy of 0.769 (± 0.038). In the downstream task, transfer learning enhanced balanced accuracy relative to training from scratch (0.778 ± 0.023 vs. 0.758 ± 0.019), yielding a mean paired improvement of +0.020 (± 0.016), and outperformed the baseline in 32 out of 50 outer folds.

Significantly, transfer learning modified decision-making behavior beyond mere accuracy. The distributions of decision scores exhibited significant differences (Kolmogorov–Smirnov p < 0.001; Wasserstein distance = 0.64), with a higher margin separation (Wilcoxon p = 0.0015). The transfer model demonstrated less overlap around the decision border and enhanced class separation, signifying increased decision stability.

Feature attribution analyses indicated that transfer learning enhanced dependence on biologically stable signals common across datasets, such as inflammatory markers (e.g., IL1ra, MPO, NGAL), metabolic and endocrine markers (e.g., IGFBP2, proinsulin, cortisol), and specific neuro-related factors (e.g., BDNF), while diminishing the impact of less consistent features.

In conclusion, cross-modal transformer pretraining enables effective transfer across heterogeneous cohorts. While improvements in accuracy are modest, the model benefits from both datasets by learning shared representations that enhance decision stability and class separation. These findings suggest that transfer learning captures transferable biological structure across neurobiologically informed datasets. In addition, this approach provides an integrated framework for analysing multi-dimensional, multi-level neurobiological and clinical data across medium to large psychiatric cohorts. Future work should examine whether these transfer-informed features improve prediction of clinically relevant outcomes, such as recovery and remission trajectories, when combined with baseline clinical models.

References

  1. Rajendran S, Pan W, Sabuncu MR, Chen Y, Zhou J, Wang F. Learning across diverse biomedical data modalities and cohorts: challenges and opportunities for innovation. Patterns. 2024;5(2):100913. doi:10.1016/j.patter.2023.100913
  2. Hu HT, Li MD, Lin XX, et al. Addressing data heterogeneity in distributed medical imaging with heterosync learning. Nature Communications. 2025;16:9416. doi:10.1038/s41467-025-64459
  3. Cheng C, Messerschmidt L, Bravo I, et al. A general primer for data harmonization. Scientific Data. 2024;11:152. doi:10.1038/s41597-024-02956-3
  4. Wu X, Manton JH, Aickelin U, Zhu J. On the generalization for transfer learning: an information-theoretic analysis. IEEE Transactions on Information Theory. 2024;70(10):7089–7124. doi:10.1109/TIT.2024.3441574