2026 ECNP Congress

PS01-1171

Data-driven stratification of perinatal depression risk: integrating metabolic and psychosocial factors to disentangle heterogeneity in perinatal depression

Federica Colombo — M. Godara, C. Verga, A. Cattaneo, S. Entringer, C. Buss, F. Benedetti, B. Vai · IRCSS Ospedale San Raffaele, Unit of Psychiatry and Clinical Psychobiology- Division of Neuroscience, Milan, Italy

2026 ECNP Congress

Background

Peripartum depression (PPD) is a heterogeneous condition whose risk factors remain incompletely understood, limiting early identification and personalized intervention [1]. Beyond psychosocial risk factors, growing evidence points to metabolic alterations involving amino acid, fatty acid, and steroid metabolism pathways as potential markers of individual risk trajectories for PPD [2]. Stratifying women based on multidimensional data could unravel PDD heterogeneity and guide personalized interventions for these women [3].

Aims

Through a data-driven approach, we aim to identify biologically and clinically meaningful subtypes of PPD combing psychosocial and metabolic markers from a population-based cohort study.

Methods

We used data from 369 pregnant woman who participated to the Berlin Birth Cohort Study (BBCS). Depression severity was evaluated both antenatal (20 gestational weeks) and postnatal (6 months after birth) through the Center for Epidemiological Studies-Depression scale (CES-D) and the Edinburgh Postnatal Depression Scale (EPDS), dichotomized based on validated cutoffs. Women were stratified leveraging demographic and psychosocial variables (e.g., lifestyle psychiatric history and adverse childhood experiences) combined with 11 metabolic markers measured in early pregnancy, including amino acids (alanine, pyruvate, phenylalanine, histidine, glycine, glutamine), lipid-related indices, albumin, and glycoprotein A (GlycA). Subtyping analysis was performed with a stability-based relative clustering approach [4], considering K-Means as the clustering algorithm, while ElasticNet logistic regression was trained to predict clusters’ labels. All preprocessing steps (including standardization and data imputation) were embedded into a 100x2 repeated cross-validation to identify the optimal clustering solution (i.e., minimum prediction error), iterating over 2-5 clusters. Statistical significance was evaluated with 10,000 Monte Carlo simulations of a single multivariate Gaussian distribution. Subtypes were tested for differences in input variables and depression rates through linear models and chi-square tests.

Results

Three distinct subtypes were identified with accuracy=60% (p<0.001). Clusters differed significantly for the proportion of women within clinically relevant depression antenatally (χ²=14.94, p<0.001), with Cluster 2 carrying the highest depression burden (50%), followed by Cluster 1 (40.2%) and Cluster 3 (26.2%). This gradient was mirrored by distinct psychosocial and metabolic profiles. The highest-risk subtype (Cluster 2, N=66) showed the greatest rates of childhood trauma across all domains (32.2%-89%) and psychiatric history (79.7%), alongside intermediate metabolic dysfunction (pyruvate: 0.065±0.015 mmol/L; GlycA: 0.920±0.084 mmol/L). Cluster 1 (N=82) was defined by the highest metabolic burden, with elevated body mass index (BMI) (27.7±6.5 kg/m²) and metabolic markers (alanine: 0.351±0.054 mmol/L, pyruvate: 0.089±0.024 mmol/L, glucose: 5.54±0.37 mmol/L, GlycA: 0.968±0.100 mmol/L). Cluster 3 (N=221) represented the most favorable profile, with the lowest metabolic risk (BMI: 21.8±2.9 kg/m², GlycA: 0.838±0.078 mmol/L; glucose: 5.14±0.30 mmol/L; pyruvate: 0.058±0.013 mmol/L), trauma exposure (<10%) and psychiatric history (35.3%).

Conclusions

These findings suggest that, with an entirely data-driven approach, unique subtypes of PPD can be identified based on psychosocial and metabolic variables. This stratification allows the identification of potential at-risk subpopulations in which PPD may arise through distinct mechanistic pathways, which could serve as targets for personalised interventions. If replicated in external cohorts, these findings could advance the understanding of PPD heterogeneity and support the development of precision psychiatry approaches to perinatal mental health.

References

  1. Wang, Z., Liu, J., Shuai, H., Cai, Z., Fu, X., Liu, Y., Xiao, X., Zhang, W., Krabbendam, E., Liu, S., Liu, Z., Li, Z., Yang, B.X., 2021. Mapping global prevalence of depression among postpartum women. Transl. Psychiatry 11 (1), 543.
  2. Li, X., Liu, Y., Wang, Y., Li, X., Liu, X., Gao, M., 2022. Exploration for biomarkers of postpartum depression based on metabolomics: A systematic review. J. Affect. Disord. 316, 144–150.
  3. Galea, L.A., Frokjaer, V.G., 2019. Perinatal depression: embracing variability toward better treatment and outcomes. Neuron 102 (1), 13–16.
  4. Landi, I., Mandelli, V., Lombardo, M.V., 2021. reval: A Python package to determine best clustering solutions with stability-based relative clustering validation. Patterns 2 (4), 100228.