Risk stratification of 12-month suicidal risk in bipolar disorder with three different clustering algorithms

Introduction

The lack of validated biomarkers represents a significant obstacle in the clinical prediction of suicidal behaviors among individuals with bipolar disorder (BD) [1]. Developing robust risk stratification models to stratify subjects by individual vulnerability could substantially enhance clinical decision-making.

To address the inherent heterogeneity within psychiatric diagnoses,unsupervised and semi-supervised machine learning (ML) techniques are valuable tools for identifying latent, homogeneous subgroups [2];.

Aims

The first aim of the study is the definition of risk subgroups among individuals with BDbased on neuroanatomical features (grey matter (GM) volumes), by uncovering homogenous and stable clusters that correlate with 12-months suicide risk.

Second, we assessed the clinical differences between the clusters and the stability of the clusters.

Methods

We enrolled 163 individuals with BD (53% females, mean age 44.8, SD 15.3) during an acute admission.

At baseline, clinical and demographic data were obtained; moreover, patients underwent MRI scan and were then followed-up for 12 months.

GM features from 136 brain areas of the Neuromorphometrics atlas were statistically adjusted for key covariates including age at MRI, sex, total intracranial volume (TIV), and MRI scanner group.

To identify neuroanatomical patterns linked to suicidal status, a Partial Least Squares Discriminant Analysis (PLS-DA) was performed. A six-component solution was estimated, yielding latent scores that best capture the variance separating suicidal from non-suicidal individuals.

Unsupervised clustering was performed on PLS components to investigate whether individuals formed neuroanatomically distinct subgroups and whether these groups reflected different levels of suicide risk. We tested three different algorithms, namely k-means, PAM (Partitioning Around Medoids), and spectral clustering. Finally, the 3 algorithms were compared to evaluate the robustness of the high-risk subgroup, with Jaccard similarity indices.

Results

During the 12-months observation, 10 patients (6.13%) attempted suicide.

PLS1 showed a strong and consistent association with suicidal behaviour. The most associated features were in the frontal, orbital and cerebellar cortices, in line with previous literature [3].

All the algorithms showed a coherent 3-clusters solution, with a high-risk clusters composed by 63 subjects where 9 out of 10 suicidal patients belonged. Cluster assignments showed a high level of agreement across algorithms, with Jaccard similarity indices ranging from 0.95 to 0.96. This strong concordance suggests that the definition of subgroups is primarily driven by the structure of the PLS-derived latent space rather than by the specific clustering algorithm employed.

Regarding the clinical and demographic variables, the clusters significantly differ for family history of psychiatric disorders across all three algorithms; moreover, significant differences between the clusters were obtained for number of admissions lifetime and in the last 12 months, abuse of recreational substances, suicide attempts in the last 12 months, use of lithium (with the subjects classified in the high risk group showing less use of lithium), in two out of three clustering solutions.

Conclusions

We successfully identified a three-cluster model based on neuroanatomical features, stable across different clustering solutions.The areas involved in the clustering include frontal and cerebellar cortices, as already described in the literature [3].The clusters showed different levels of suicide risk and presented some significant differences in clinical and demographic characteristics.

References

1-Pigoni A, et al., Machine learning and the prediction of suicide in psychiatric populations: a systematic review. Transl Psychiatry, 2024, 14, 140. https://doi.org/10.1038/s41398-024-02852-9 2-Dwyer et al., Psychosis brain subtypes validated in first-episode cohorts and related to illness remission: results from the PHENOM consortium, 2023, Molecular Psychiatry 28. https://doi.org/10.1038/s41380-023-02069-0 3-Pigoni, et al. Multimodal Machine Learning Prediction of 12-Month Suicide Attempts in Bipolar Disorder. Bipolar Disorders, 2025, https://doi.org/10.1111/bdi.70011